Disclosure: the author holds none of the companies named here.

Why this matters if you own any AI stock

Many AI stocks depend on the same thing: a handful of very large companies spending money on data centers.

That spending is someone else’s income. When Meta or Amazon buys servers, cables, cooling and power equipment, the companies that make those things get paid. Chip makers, fiber makers, parts makers and power companies all sit downstream of that money.

So the question is simple. Where is the money for the build coming from? If it comes out of what these companies earn, the build can keep going as long as their businesses do well. If more of it comes from borrowing, the build starts to depend on lenders too.

In mid-October, lenders showed how they treat a smaller AI company that borrows heavily to build. They charged CoreWeave far more to carry its debt, and its shares lost a quarter of their value in five trading days. This piece is about the giants.

Four terms anyone can read

Each of the four builders publishes the figures behind these terms every three months. Here is what each one means, in dollars.

Capital spending is money spent on buildings and equipment that will be used for years. A data center is capital spending.

Cash from the business, which companies call operating cash flow, is the cash that came in from running the business during the quarter, after paying staff, suppliers and taxes.

Cash after spending is the first number taken away from the second. If a company brings in $40 billion of cash and spends $45 billion building, its cash after spending is negative $5 billion. It had to cover that $5 billion from savings or by borrowing.

Long-term debt is money the company owes that is not due within the next year. Debt due within the year is listed separately, and the two together are total debt. If debt goes up, the company usually borrowed more than it paid back.

Those four terms are all you need through Friday 30 October.

What the four latest reports show

Each figure below comes from the company’s own latest quarterly results or quarterly report. The quarters are not the same for all four, so each one is dated.

Meta, in its quarter ended 30 September 2026: cash from the business $32.79 billion, capital spending $33.94 billion. Cash after spending: negative $1.16 billion.

Alphabet, in its quarter ended 30 June 2026: $39.07 billion in, $44.92 billion spent. Cash after spending: negative $5.86 billion.

Amazon, in its quarter ended 30 June 2026: $45.39 billion in, $53.08 billion spent after small sales of equipment. Cash after spending: negative $7.69 billion.

Microsoft, in its quarter ended 30 June 2026: $55.44 billion in, $35.80 billion spent. Cash after spending: about positive $19.64 billion.

Three of the four spent more on building than their businesses brought in.

Meta: the plan went up, the cash went the other way

Meta’s results for the quarter to 30 September 2026 show both halves of the story. It now plans to spend between $135 billion and $145 billion in 2026, counting lease payments, up from a floor of $130 billion.

By Meta’s own count, which also subtracts those lease payments, cash after spending came to minus $2.12 billion. A year earlier the same measure was plus $10.63 billion.

Its long-term debt stood at $80.92 billion on 30 September, against $58.74 billion when 2025 ended.

Meta’s shares fell 6.98 percent on 22 October, the day after the results, Yahoo Finance reported. Meta announced a 2-for-1 stock split the same day, so the report did not tie the fall to one cause, and neither will I.

Alphabet: $56.23 billion borrowed in six months

Alphabet’s cash flow statement shows it raised $56.23 billion by issuing new debt in the six months to 30 June 2026. Its long-term debt went from $46.55 billion at the end of 2025 to $98.17 billion. That is more than double.

Its spending plan went up too. On 23 July, Bloomberg News reported that Alphabet raised its 2026 forecast to between $195 billion and $205 billion. In a presentation dated 3 June, Alphabet had put it at $180 billion to $190 billion.

Alphabet reports its next quarter on 28 October.

Amazon: debt nearly doubled

Amazon’s long-term debt was $65.65 billion at the end of 2025. On 30 June 2026 it was $128.89 billion, according to its quarterly report. That is nearly twice as much.

Amazon publishes its own cash after spending as a twelve-month figure. For the twelve months to 30 June 2026, it was negative $7.60 billion. A year earlier, the same figure was positive $18.18 billion.

Amazon’s results release did not give a dollar plan for 2026 spending, so the test reads what it says in its next report on 29 October.

Microsoft: the exception, with one caution

Microsoft brought in more cash than it spent. Its total debt, counting the part due within a year, was $40.26 billion at the end of 2025 and $40.29 billion on 30 June 2026. In other words, it barely moved.

The caution: the spending figure here counts equipment a company buys. It leaves out equipment rented under long leases. Meta’s report shows those lease payments came to $964 million in its latest quarter. Microsoft’s results release gives no lease figure, so this count may flatter Microsoft, and I cannot say by how much.

Microsoft reports on 28 October, after the market closes.

The hidden assumption

The hidden assumption behind the AI build story is that the giants pay for it out of what they earn. On their latest reports, three of the four did not. They borrowed or used savings to cover the gap.

Borrowing to build is not a warning sign by itself. What changes is who carries the risk if the build does not pay off.

That is not a forecast. These are very large businesses, and a few quarters of heavy building can be paid back. What these numbers do show is that the build now rests on lenders as well as earnings, for three of the four.

The week ahead, and what each day answers

Tuesday 27 October. Corning, which makes the optical fiber used in data centers, reports its results, with a call at 8:30 a.m. Eastern. The day’s notes ask whether a supplier’s sales kept growing while some of its customers borrowed to pay for the build.

Wednesday 28 October. The second piece of the week: who gets the money. It shows how to find, in a supplier’s own reports, how much it depends on these four builders. Amphenol, which makes connectors and cables used in data centers among other products, reports the same day. The Federal Reserve announces its interest rate decision that afternoon, and Alphabet and Microsoft report after the market closes.

Thursday 29 October. Notes read Alphabet’s and Microsoft’s numbers against the test below. Amazon reports after the close.

Friday 30 October. The scorecard for all four, what it means, and the lesson.

The test, set before the results

Here is how Friday will score each builder. I am publishing it now so the results cannot choose it.

1. The spending plan: raised, held or cut, against the last plan the company gave.

2. Cash after spending for the quarter: positive or negative.

3. Debt, counting the part due within a year, against the previous quarter: up or down.

Each builder then gets one of three readings. Paid from earnings, if cash after spending is positive. Paid partly by borrowing, if it is negative and debt went up. Paid from savings, if it is negative and debt did not go up.

You can score it yourself. Most of these numbers are in each company’s results release on the day it reports. Microsoft’s release says it gives its outlook on its results call, so for Microsoft I will read the call transcript.

The question to carry through 30 October: of Alphabet, Amazon and Microsoft, which do you expect to report the biggest gap between what it spends and what its business brings in?

Not investment advice. The subscriber decides.

Disclosure: the author holds none of the companies named here.

What happened, in two minutes

CoreWeave rents out computer power to companies building AI. To build its data centers, it borrowed. Its latest quarterly report, called a 10-Q, shows $35.07 billion of debt at 30 June 2026. A 10-Q is the report every US public company files with regulators every three months.

That report shows something else. In the three months to 30 June, CoreWeave paid $640 million in interest on that debt. Its loss for the same three months was $626 million. In other words, the interest bill alone was enough to wipe out the quarter.

Lenders noticed. In the week of 12 October they started charging much more to carry CoreWeave’s risk, and its stock fell 25.2 percent between the 9 and 16 October closes. Five other AI-related stocks fell with it, including Fabrinet (NYSE: FN), a parts maker with no debt at all.

That raised the real question. Was the fear spreading through the whole lending system, or stuck on one company?

Why one company’s debt can reach other companies

Think about what happens when a bank gets nervous about lending.

It charges more. Companies that depend on borrowing slow down their building. The suppliers that sell them parts and power see fewer orders. That is how a problem at one borrower could, in theory, reach companies that borrowed nothing.

So the test is simple. If the fear is spreading, the cost of borrowing should rise for many companies, not just one, and the big AI spenders should start cutting back. Four numbers show whether that is happening.

Number one: the price of insuring a loan

Lenders can buy insurance against a borrower not paying them back. It is called a credit default swap. The buyer pays a yearly fee, and if the borrower fails to pay, the insurer covers the loss.

The fee is a good thermometer. When lenders are calm, it is low. When they are scared, it jumps.

CNBC reported on 16 October that insuring CoreWeave’s debt cost more than 1,000 basis points a year. A basis point is one hundredth of one percent, so 1,000 basis points is 10 percent.

Put simply, a lender paid more than $10 a year to protect every $100 it had lent to CoreWeave.

Reading: warning, for one company. These prices sit behind paid professional services, so this is the one number you cannot check for free. News outlets report it when it moves.

Number two: the extra interest companies pay

Every company pays more to borrow than the US government does, because the government is treated as the safest borrower there is. That extra amount is called a credit spread.

Here is how it works in dollars. If the government pays 5 percent and a company pays 5.85 percent, the spread is 0.85 points. The company pays 85 cents more a year for every $100 it borrows.

The safest companies, called investment grade, saw their spread move from 0.83 to 0.85 points between 9 and 16 October. That is almost nothing.

Riskier companies, called high yield, saw theirs rise from 3.11 to 3.31. That move was ten times bigger.

Reading: watch. Worry reached the riskier borrowers. It did not reach the strongest ones. Both numbers are free every business day on FRED, the St. Louis Federal Reserve’s public data website.

Number three: what the US government pays to borrow

The interest rate on a 10-year US government bond is called the 10-year Treasury yield. It matters to almost every borrower, because many other loans are priced off it, from company debt to home mortgages.

If the debt scare were spreading through the whole system, you might expect this rate to rise. It did the opposite. It fell from 5.24 percent on 9 October to 5.09 percent on 16 October, according to the Treasury’s own daily figures. CNN reported money moving out of chip stocks and into government bonds, which pushes the rate down.

Reading: calm, during the scare. One caution. A rate above 5 percent is still expensive money. It just did not get worse during the scare. The Treasury publishes this number free every business day.

Number four: what the big AI builders plan to spend

Capital spending is money a company spends on buildings and equipment. For the biggest AI companies, it means data centers. That spending flows to the companies that supply data center parts, such as Fabrinet, whose customers include Nvidia and Amazon.

So this is the number that decides whether a debt scare reaches the suppliers’ actual sales, not just their stock prices.

Meta reported on 21 October. It spent $34.91 billion on buildings and equipment in the three months to 30 September 2026. It raised the low end of its plan for the year to $135 billion, with the top still at $145 billion. In July the plan started at $130 billion.

Reading: calm, for spending. Microsoft reports on 28 October and Amazon on 29 October. Each will update this number.

One thing to watch inside Meta’s report. Free cash flow is the cash a business has left after paying for its buildings and equipment. Meta’s was negative $2.12 billion for the quarter. In plain terms, it spent more on building than its business brought in. A gap like that has to be covered from cash on hand or new borrowing. In the quarter before, it had $784 million left over.

What the four numbers say together

One number gave a warning, one said watch and one stayed calm. The fourth, the AI builders’ spending, kept growing, though Meta now spends more than its business brings in.

Put plainly, the fear was real but narrow. The stock prices of six companies fell together, as if one company’s debt problem belonged to all of them. The lending numbers say the trouble stayed mostly with one heavy borrower and the riskiest corner of the loan market.

The hidden assumption this week tested was that a debt problem at one company spreads to every company near it. The share prices acted as if it had. The four numbers say that, so far, it has not.

That is a reading of one week, not a prediction. Any of these numbers can change, and each one is published on a schedule you can follow.

The lesson: a falling price is a question, not an answer

A falling price tells you that people are afraid. The company’s own reports tell you whether they are right.

Fabrinet shows why. Its annual report shows no debt. It still fell as far as CoreWeave, which owes $35.07 billion. The stock chart could not tell the two apart. The reports could, in an afternoon.

Corning (NYSE: GLW) shows the second half of the lesson, which is about timing. A common way to judge a stock’s price is to ask how many dollars you pay for each $1 of the company’s yearly profit. On 6 October, a Corning share cost about $59 for every $1 of its yearly profit. By 16 October it cost about $46.

The investor Peter Lynch used a simple rule of thumb: a fair price is roughly equal to how fast profit is growing. A company growing profit 30 percent a year is fairly priced at around $30 for every $1. Corning’s profit grew about 30 percent in its latest reported quarter, using the adjusted measure Corning reports, which it calls core earnings. So the scare made Corning cheaper. It did not make it cheap by that rule.

An investor who had worked that out before the scare knew within a minute whether the drop mattered. One who had not was doing the math in a hurry, which is how people talk themselves into bad decisions.

One thing I got wrong

In August I wrote that the US government’s borrowing cost was the place this kind of pressure would show up first, partly because foreign governments had been selling US bonds. During the scare, that number went the other way. The Treasury’s latest monthly report, released 16 October, shows foreign governments bought $18.1 billion of US government bonds in August.

I am printing that because the method only works if I show you the numbers that disagree with me, not just the ones that agree.

How to watch this yourself

You can follow all four from here. Three are free.

1. The price of insuring CoreWeave’s debt: reported by CNBC and other outlets when it moves.

2. Credit spreads: free on FRED every business day. Search “ICE BofA US Corporate” and “ICE BofA US High Yield.”

3. The 10-year Treasury yield: free on the US Treasury’s website every business day.

4. AI builders’ spending plans: Microsoft on 28 October and Amazon on 29 October, in each company’s quarterly results.

The Federal Reserve, which sets short-term interest rates, also meets on 27 and 28 October. Its decision moves number three.

One question to carry into the weekend: which of these four numbers would have to change before you felt comfortable buying any of these companies?

Not investment advice. The subscriber decides.

Disclosure: the author holds none of the companies named here.

The question from Tuesday

On Tuesday I asked which of four companies was hit hardest by debt that sits on someone else’s balance sheet. Fabrinet, Bloom Energy (NYSE: BE), CrowdStrike (Nasdaq: CRWD) and Corning (NYSE: GLW) all fell in the week of 12 October. Here is each one, set against its own debt and cash, from its latest filing.

Read the graphic from top to bottom and one thing stands out. The size of each fall does not line up with the size of each company’s debt. The company with the most debt and the company with none fell the same amount. The filings show three different routes the fear took.

Route one: your revenue depends on borrowed spending

Fabrinet makes optical and precision parts for other companies. Its 10-K for the year ended 26 June 2026 lists four customers that each bought more than 10 percent of its output: Cisco at 19.9 percent, Nvidia at 16.3, Nokia at 10.7 and Amazon at 10.5. Together they made up 57.4 percent of revenue, and data center customers made up 47.9 percent.

Nvidia sells CoreWeave its chips and is a large CoreWeave shareholder, according to CNBC. Amazon runs the AWS cloud. When lenders start pricing doubt about how the AI build gets paid for, Fabrinet’s revenue sits downstream of that build.

Owing nothing protects your balance sheet. It does not protect your customers’ order books.

The filing cannot tell you why anyone sold Fabrinet. It can tell you whose spending Fabrinet’s revenue depends on. That is the answer to Tuesday’s question.

Bloom Energy: profitable, and concentrated

Bloom Energy makes fuel cells that generate electricity on site. Its 10-Q for the quarter ended 30 June 2026 shows revenue of $1.07 billion, against $401.2 million a year earlier. It earned $196.3 million, after losing $42.6 million in the same quarter of 2025. Its $2.67 billion of cash exceeds its $2.48 billion of debt.

Then the customer note. One customer made up about 44 percent of the quarter’s revenue and a second about 21 percent. The note gives percentages, not names. Two buyers produced roughly two thirds of the quarter.

Bloom sits on route one only if those two buyers are part of the AI build, and the filing does not say who they are. On that point the answer is NOT PROVEN. What the filing does show is two buyers carrying two thirds of a quarter. Its stock fell 21.1 percent between the 9 and 16 October closes.

Route two: your price assumed cheap money

CrowdStrike, the cybersecurity company, held $5.01 billion of cash against $746.2 million of long-term debt at 31 July 2026, per its 10-Q. Revenue for that quarter was $1.47 billion, up 25.8 percent. Net income was $5.3 million.

At the 16 October close of $227.39, its 1,023,934,842 shares were worth about $232.8 billion together. That price rests on profits that have mostly not arrived yet.

A high price is a bet on profits far away. When safe government bonds pay more, a profit ten years out is worth less today, whatever the company owes. My August piece called this the second route by which rising rates reach AI stocks. CrowdStrike fell 17.3 percent.

CrowdStrike also had news of its own that week, and its fall had more than one cause in the coverage I read. It carries the least weight of the four.

Route three: a borrower that earns

Corning fell from $156.68 at the 9 October close to $132.00 on 16 October, a drop of $24.68 a share, or 15.8 percent. Those were the same five days The Long View spent on Corning.

Corning does carry debt. Its 10-Q shows $7.76 billion of long-term debt and $668 million of current debt and short-term borrowings at 30 June 2026, about $8.42 billion together, against $2.50 billion of cash. That is more than Fabrinet, Bloom and CrowdStrike carry combined.

Corning carries the most debt of the four and fell the least.

It also earns. Net income for the quarter ended 30 June 2026 was $559 million. Optical communications products brought in $2.07 billion of its $4.51 billion of sales, about 46 percent. My reading is that a borrower paying its interest out of profit sits on a different route from a borrower paying it out of new loans. That is a reading, and Friday’s gauges test it.

The hidden assumption

The hidden assumption this week tests is simple: a company with a clean balance sheet sits outside a debt scare.

Fabrinet’s 25.2 percent says it does not. A company can owe nothing to lenders and still depend on customers who are spending borrowed money. A company can owe very little and still be priced on cheap money continuing.

The line in a 10-K that shows your route is the customer note. Search the filing for “accounted for” and read the percentages. If a few customers produce most of the revenue, find out how those customers are paying for what they buy.

What one week cannot tell you

Three limits again. Nothing in a filing says why any particular investor sold. Five days of prices can reverse, and a fall is not a forecast. And Corning’s numbers suggest that debt size alone did not decide who fell most. That is one week of evidence, not a rule.

What the filings do show is the order of exposure: the borrower, the companies selling to borrowers, and the companies priced for patience.

Spreading or fading

Five companies, three routes. The open question is whether this spreads or fades, and you do not need a forecast to find out. Four published numbers will show it, each on its own schedule: the price of insuring AI borrowers’ debt, investment-grade credit spreads, Treasury yields, and the spending plans of the companies building AI data centers.

Tomorrow’s notes take those four one at a time, with where each is published. Friday’s piece scores them.

Not investment advice. The subscriber decides.

Disclosure: the author holds none of the companies named here.

The week the lenders raised their price

The week started with an inflation report. On Wednesday 14 October the Bureau of Labor Statistics reported that consumer prices rose 3.7 percent in the 12 months ending September, up from 3.4 percent for the 12 months ending August, with gasoline accounting for more than half of the monthly increase. Inflation that refuses to fall keeps another rate rise on the table, and a borrower feels that before anyone else.

CoreWeave, which rents out AI computing power, fell 15 percent that day, according to Forbes. By the end of the week, CNBC reported that insuring CoreWeave’s debt cost more than a tenth of the amount insured, every year, alongside a jump in the yields on its bonds. A rising yield means the bonds’ price fell.

Nvidia’s chief executive, Jensen Huang, told Bloomberg that CoreWeave’s position was stable and improving, and the stock closed higher that day. On Friday it fell again, by nearly 12 percent, and the Nasdaq Composite finished the week down about 4.5 percent, per CNBC. From $82.08 at the 9 October close to $61.39 at the 16 October close, CoreWeave’s shares fell 25.2 percent.

Nebius (Nasdaq: NBIS), which runs the same kind of business, fell from $221.07 to $165.47 between the 9 October and 16 October closes, also 25.2 percent. Its most recent quarterly results were released on 12 August, per the company.

In August I wrote that the AI build was increasingly paid for with borrowed money, and that credit spreads were one of four gauges worth watching. This was the week that gauge moved, for one company.

What a credit default swap measures

A credit default swap is insurance on a loan. The buyer pays a yearly fee, and if the borrower fails to pay, the seller covers the loss. The fee is quoted in basis points, hundredths of a percentage point, so 1,000 basis points means 10 percent of the amount insured, every year.

A stock price is what buyers will pay for the upside. A swap price is what lenders charge to carry the downside.

The lenders’ price comes from the same filing anyone can download. So I opened it.

What CoreWeave’s 10-Q shows

CoreWeave’s Form 10-Q for the quarter ended 30 June 2026 tells a growth story first. Revenue was $2.58 billion, against $1.21 billion in the same quarter of 2025, an increase of 112 percent.

Then read further down the income statement. Interest expense, net, was $640 million, against $267 million a year earlier, about 2.4 times as much. The business itself ran an operating loss of $49 million. The net loss for the quarter was $626 million.

The quarter’s net loss was almost exactly the quarter’s interest bill.

Revenue grew fast. The cost of the money that built the capacity grew faster. That gap is the reason a growing company can still frighten its lenders.

How the debt got there

The debt note in the same filing puts recourse debt at $31.41 billion and non-recourse debt at $3.66 billion on 30 June 2026, about $35.07 billion in total. On 31 December 2025 the same two lines totalled $21.37 billion. Recourse debt is money the company itself must repay. On non-recourse debt, lenders can look only to specific assets.

Add $16.54 billion of lease liabilities and you reach $51.6 billion, which matches the more than $51 billion of debt and obligations in CNBC’s coverage. When you see that number, know that it includes leases.

The cash flow statement shows why the debt grew. In the six months ended 30 June 2026, CoreWeave spent $14.12 billion on property and equipment. Its operations brought in $3.66 billion. It covered the gap by raising $16.75 billion of new debt, repaying $5.22 billion, and selling $2.98 billion of stock.

The price of that money is in the note too. The delayed-draw loan facilities carry effective rates of 9 to 15 percent, the senior notes 9 to 10 percent, and the convertible notes 2 percent. Of the total, $7.51 billion was classed as due within a year of 30 June 2026.

The hidden assumption

Against that debt sits a promise. The 10-Q reports $103.7 billion of remaining performance obligations at 30 June 2026. That is contracted work not yet delivered. The company expects 41 percent of it, about $42.5 billion, to become revenue in the 24 months ending 30 June 2028.

So here is the hidden assumption, stated plainly: CoreWeave’s customers pay on schedule, and that money arrives faster than the debt costs and comes due.

The filing tests part of it. One customer made up 36 percent of revenue in the quarter ended 30 June 2026, a second 26 percent and a third 10 percent. The revenue note labels them only as Customer A, B and C. Three customers produced 72 percent of the quarter’s revenue.

$103.7 billion of promises is not $103.7 billion of cash.

Some of it has been paid in advance. The same filing shows $9.7 billion of deferred revenue, customer money received for work not yet done. The rest arrives only as the work is delivered.

The filing tells you how much is owed to CoreWeave and by roughly when. It cannot tell you whether those three customers will keep paying.

What the bond market did not say

Three limits belong here. A default swap price is not a forecast of default. It is a price, and it can reverse in a day, as Thursday showed. This was also one borrower. On 9 October, the Friday before, ProShares noted that investment-grade credit spreads had barely moved, with nearly the whole rise in corporate bond yields coming from Treasury rates. Whether that held through the week is one of the gauges Friday’s piece checks. And nothing here tells you why any particular investor sold.

What the week does show is where the stress appeared first: in the price of lending to a company whose quarterly loss matches its interest bill.

The company that owes nothing

Here is what CoreWeave’s filing cannot explain. Fabrinet (NYSE: FN), which does optical and precision manufacturing for communications and data center customers, reported $875.1 million of cash and short-term investments and no outstanding debt at 26 June 2026, in its Form 10-K. It has no bonds for lenders to reprice.

Its stock fell anyway, from $486.28 at the 9 October close to $363.51 on 16 October. That is a fall of 25.2 percent, the same as CoreWeave’s to one decimal place.

Debt that is not yours can still reach you. Tomorrow’s notes lay out the clues. Wednesday’s piece follows the route from the borrower to the companies around it: Fabrinet, Bloom Energy, Corning and CrowdStrike.

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in LUMN, DY, APH, LITE or COHR.

The five: Lumen (Nasdaq: LUMN) owns long-haul fiber. Dycom builds and repairs fiber networks. Amphenol makes the plugs and cable inside data center racks. Lumentum (Nasdaq: LITE) and Coherent (NYSE: COHR) turn electrical signals into light. Prices are the 12 October 2026 closes, per CNBC.

The verdict on all five

On Monday I set a bar before any company faced it. Earnings I can defend from the company’s own filings. A price that is not too high for its growth. And more cash coming in than going out last year.

Dycom and Amphenol passed all three. Both are classed and deserves serious review.

Lumen and Lumentum reported losses, so there were no earnings to test. Both are NOT PROVEN, which means the test could not run, not that they failed. Coherent had earnings but spent $1.02 billion more cash than its operations brought in. It is keep on watch.

Thursday asked whether passing two companies meant the bar was too easy. At Peter Lynch’s stricter line, from One Up on Wall Street, Amphenol still passes and Dycom does not. And both passes need growth to stay about where it is today.

The other promises, kept

Tuesday showed what Lumen kept after selling its home fiber to AT&T: its long-haul network, its business customers and its copper. It also found that Dycom’s own 10-K calls it “a leading provider,” not the largest, whatever Google’s AI says. Thursday read Coherent’s own filings first. Monday checked the hyperscalers’ spending plans where the companies published them.

The last promise was who is paying for all this fiber. The hyperscalers, the few giant cloud companies, are the biggest and fastest-growing buyers. They are not the only ones. Most of Lumen’s revenue comes from ordinary businesses and government. Dycom’s biggest customers are phone companies. No single customer buys even 10 percent of Amphenol’s sales.

IDC counted data center switch sales up 64.5 percent in the second quarter of 2026, and office switch sales up 15.7 percent. Everyone else is upgrading. The data centers are upgrading about four times faster.

How to read what each company has to do

A stock price goes up in only two ways. The company earns more, or buyers pay more for each dollar it earns.

What buyers pay for each dollar of yearly earnings is the price-to-earnings ratio. Today it is high for all of these companies. If it comes down, earnings have to grow just to keep your money where it started.

So here is the test for each company. Five years from now, in October 2031, assume buyers pay 20 times earnings. Lynch would call that fair for a company growing 20 percent a year. That 20 is my assumption, not a forecast. Then ask: how much would earnings have to grow for your five thousand dollars to be worth five thousand dollars again? Anything above that is your profit.

Then check the company’s own record since 2015. Has it ever grown that fast for five years?

Dycom: earnings up about 21 percent

Five thousand dollars bought 18.73 shares at $266.99.

Dycom earned $10.99 a share over its last four quarters. For your money to be whole in October 2031, it needs to earn about $13.35. That is growth of about 4 percent a year. Everything above that is profit.

Its record says it can grow much faster than that. In each five-year stretch I could measure, ending in fiscal 2024, 2025 and 2026, Dycom grew earnings by 30 percent a year or more.

The catch is where those stretches started: from bad years. The last time Dycom started from a good year, in fiscal 2017, its earnings per share fell from $4.92 to $1.07 by fiscal 2021. Today it starts from its best year in the records I read. Its order book is strong: $12.24 billion of backlog at 1 August 2026, up 53.2 percent in a year. But its buyers are mostly phone companies, and AT&T alone is a quarter of its revenue.

Dycom needs the least. The risk is a downturn, not a missed growth target.

Amphenol: earnings more than double

Five thousand dollars bought 58.01 shares at $86.19.

Amphenol earned about $2.00 a share over its last four quarters, adjusted for its stock split. For your money to be whole in October 2031, it needs to earn about $4.31. That is more than double, or growth of 16.7 percent a year for five years.

Its record says that is possible but not easy. In six five-year stretches since 2015, it grew that fast twice. Over the full ten years it grew 18.7 percent a year, just above the bar.

Right now it is growing much faster. In 2025, its earnings per share jumped 74 percent. But growth without acquisitions has slowed for three quarters in a row, from 41 percent to 30 percent. No customer buys 10 percent of its sales, the widest spread of buyers of the five. Amphenol reports again on 28 October.

Amphenol has to keep growing at a pace it has hit only twice in ten years, just to give your money back.

Coherent: earnings up three and a half times

Five thousand dollars bought 17.36 shares at $288.02.

Coherent earned $4.12 a share in its last fiscal year. For your money to be whole in October 2031, it needs to earn about $14.40. That is 3.5 times as much, or growth of 28.4 percent a year for five years.

Its record has never done that. In seven five-year stretches since 2015, its best was 17.9 percent a year. It reported a loss in four of its last seven fiscal years.

Coherent also has to stop burning cash. Last year it spent heavily on factories and inventory. If those orders turn into sales, the cash comes back. Coherent guides to sales between $2.2 billion and $2.4 billion for the quarter to September 2026, up from $1.58 billion a year before. Its two biggest customers make up 32 percent of its revenue, and neither is named in the filings I read.

Lumentum: earnings up nearly six times

Five thousand dollars bought 5.02 shares at $995.41.

Lumentum reported a loss last year, almost all of it one charge on its debt. On its own adjusted figures it earned $8.67 a share. For your money to be whole in October 2031, it needs to earn about $49.77. That is 5.7 times as much, or growth of 41.8 percent a year for five years.

Its record has never done that either. It reported a loss in six of its last eleven fiscal years. Its earnings rise and fall with the industry’s cycle.

It is growing fast now. It expects $4.05 to $4.35 a share of adjusted earnings in the quarter to September 2026. If it delivers, that one quarter would equal about half of what it earned all last year. Two customers make up 41.6 percent of its revenue.

Lumen: first, a profit

Five thousand dollars bought 990 shares at $5.05.

Lumen lost money over its last four quarters. For your money to be whole at 20 times earnings in October 2031, it needs to earn about $0.25 a share each year. So the first job is simple to say and hard to do: turn a loss into a steady profit.

Its new business is growing, but its old one is shrinking faster. It owes $13.21 billion, about 2.5 times what all its shares are worth. Its chief executive bought stock this year at about $6 a share, above the $5.05 close.

Which one gets the deep dive, and your routine

Monday said only one company would earn a deep dive, the weeks of reading I do before any money moves. It goes to Dycom. It needs the least growth to give your money back, and more of its price is backed by what it earns today. At a 10 percent return, Dycom’s last year of earnings is worth about 41 percent of its share price. For Amphenol, it is 23 percent. Corning’s piece asked which of these comes closest to the Buffett and Lynch playbook. Dycom comes closest to Buffett. Amphenol comes closest to Lynch.

This is my reading, not a finding. The AI buildout is real money for these companies. What no filing can show is earnings that keep growing for a decade after it ends.

The risk is not that the AI buildout stops. The risk is that it is already in the price.

Your routine for any stock:

1. Write down the hidden assumption: what has to be true for the story to hold.

2. Find what the company earned last year, under standard accounting.

3. Work out what it must earn in five years for your money to be whole.

4. Check whether it has ever grown that fast, and from where.

5. Find out who the customers are before you believe any growth rate.

Reply with a ticker you own or are watching, and I will run it through these five steps. I read every reply.

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in LITE or COHR.

Friday’s Corning piece promised that Coherent’s own filings would come first when its day arrived. So Coherent goes first.

The Firewall’s hidden assumption is the same for both companies: that AI demand for their optical parts stays above their capacity long enough to turn scarcity into durable, high-margin growth. The filings can test today’s half of that sentence. Neither can test the long half.

Coherent, from its own filings

Coherent sells what its Nvidia release calls “advanced laser and optical networking products.” In its fiscal year ended 30 June 2026, its Datacenter and Communications segment brought in $5.27 billion, up 40 percent, and its Industrial segment $1.84 billion, down 10 percent. Datacenter and Communications was 74 percent of revenue.

Nvidia bought 7,788,161 Coherent shares at $256.80 each, according to Coherent’s 8-K. The 10-K adds terms the 8-K left out: a six-month lock-up from the closing date, and a price protection provision for the same period. The press release describes “an NVIDIA multi-billion-dollar purchase commitment” but gives no amount and no length, and the filings I read add none.

Coherent closed at $288.02 on 12 October 2026, per CNBC, 12.2 percent above what Nvidia paid.

Who pays the light makers

Coherent’s 10-K: “One major customer accounted for 20%” of revenue in fiscal 2026, and “a second major customer accounted for 12%.” Revenue from them “was primarily attributable to our Datacenter & Communications segment.” Neither is named.

Lumentum’s is steeper. Its largest customer, named only as Customer A, accounted for 26.6 percent of its revenue in the year ended 27 June 2026, up from 15.4 percent the year before. Customer B was 15.0 percent. Two buyers, 41.6 percent of the company.

Set that beside the rest of the week. Dycom’s largest customer, AT&T, was 25.4 percent of its revenue. Amphenol says no customer reached 10 percent.

In this week’s filings, the closer a company sits to the light inside a data center, the fewer buyers it depends on.

That is the answer this week has been building toward. The optical layer is where customer concentration shows up in these filings. Neither company names its customers, so whether they are hyperscalers, Nvidia or someone else is NOT PROVEN.

Coherent’s earnings and its cash

Wednesday’s piece said Coherent’s cash flow tells a different story from its earnings. Here it is.

Coherent’s net earnings for fiscal 2026 were $786.9 million. Its operations brought in $79.5 million of cash. The year before, operating cash flow was $633.6 million.

The cash flow statement shows where it went. Inventories absorbed $1.18 billion and receivables $367.6 million. Then Coherent spent $1.10 billion on property and equipment, up from $440.8 million, with another $371.8 million of equipment received but not yet paid for. Operating cash minus capital spending comes to negative $1.02 billion, by my arithmetic. The release does not explain the swing, and I could not reach the part of the 10-K that would.

The earnings also include one-off gains: $124.1 million from selling businesses, including its aerospace and defense unit, and $74.0 million from selling an investment.

Coherent earned $786.9 million and its operations produced $79.5 million of cash. Both numbers are true, and they answer different questions.

The hidden assumption, tested

The scarcity shows up in Lumentum’s numbers. Revenue rose 83 percent in the year to $3.01 billion. Fourth-quarter revenue was $1.006 billion, and the company guides to $1.225 billion to $1.275 billion for the quarter after. Gross margin under standard accounting was 47.4 percent in the fourth quarter, against 28.0 percent for the year before.

Coherent’s numbers are less clear. Datacenter and Communications grew 40 percent, but gross margin under standard accounting was 38.5 percent in the fourth quarter, up 2.77 points from a year earlier. Inventory rose from $1.44 billion to $2.58 billion, most of it work in progress. That can mean parts waiting for demand or parts racing to meet it. The filings I read do not say which.

Coherent against the bar

Test one, defensible earnings: $4.12 a share under standard accounting, gains included. Test two, price against growth: Coherent lost $0.52 a share the year before, and growth from a loss cannot be measured as a rate, so this test cannot run. Test three, cash: negative $1.02 billion. It fails.

Coherent is classed keep on watch. A company that is building as fast as this one may be making the right decision. The bar does not judge that. It measures what a buyer gets for the price today.

At $288.02, Coherent is valued at about $56.7 billion, 69.9 times its standard earnings. A 10 percent return on those earnings supports about $41.20 a share. Five thousand dollars buys 17.36 shares. Ignoring any dividend, a total gain of 30 percent by October 2031 needs about $374.43, and 30 percent a year needs about 3.7 times today’s price.

The balance sheet can carry the build for now. At 30 June 2026, Coherent held $1.16 billion of cash and $825.0 million of short-term investments, against $3.22 billion of debt. Its outlook for the September quarter calls for revenue of $2.2 billion to $2.4 billion, against $1.58 billion a year earlier. If that arrives, the inventory may be turning into sales. The next cash flow statement is the one to read.

Lumentum, the same deal at a higher price

Lumentum’s version of the Nvidia deal was 2,876,415 shares of convertible preferred stock at $695.31 each, according to its 8-K. Lumentum closed at $995.41 on 12 October 2026, per CNBC, 43.2 percent above that price.

The purchase commitment is harder to find. Its press release says “the non-exclusive agreement includes an NVIDIA multi-billion purchase commitment and future capacity access rights.” The 8-K’s own items say nothing about it, and neither do the notes to the 10-K that I read. The $2.40 billion of “legally-binding purchase commitment obligations” in Lumentum’s 10-K is Lumentum’s own commitment to its suppliers, not Nvidia’s to Lumentum.

Which earnings for Lumentum

For the year ended 27 June 2026, Lumentum reported a loss of $92.96 a share under standard accounting. Its operating income was positive, $524.8 million. The loss came from a $7.76 billion charge when convertible notes were swapped for shares, which its release calls “one-time, non-cash.” On its own adjusted basis, Lumentum earned $8.67 a share.

The bar has used standard-accounting earnings for every company this week. On that basis Lumentum has no earnings to divide by, so the price test cannot run. That is the same answer Lumen got on Tuesday, for a very different reason.

A cyclical in a fast grower’s clothes

Run the adjusted figure anyway and Lumentum looks remarkable. Adjusted earnings rose from $2.06 to $8.67, up 321 percent. At $995.41, that is 114.8 times adjusted earnings, or 0.36 times the growth rate, comfortably inside the bar.

The longer record says why that number cannot be trusted. Lumentum’s adjusted earnings were $5.84 a share in fiscal 2021 and $6.05 in fiscal 2022. They fell to $4.56 in 2023 and to $0.44 in 2024, as later restated. Revenue went from $1.77 billion in 2023 to $1.36 billion in 2024. Standard-accounting operating income swung from $527.0 million in 2021 to a loss of $434.0 million in 2024.

Measured from fiscal 2022’s $6.05 instead, adjusted earnings grew 43 percent in four years. Lumentum changed its adjusted method twice in that span, so the older figures are not strictly comparable. The shape is clear anyway. Measured from the bottom of a cycle, a cyclical looks like a fast grower. Lynch’s price-to-growth yardstick was built for fast growers, and my reading is that it misleads here.

Lumentum against the bar

Test one, defensible earnings: a standard-accounting loss, almost all of it one non-cash charge. Test two: cannot run. Test three, cash: operating cash flow of $751.4 million less $451.3 million of capital spending leaves $300.1 million, positive.

Lumentum is classed NOT PROVEN. At $995.41, it is valued at about $92 billion counting Nvidia’s preferred as converted. Even on its own adjusted figure, a 10 percent return supports about $86.70 a share. Five thousand dollars buys 5.02 shares. Ignoring any dividend, a total gain of 30 percent by October 2031 needs the shares 30 percent higher, and 30 percent a year needs about 3.7 times today’s price.

Beside the price sit two more figures. Lumentum’s outlook for the September quarter is adjusted earnings of $4.05 to $4.35 a share, about half of last year’s total in a single quarter. And its adjusted share count for the June quarter, 101.1 million, already counts about 14.0 million shares from convertible notes that have not yet converted.

Tomorrow

Tomorrow is the verdict on all five. Dycom and Amphenol passed the bar. Lumen and Lumentum could not be measured on it. Coherent failed on cash. Friday asks whether a bar that passed two of five was set too loose, and sets the customer filings beside the industry figures to answer who is paying for this buildout.

One question to sit with until then: would you rather own the company with one customer at 26.6 percent of revenue, or the one with no customer above 10 percent, if both cost the same?

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in APH.

Amphenol is the second company this week to pass the bar, after Dycom on Tuesday. A bar that passes two companies in a row deserves a harder look, not a softer one, so today’s piece spends as long on the caveat as on the pass.

Where Amphenol sits on the road

Amphenol makes the connectors, cables and assemblies that join everything in a data center rack, and a great deal outside data centers too. Its 10-K for 2025 describes “active and passive fiber optic interconnect technologies” for the IT datacom market and lists optical fiber among its raw materials.

On 9 January 2026 it completed the purchase of CommScope’s Connectivity and Cable Solutions business for about $10.5 billion. Its closing release says the deal “adds significant fiber optic interconnect capabilities for the IT datacom and communications networks markets,” along with products “for the building infrastructure connectivity market.” Its 29 July 2026 release now expects that business to bring in $4.6 billion of sales this year, up from an earlier $4.1 billion.

From the day it closed to 30 June 2026, the CommScope business contributed $2.10 billion of sales and $190.2 million of net income, according to Amphenol’s second-quarter 10-Q.

Its 10-K names its primary competitors. Corning is on the list, alongside TE Connectivity, Molex, Belden and others.

Who pays Amphenol

This is where Amphenol differs from every other company this week. Its 10-K says: “No single customer accounted for 10% or more of the Company’s net sales.”

Its sales spread across markets. In 2025, by its 10-K, IT datacom was about 36 percent of net sales, industrial 19 percent, automotive 15 percent, communications networks 10 percent, defense 9 percent, mobile devices 6 percent and commercial aerospace 5 percent.

IT datacom, the market that includes data centers, is the largest slice. The growth is concentrated in the segment that serves it. In the second quarter, Communications Solutions sold $5.38 billion against $2.91 billion a year earlier, with 42 percent organic growth. Harsh Environment Solutions grew to $1.86 billion from $1.45 billion, and Interconnect and Sensor Systems to $1.52 billion from $1.30 billion. No single buyer carries the weight that Lumentum’s largest customer does: one unnamed customer was 26.6 percent of Lumentum’s revenue last fiscal year.

The building connectivity Friday’s Corning piece pointed to sits inside the CommScope business. Amphenol does not split that business by market in the documents I read, so how much of it goes into ordinary buildings rather than data centers is NOT PROVEN.

Which earnings, after the split

Amphenol split its shares two for one, with the new shares distributed on 2 September 2026. Every per-share figure it printed before that date is on the old basis, including all four of its last quarterly results. Halving them is my arithmetic, not a company figure.

On that basis, standard-accounting earnings over the last four quarters add up to $3.99 a share before the split, or about $2.00 after it. Amphenol’s adjusted figure is $4.31, or about $2.16. The gap is mostly the first quarter of 2026, when discrete tax items and acquisition costs cut standard earnings to $0.72 against $1.06 adjusted.

Amphenol’s filings put its first-half 2026 tax accruals at $329.0 million, including $130.0 million from China tax determinations and $160.0 million from reassessing prior years. Its adjusted figure removes discrete tax items of this kind. Standard accounting keeps them.

One more item sits inside both. The second quarter included “an $80 million ($0.04 per share) net benefit related to the recovery of IEEPA tariffs.” Its third-quarter guidance, by its own words, “does not include any additional tariff recoveries.”

Amphenol’s earnings need two corrections before they can be priced: halve them for the split, and set the tariff refund aside.

The hidden assumption, tested

The Firewall’s hidden assumption for Amphenol is that its products are so embedded in AI data centers that demand stays structurally elevated, with above-average organic growth and expanding margins for years.

Organic growth strips out acquisitions and currency. Amphenol’s releases give it quarter by quarter: 41 percent in the third quarter of 2025, then 37, then 33, then 30 percent in the quarter to 30 June 2026. Still far above average. Also slower each time.

Orders point the same way. Amphenol booked $10.7 billion of orders in the second quarter against $8.76 billion of sales, a book-to-bill of 1.23 to 1. More orders came in than went out the door.

Margins are expanding. Adjusted operating margin was 29.8 percent in the second quarter, against 25.6 percent a year earlier. Take out the $80 million tariff refund and it is about 28.9 percent, still higher.

So the assumption holds today, on both counts. “For years” is the part no filing can test yet, and the organic trend is the number that will answer it first.

The growth rate the bar uses

On Tuesday, Dycom was measured on its standard-accounting earnings growth in its last fiscal year. Amphenol gets the same treatment. Its 2025 earnings per share were $3.34, against $1.92 in 2024, a rise of 74.0 percent.

Measured over the last four quarters instead, standard-accounting earnings per share grew 59 percent and adjusted earnings 73 percent against the four quarters before. Every version of the number is high.

That growth includes acquisitions and the AI demand wave at once. Sales rose 52 percent in 2025, 38 percent of it organic, by Amphenol’s fourth-quarter release. Earnings per share grew faster than sales.

Peter Lynch, in One Up on Wall Street, described fast growers as small, aggressive companies growing 20 to 25 percent a year. Amphenol is valued at about $212.5 billion and grew earnings 74 percent. That combination is rare, and my reading is that Lynch would ask first how long it can last.

Amphenol against the bar

Test one, defensible earnings: yes, about $2.00 a share of standard-accounting earnings over the last four quarters, after the split, with the tariff refund noted. Test two, price against growth: at $86.19, the shares trade at 43.2 times those earnings. Against 74.0 percent growth, that is 0.58 times the growth rate. The bar allows 1.5. Test three, cash: free cash flow was $4.39 billion in 2025, positive.

Amphenol passes all three tests of the bar. It is classed deserves serious review, the second company this week to earn the deep dive.

The pass is real. So is the condition attached to it: the price needs growth to stay above about 29 percent a year.

The caveat, in one number

At 43.2 times earnings, Amphenol stays inside the bar as long as growth stays above 28.8 percent. Below that, it fails.

Organic sales growth was 30 percent in the latest quarter. Sales growth and earnings growth are not the same measure, but they have moved together here, and the organic line is falling toward the edge of the bar. On that measure, the same price is 1.44 times growth. At 25 percent, it would be 1.73, and Amphenol would fail.

The near term argues the other way. Amphenol’s guidance for the quarter to 30 September calls for sales of $9.3 billion to $9.4 billion, up 50 to 52 percent, and adjusted earnings per share up 51 to 53 percent on the old share basis. If that arrives, growth stays far above the line for at least one more quarter. Guidance is a forecast, and the release on 28 October will settle it.

Buffett’s first question gives the same answer from another side. At a 10 percent return, the last four quarters of earnings support about $19.95 a share. The remaining $66.24 is paid for growth still to come.

Five thousand dollars

Five thousand dollars at $86.19 buys 58.01 shares. Counting a dividend of $0.125 a quarter, a total gain of 30 percent by October 2031 needs the shares at about $109.55. Thirty percent a year needs about $317.52. The dividend is small beside the price: Amphenol declared $0.745 a share for 2025 and $0.55 for 2024, both on the old share basis, about $0.37 and $0.28 after the split.

Every $100 of Amphenol stock is backed by about $2.07 of 2025’s free cash flow. Debt was $18.81 billion at 30 June 2026, against $5.42 billion of cash and short-term investments, some of it raised to pay for CommScope, according to its 10-Q.

Amphenol reports its third quarter on 28 October, per its investor site. The organic growth rate in that release is the number this piece says to watch.

Tomorrow

Tomorrow is Lumentum and Coherent, the companies that make the light that travels down the fiber. On 2 March 2026, Nvidia put $2 billion into each of them. Their customers are far more concentrated than Amphenol’s, and Coherent’s cash flow tells a different story from its earnings.

One question to carry until then: when a company’s price only works if growth stays above a line, how close to that line would you let it get before you stopped calling it a pass?

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in LUMN or DY.

The Firewall’s hidden assumption for Lumen is that its new business grows faster than the old one shrinks. That is a sentence the filings can test, and they test it quarterly.

What Lumen still owns

Lumen sold its consumer fiber business, in 11 states, to AT&T for $5.75 billion. The sale closed on 2 February 2026, and about $4.76 billion of the proceeds went to retiring debt, according to Lumen’s filings.

What it kept is the network most people never see. Its 10-K for 2025 counts over 340,000 route miles of fiber, most of it long-haul links between cities, plus metro fiber connecting buildings. Its Investor Day release says Lumen had deployed 17 million intercity fiber miles by the end of 2025 and plans to reach approximately 58 million miles by 2031. Route miles count the paths. Fiber miles count every strand in them.

It also kept its business customers, its wholesale customers and its copper network. In the quarter to 30 June 2026, Lumen reported revenue of $2.81 billion, against $3.09 billion a year earlier. The consumer business that went to AT&T is most of the difference.

Who pays Lumen

Lumen’s second-quarter revenue, by customer: large enterprises $794 million, wholesale $653 million, the public sector $490 million, mid-sized businesses $435 million, international $72 million, and what remains of its consumer business $361 million.

Lumen sells to hyperscalers and AI companies through what it calls Private Connectivity Fabric. Its Investor Day release counts “nearly $13B in total PCF deals to date, including hyperscalers and AI leaders such as Anthropic.” Its second-quarter slides show $91 million of PCF revenue in the quarter, about 3.2 percent of the total.

Lumen’s revenue rests on ordinary businesses and government. The hyperscaler line is new, and still small.

Its 10-K adds a caution worth reading in full: “certain hyperscaler customers have developed infrastructure that has reduced their reliance on our network.”

The hidden assumption, tested

Lumen splits its business revenue into strategic, the products it is growing, and legacy, the older services it is managing down.

In the quarter to 30 June 2026, strategic revenue was $1.29 billion against $1.13 billion a year earlier, up $159 million. Legacy revenue was $1.16 billion against $1.36 billion, down $205 million. Over the first half, strategic rose $266 million and legacy fell $392 million.

The new business is growing at 14 percent. The old one is shrinking faster in dollars. Total business revenue fell from $2.49 billion to $2.44 billion.

The assumption has not been met yet. It may be. The crossover is a number to watch each quarter, and it has a date only in Lumen’s plans, not in its results.

Which earnings for Lumen

Lumen has no earnings to put a multiple on. Under standard accounting it lost $0.62, $0.00, $0.20 and $0.20 a share in its last four quarters. With no profit, there is nothing to divide the share price by.

Cash is the defensible figure, and it needs care. Lumen guides to free cash flow of $1.9 billion to $2.1 billion for 2026, excluding what it calls Special Items. Its own footnote says that guidance includes $729 million of proceeds from the AT&T sale, booked as operating cash. The same outlook reflects a $400 million tax refund, and nothing I read in the filings says whether the refund sits inside the cash figure.

So there are three honest readings: $0.36 a share of reported free cash flow in 2025, $1.01 excluding Special Items, and $1.23 for the 2026 guidance midpoint once the AT&T money comes out, each over today’s 1.03 billion shares. Against $5.05, a 10 percent return on those three figures implies $3.60, $10.10 and $12.30 a share.

Two more numbers sit beside the cash guidance in the same outlook: capital spending of $3.2 billion to $3.4 billion, and adjusted EBITDA of $3.1 billion to $3.3 billion. Capital spending about the size of operating earnings, with free cash flow still positive, means cash is arriving from somewhere besides profit. One place it can come from is customers paying ahead: Lumen carried $8.18 billion of noncurrent deferred revenue at 30 June 2026, money billed or received for service it has not yet delivered.

Pick the 2025 reported figure and Lumen looks expensive. Pick the 2026 guidance and it looks cheap. The choice is the analysis.

Can Lumen survive the fix

Lynch judged a turnaround on whether it could last long enough for the repair to work. For Lumen, the answer sits in its 10-Q maturity table.

At 30 June 2026, Lumen owed $46 million of principal for the rest of 2026, $23 million in 2027, $209 million in 2028, $675 million in 2029 and $144 million in 2030. Everything else, $12.25 billion, falls due in 2031 or later. It held $1.88 billion of cash and had $660 million available on its credit line.

Through 2030, the cash covers what comes due. The wall is real, and it is five years out. The chief executive’s purchases, flagged on Monday, came at weighted average prices of $6.35 a share in February and $6.13 in August, per her two Form 4 filings.

Lumen against the bar

Test one, defensible earnings: there is a loss, and the cash figure needs two adjustments to defend. Test two, price against growth: it cannot run, because there are no earnings to divide by. Test three, cash: Lumen’s reported free cash flow for 2025 was $371 million, positive.

At its 12 October 2026 close of $5.05, per CNBC, Lumen was valued at $5.21 billion, below the $6.35 and $6.13 its chief executive paid. Five thousand dollars buys 990 shares. A total gain of 30 percent by October 2031 needs $6.57 a share, and 30 percent a year needs $18.75.

Lumen is classed NOT PROVEN. That is not a judgment against the turnaround. It means this bar cannot measure it until there are earnings to measure.

Dycom, and the word largest

Google’s AI called Dycom “the absolute largest provider of specialized outside-plant (OSP) wireline construction and emergency restoration services in North America.” Friday’s piece promised Dycom’s own filings would have to carry that word.

They do not use it. Item 1 of Dycom’s 10-K for the year ended 31 January 2026 opens: Dycom “is a leading provider of specialty contracting services focused on the digital infrastructure, telecommunications and utilities industries throughout the United States.” In the parts of the 10-K and the latest 10-Q I read, including Item 1, the word largest does not appear, and neither does restoration.

Leading is a claim. Largest would be a ranking. The filing makes the first and not the second.

Who pays Dycom

In its fiscal year ended 31 January 2026, Dycom’s revenue came 25.4 percent from AT&T, 14.0 percent from Verizon, 10.8 percent from Lumen and 7.4 percent from Comcast. Four phone and cable companies made up 57.6 percent of its revenue. Its 10-K says plainly: “Our customer base is highly concentrated.”

That mix is shifting. In December 2025 Dycom bought Power Solutions, an electrical contractor whose revenue its deal release put at over 90 percent data centers. That new Building Systems segment was 1.7 percent of fiscal 2026 revenue, after one partial quarter, and 19.8 percent in the quarter to 1 August 2026. Dycom also says its contracted backlog for long-haul, middle-mile and data-center fiber builds “exceeds $1 billion,” out of $12.24 billion in total.

Dycom’s buyers are mostly phone companies today. Its new Building Systems business, mostly data center electrical work, is already a fifth of revenue.

Dycom against the bar

Dycom has been here before. Its fiscal 2002 10-K reports revenue down 24.5 percent to $624.0 million and a net loss of $123.0 million, after “certain segments of the telecommunications industry suffered a severe downturn.” Several customers, including WorldCom, filed for bankruptcy. By Lynch’s categories that history is a cyclical’s. Its latest year reads like a fast grower’s.

Test one, defensible earnings: standard-accounting earnings over the last four quarters add up to $10.99 a share. The growth rate is 20.7 percent, the rise in earnings per share in fiscal 2026, a year that had 53 weeks instead of 52. Test three, cash: free cash flow was $435.3 million in fiscal 2026, positive, though the first half of this year ran $54.2 million negative, as the first half did a year earlier.

At its 12 October 2026 close of $266.99, per CNBC, Dycom was valued at $8.05 billion, or 24.3 times $10.99 of trailing earnings. Against 20.7 percent growth, that is 1.17 times the growth rate. The bar allows 1.5. Five thousand dollars buys 18.73 shares. A total gain of 30 percent by October 2031 needs $347.09 a share, and 30 percent a year needs $991.32.

Dycom passes all three tests of the bar. It is the first company this week classed deserves serious review, which earns it the deep dive: the long filings work, not a purchase. The pass rests on the growth rate. Fiscal 2026 had an extra week, and in the year before it, earnings per share grew 7.5 percent. At that rate, the same price would be 3.2 times growth, and Dycom would fail. Its backlog of $12.24 billion, up 53.2 percent in a year, is the evidence for the faster rate.

Tomorrow

Tomorrow is Amphenol, the connectors and cable around the fiber. No customer reaches 10 percent of its sales, and a stock split in August changed the basis of every per-share figure it printed before, so the earnings figure has to be chosen with care before the bar can be applied.

One question to sit with before then: when a turnaround’s new business grows 14 percent and its old business shrinks faster, how many quarters would you give it before the crossover has to show up?

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in LUMN, DY, APH, LITE or COHR.

The obvious company on this road is Corning, and at its 6 October 2026 close its share price stood at 58.9 times the core earnings of the prior four quarters. The question for this week: is there anywhere else along the same road that a value investor could take seriously at today’s price?

What Friday promised, and where each promise lands

Friday’s piece on Corning made seven promises about this week. Here they are, with the day each one is kept.

• Five companies, the same five thousand dollars, and one of three classifications for each: deserves serious review, keep on watch, or NOT PROVEN. All week.

• The first question for Lumen, what it still owns after selling its consumer fiber business to AT&T. Tuesday.

• Whether Dycom’s own filings support Google’s AI calling it the largest restoration provider in North America. Tuesday.

• Coherent’s own filings, read before anything said about it. Thursday.

• The spending figures for Amazon, Alphabet and Microsoft, checked against the companies’ own documents. Today, below.

• Whether fiber demand reaches beyond the hyperscalers. Each company’s own customer disclosures, Tuesday to Thursday, and industry-wide figures on Friday.

• The closing question, which of the five comes closest to the Buffett and Lynch playbook at today’s price. Friday answers it.

What this week is looking for

This week is not looking for a stock to buy. It is looking for a company that earns a deep dive.

A deep dive is the work I do before money moves. A decade of annual reports, the proxy, the people running the business, what they paid for every share they bought back, and a price I would pay written down in advance. It takes weeks, and most companies never earn it.

So this week is a filter. Five companies go in. At most one comes out with a reason to do the long work. If none does, that is the answer.

The hidden assumption under the whole week: that somewhere along the AI fiber road, at least one company offers the buildout at a price the numbers can defend.

The story test: two clear, three on watch

The Firewall names the hidden assumption inside each company’s story, the thing that has to be true for it to hold, and classifies what the evidence supports.

Amphenol makes the connectors and cable systems that join a data center together. Its hidden assumption is that demand stays structurally elevated for years. Dycom builds and maintains fiber networks in the field, and since December 2025 also does electrical work inside data centers. Its hidden assumption is that a backlog of $12.24 billion at 1 August 2026 turns into double-digit organic growth for three to five years. Both came back clear for deeper research.

Lumen owns fiber networks in the ground and was run as a turnaround: can strategic revenue outgrow the legacy decline. Lumentum makes the lasers that put light into fiber, and Coherent makes optical transceivers. For both, the assumption is that AI demand outruns their capacity long enough to hold prices up. All three came back watch. Watch is not a fail. It means real ingredients and unresolved risks.

A story test is not a price test

The Firewall asks whether the story holds together. It does not ask what you are paying for it. That is how it is built to be used: it gets you to the right questions, and to a first decision about whether the evidence gives more reason to move forward or only to watch. The filings then answer the questions.

Its own reports say so. On Dycom: “The classification does not mean the stock is cheap or that the growth will last.” On Amphenol, it says an investor needs to separate structural moat from cyclical tailwind “before committing at a valuation that already prices in continued exceptional growth.”

A company can pass the story test and still be the wrong purchase. The price gets its own test.

The earnings you start from decide the answer

Every price test divides a price by some measure of what a business earns. Choose the wrong measure and the test answers a different question.

Lumentum shows how far apart the choices can sit. Its results release for the fiscal year ended 27 June 2026 reports a loss of $92.96 a share under standard accounting, called GAAP, the rules every US company must report under. On its own adjusted basis it earned $8.67 a share. Almost all of the gap is one non-cash charge of $7.76 billion, booked when holders of its convertible debt swapped their notes for shares.

The others need the same care. Lumen’s 2026 free cash flow guidance includes $729 million from the AT&T deal, by its own footnote. Amphenol declared a two-for-one share split in August 2026, so earlier per-share figures must be halved. Dycom changed how it calculates adjusted earnings from its fourth quarter of fiscal 2026. Coherent’s version is Thursday’s.

Before you ask whether a stock is cheap, ask which earnings you are measuring it with.

The bar, set before any test runs

A bar set after the results is not a bar. So here it is, before Tuesday.

A company is classed deserves serious review, and earns the deep dive, only if three things are true. Its earnings can be stated in a figure I can defend from its own filings. Its price-to-earnings ratio is no more than one and a half times its growth rate. And it produced positive free cash flow over its last full fiscal year.

In One Up on Wall Street, Lynch wrote that the price-to-earnings ratio of any fairly priced company will equal its growth rate. I am allowing half again above his line. That is my choice, not his. For reference, Corning scores about 3.1 against its own 19 percent growth plan.

A company that fails a test is keep on watch. Where a test cannot run, the result is NOT PROVEN, not a pass. Lumen still reports losses, so it has no price-to-earnings ratio. Lynch judged turnarounds on whether the company can survive long enough for the fix to work, and Tuesday applies that question too.

Friday’s spending figures, checked at the source

Friday’s piece set four 2026 spending plans beside the 2000 buildout and said two of the figures came from reporters, not the companies. I went to the companies.

Meta’s results release of 29 July 2026 states $130 billion to $145 billion. Microsoft’s chief financial officer, Amy Hood, said on its 29 July 2026 earnings call, in the transcript on Microsoft’s site, that the calendar 2026 figure adjusts to “approximately $175 billion” after a lease accounting change. Both confirmed.

Amazon’s and Alphabet’s second-quarter results releases give no 2026 figure. Their numbers were given on earnings calls, and I could not find a transcript on either company’s own site. They stay attributed to Fortune and Bloomberg. Amazon’s release does say its purchases of property and equipment over the trailing twelve months rose by $66.1 billion, and that the increase “primarily reflects investments in artificial intelligence.” None of the four documents mentions fiber.

Who is paying: the hyperscalers or everyone else

In 2000 I watched ordinary businesses rewire their own buildings with fiber, and that spending never made a headline. Today the hyperscalers’ budgets are counted every quarter. If ordinary companies are upgrading too, the demand under these five stocks is wide. If a handful of data center owners are the buyers, it is narrow, and every price this week carries that concentration.

The early numbers lean narrow. Cisco’s results slides for its fiscal year ended 25 July 2026 report $9.3 billion of AI infrastructure orders from hyperscalers, against $1.3 billion from neocloud, sovereign and enterprise customers combined. And a figure that looks like ordinary demand often is not: Corning’s 10-K counts hyperscale data centers inside its Enterprise Networks business.

So each day this week reads one more thing in each company’s filings: who the customers are, and how much of the revenue they account for. Friday puts those disclosures beside the industry figures and answers how much of this buildout rests on a few buyers.

How the week runs

Tuesday takes Lumen and Dycom together, the fiber in the ground and the company that builds and maintains it. Lumen was 10.8 percent of Dycom’s revenue in its last fiscal year, according to Dycom’s 10-K. Lumen’s chief executive, Kate Johnson, bought 178,685 Lumen shares with her own money in February and August 2026, according to her Form 4 filings. Lumen’s Investor Day release of 25 February 2026 says it plans to expand to approximately 58 million fiber miles by 2031. Tuesday asks what the income statement says beside both.

Wednesday is Amphenol. Thursday pairs Lumentum and Coherent, which both took $2 billion from Nvidia under 8-Ks dated 2 March 2026. Friday is the verdict against today’s bar, and the answer on who is paying.

Nothing here is a buy or sell call. The week does not put a value on Lumen’s network, and its figures come from second-quarter filings. Amphenol reports again on 28 October and Lumentum on 5 November.

Before Tuesday, put any of these five, or a company of your own, into the free Stock Story Firewall at firewall.readthelongview.com. Then hold its hidden assumption up against the price test this week runs.

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in GLW.

This was never a week about whether AI is right or wrong. It was about what happens when you take an AI’s answer and ask the questions a filing can settle. Some of the answer held up. Some of it fell apart. The most useful findings came from places neither AI looked.

The scorecard: the contracts proved out, the crown did not

The week put six questions to the filings. Here is where each one landed.

Size. NOT PROVEN. No rival on Corning’s 10-K list that reports in dollars or euros has a larger optical business. Sumitomo reports in yen on a different profit measure, and I did not read Fujikura’s results, so “largest” against all four stays unproven. Corning’s 10-K claims cost advantages relative to several competitors, not the top spot.

Contracts. PROVEN. Corning has announced seven agreements since 23 September 2025. One buyer has already put $1.0 billion down, per the 10-Q. Meta’s figure caps out at $6 billion, and AT&T’s starts above $3 billion.

Pricing power. NOT PROVEN. Corning’s optical margin rose, and so did Sumitomo’s and Prysmian’s. Sumitomo’s report credits its info communications profit to higher sales, product mix and productivity. Price is not on its list.

Capacity. A RISK. Corning has committed to more than 50 percent more US fiber capacity, and Prysmian intends to more than double its own. After the last buildout, Corning’s 2002 annual report shows fiber and cable sales falling 70 percent that year.

The one door. NOT PROVEN. Set beside the four rivals its 10-K names, Corning alone is both listed in the US and reporting optical as its own segment. Whether its price carries a premium for that cannot be measured until the other doors are priced the same way.

Price. DEMANDING. At $168.97, Corning trades at 58.9 times its past four quarters of core earnings, and 24.2 times what its own plan says it will earn in 2030.

Google and ChatGPT: the numbers held, the words did not

Monday checked the first claims one at a time. With the week finished, the full tally shows a pattern worth more than any single claim.

Most of the numbers held. Nvidia’s $500 million, AT&T’s agreement of more than $3 billion, Verizon’s 80 million fiber miles and Amazon’s agreement each match a Corning release or filing. ChatGPT’s figures held on every claim I checked, from Corning’s second-quarter growth to the $2 billion Nvidia investment in Lumentum, which Lumentum’s own 8-K exhibit of 2 March 2026 confirms.

The errors sat in the words around the numbers. Google dropped “up to” from Meta’s agreement. It said Contour packs double the fiber, where Corning’s release says significantly more. It said Nvidia’s partnership prompted Corning to expand its US optical manufacturing capacity tenfold. Corning and Nvidia’s release of 6 May 2026 commits to tenfold for connectivity and more than 50 percent for fiber. And it set the word monopoly beside a 19.5 percent market share that none of the Corning filings I read contains.

The numbers in Google’s answer mostly held. The words around them did not.

Two gaps no wording could fix. Neither answer named Fujikura or Sumitomo, two of the four rivals Corning’s 10-K lists. And neither put a price on Corning. Google’s answers I saved never raised price. ChatGPT closed by saying the question of which was the better business at today’s valuation still needed investigating. That was the question this week most needed answered, and both left it open.

The money behind the buildout, then and now

I worked inside the last buildout. Around 2000 I was installing ATM switches on fiber, and watching businesses move from token ring to Ethernet, with fiber running to switches in building closets. So the comparison I most wanted was between that spending and today’s.

A Federal Reserve Board staff paper by Rochelle Antoniewicz, a draft dated 23 April 2001, puts 2000 capital spending by publicly traded US telecom service providers at $82 billion for investment-grade firms and around $33 billion for speculative-grade firms. That is about $115 billion together.

Now today’s plans. Meta’s results release of 29 July 2026 anticipates 2026 capital spending of $130 billion to $145 billion. Amanda Gerut of Fortune reported on 30 July 2026 that Amazon expects $220 billion. Davey Alba and Julia Love of Bloomberg reported on 23 July 2026 that Alphabet raised its range to $195 billion to $205 billion. Reuters reported on 31 July 2026 that Microsoft expects $175 billion for calendar 2026, below its earlier $190 billion, partly because of a change in lease accounting.

At the midpoints, those four plans add up to about $730 billion, more than six times what those carriers spent in 2000.

Three caveats travel with that comparison. Fiber is one part of today’s spending: Meta’s release names servers, data centers and network infrastructure, and none of the four sources I read breaks out the fiber slice. The 2026 figures are plans. And four companies spending worldwide stand against one industry of US carriers, in each year’s own dollars.

With those caveats, the direction holds. This buildout is being funded at several times the scale of the last one. That supports the demand side of Corning’s story. It says nothing about the price of its shares.

Where the documents stopped me

A week that only shows the tests that worked is an advertisement. So here is every place an answer could not be found, and whose wall it was.

Corning’s wall: its filings count fiber sold in dollars. I found no count in miles, so Corning’s fiber output is NOT PROVEN. The 19.5 percent share was in none of them either.

Sumitomo’s wall: its segment profit is in yen and counted before tax. Corning’s is in dollars and after tax. Lining them up needs a dated exchange rate and one profit measure.

Amphenol’s wall: none of the three segments in its quarterly filing is fiber, so Amphenol’s fiber sales stay NOT PROVEN.

The spenders’ wall: Meta’s release and the three news reports I read do not separate fiber from servers and buildings.

My own wall: I did not read Fujikura’s results, so it is the one named rival that never testified. And Amazon’s, Alphabet’s and Microsoft’s figures came from reporters, not from the companies’ own documents. Next week opens those documents first.

On my tally, Corning gave a shareholder the most to work with: a separate optical segment, a dated growth plan, quarterly guidance, and the deposit in its 10-Q. A company that discloses more is easier to test. It is not cheaper because of it.

What Buffett’s yardstick says about the price

Buffett has long described buying a stock as buying a share of a business and what it earns. The first question that sets up, before any forecast, is what today’s earnings are worth on their own.

Over its last four quarters, Corning’s core earnings came to $2.87 a share. At $168.97, that returns 1.7 cents per dollar before any growth. A buyer who wanted 10 percent from those earnings alone would pay about $28.70. A buyer who wanted 8 percent would pay about $35.88. On free cash flow the figures are lower still.

At a 10 percent return, about $140 of Corning’s $168.97 share price pays for growth that has not happened yet.

That is not a verdict that the growth will fail. Corning’s plan has sales rising 19 percent a year to the end of 2030, and its second-quarter core earnings grew 30 percent. It is a measure of how much of the price rests on the plan. Thursday’s piece asked how much room for error you would want at this price. My own answer is more room than this price leaves.

What Lynch’s yardstick says, and what kind of company Corning is

Peter Lynch, in One Up on Wall Street, sorted companies into six kinds before judging any of them. Two matter here. A fast grower expands its earnings quickly for years. A cyclical is a company whose sales rise and fall with its industry’s cycle.

Read against his categories, Corning’s documents describe both. Its 2026 plan describes a fast grower. Its 2002 annual report, with fiber and cable sales that fell by 70 percent, describes a cyclical. The price pays for the fast grower. That is my reading, and it is the one tension in this company that no single filing resolves.

For fast growers, Lynch compared a stock’s price-to-earnings ratio with its growth rate. He treated a ratio equal to the growth rate as fair, and one twice the growth rate as a warning. At 58.9 times core earnings, Corning’s ratio is about 3.1 times its 19 percent plan and about 2.0 times its recent 30 percent pace. The plan’s 19 percent is for sales, and Corning expects earnings to grow faster, which would lower the first figure. Neither figure is near 1.

Five thousand dollars, and the verdict

The conviction test puts real money on the question. Five thousand dollars at $168.97 buys 29.59 shares. With the dividend held at $0.28 a quarter, what does Corning have to reach by 6 October 2031?

For a 30 percent gain in total, about $214. If the plan’s 19 percent pace ran one year past the end of the plan, that would be about 26 times 2031 earnings. So even a modest gain needs the plan delivered and the market still paying a high multiple five years from now.

For 30 percent a year, about $622, or about 75 times those same earnings. That needs Corning to command a higher multiple in 2031 than the 58.9 it carries today, after five years of plan growth.

Corning looks like a strong business that is a long way from the Buffett and Lynch playbook at today’s price. It belongs on the watch list.

That is a classification, not a recommendation. A lower price, or earnings that outrun the plan, would move it.

The roads beyond Corning, and how they will be judged

Following the fiber from where it is made to where it earns money gives next week its five stops.

Amphenol (NYSE: APH) makes the connectivity around the fiber, and its 29 July 2026 release puts expected CommScope sales this year near $4.6 billion. Lumentum (Nasdaq: LITE) makes laser components that create the light, with $2 billion of Nvidia’s money behind it. Coherent (NYSE: COHR) was the third company ChatGPT said it would put in a comparison, and its own filings come first when its day arrives. Dycom (NYSE: DY) installs and repairs fiber, and Google’s AI called it the largest restoration provider in North America, a claim Dycom’s own filings will have to carry. Lumen (NYSE: LUMN) owns networks already in the ground. Reuters reported that AT&T agreed to buy Lumen’s consumer fiber business for $5.75 billion, so the first question for Lumen is what it still owns.

Each gets the same five thousand dollars and one of three classifications. Deserves serious review: the price leaves a margin of safety, or growth is cheap against the multiple. Keep on watch: the business may be strong, but the price is a long way from the playbook, which is where Corning sits. NOT PROVEN: the filings do not give enough to run the tests fairly. None of the three is a buy or a sell.

The part of the buildout I could not count

In 2000 I watched ordinary businesses rewire their own buildings with fiber, switch by switch. Today the headlines count the hyperscalers’ billions. That leaves a question this week could not answer: is there a part of the AI buildout that is not being counted, in companies getting their own networks ready for AI?

The 2000 record shows why the question is worth asking. A Federal Reserve Bank of Richmond study of that boom found investment in communications equipment grew from about $62 billion a year in the first quarter of 1996 to over $135 billion a year in the fourth quarter of 2000, in 1996 dollars. The study sources that series to the Bureau of Economic Analysis and measures it against all private investment, so it reaches beyond the carriers.

Next week goes looking. Amphenol’s building connectivity, Dycom’s customers, and five companies drawn at random from the S&P 500 will show whether ordinary businesses are spending on their networks, and whether their filings say so. If none of them does, that is part of the answer.

The hidden assumption, clause by clause

The hidden assumption this week was that Corning’s size, contracts and technology turn into AI fiber revenue at margins its competitors cannot match, at a price that has not already paid for all of it.

An assumption with four clauses is only as strong as its weakest one. The contracts clause held. The size clause held against every rival I could measure. The margin clause, the one that would make Corning special rather than busy, is the one no filing I read could settle. And the price clause runs the wrong way: at 24.2 times the earnings Corning plans for 2030, much of that plan is already in the price.

Google’s AI was right that Corning matters to AI fiber. It could not tell me what Corning is worth, or what I would be paying for it. Those answers came from asking the filings a better question.

That is the idea behind my book, AI Won’t Make You a Better Investor: But the right question will. It is on Amazon at https://a.co/d/0fQzCk4V.

The question to carry into next week: of Lumen, Amphenol, Lumentum, Coherent and Dycom, which do you expect to come closest to the Buffett and Lynch playbook at today’s price?

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in GLW.

Monday promised that this piece would set the price against what the business has delivered, and ask whether Corning carries a premium for being the one clean US door into AI fiber. Wednesday found its margin rising alongside its rivals’. Today is the arithmetic, and every step can be redone with the documents named.

The price, and where it comes from

A share price with no document behind it is just a number on a screen, so this one comes with its source. CNBC shows Corning closing at $168.97 on 6 October 2026, with 861.39 million shares outstanding, which puts its market value at $145.55 billion.

There is one price in a filing to set beside it. Corning’s Form 8-K dated 6 May 2026 gives Nvidia a warrant to buy up to 15 million Corning shares at $180.00 each. On 6 October the market price sat about 6.5 percent below that strike. Nvidia holds a right to buy Corning that is worth using only if the shares rise.

That says nothing about whether Corning is cheap. It does say the market is not currently paying more than the price Nvidia agreed in May.

Core earnings, and why the word matters

Corning reports two kinds of earnings. Its results releases give earnings under standard accounting, and a measure it calls core earnings, which leaves out items the company treats as outside its underlying performance. Each release reconciles the two.

The gap is large. For 2025, Corning’s fourth-quarter results release reports core earnings of $2.52 a share and earnings under standard accounting of $1.83. In the third quarter of 2025 the pair was $0.67 and $0.50. In the fourth quarter it was $0.72 and $0.62.

The same share price gives two different answers. At $168.97, Corning trades at 67.1 times its 2025 core earnings, and at 92.3 times its 2025 earnings under standard accounting. Both figures are true.

The word core turns a stock at 92 times earnings into a stock at 67 times. Know which number you are paying for.

The rest of this piece uses core earnings, because that is the measure Corning’s own plan is built on. Read every multiple below as the lower of the two.

Three more ways to measure the price

Earnings are one lens. There are three others, each from Corning’s own figures.

Free cash flow is the cash left after the company pays for its plants and equipment. Corning’s fourth-quarter release reports adjusted free cash flow, its own measure, of $1.72 billion for 2025. Against a market value of $145.55 billion, that is $1.18 of free cash flow for every $100 of stock.

The second-quarter figure for 2026 looks much stronger, at $1.42 billion in one quarter. As Tuesday’s piece showed, $880 million of that quarter’s operating cash came from customer deposits and government incentives, so it is not a run rate.

Book value is what shareholders would have if the company’s assets were sold at the values on its balance sheet and its debts paid. Corning’s second-quarter release shows shareholders’ equity of $12.56 billion. The market values the company at 11.6 times that.

The dividend is the last lens, and the smallest. CNBC shows an annual dividend of $1.12 a share, a yield of 0.66 percent at the 6 October close. An investor buying Corning today is buying growth, not income.

What the price requires

A trailing multiple looks backward. The more useful question looks forward: what growth makes today’s price reasonable?

Corning’s own second-quarter release supplies both inputs. It guides third-quarter 2026 core earnings to between $0.85 and $0.89 a share. And it says the company expects to deliver sales growth of 19 percent a year from the fourth quarter of 2026 to the fourth quarter of 2030, with earnings growing faster than sales. Both are management forecasts, and this piece treats them as forecasts.

Take the midpoint of the guidance, $0.87, and multiply by four. That gives a starting rate of $3.48 a year. Grow it at 19 percent a year for four years, the period of Corning’s plan, and core earnings reach about $6.98 a share by the end of 2030.

At $168.97, that is 24.2 times 2030 earnings.

If Corning hits its own plan exactly, the price you pay today is still 24 times what it earns four years from now.

That is the sense in which the price assumes the plan works. The plan is not a stretch the price leaves room for. It is roughly what the price already counts on.

If growth comes in faster, or slower

The same arithmetic, run at other growth rates, shows how much depends on the plan.

At no growth from the $3.48 starting rate, $168.97 is 48.6 times earnings in 2030. At 10 percent a year it is 33.2 times. At Corning’s 19 percent it is 24.2 times. At 25 percent a year, faster than the plan, it is 19.9 times.

Corning’s plan says earnings will grow faster than sales, which would bring those multiples down. That is a forecast too, and the table does not count on it. If you believe earnings will outrun sales, run the same steps with a higher rate. One more input cuts the other way. Nvidia’s two warrants cover up to 18 million Corning shares, and any shares issued under them spread the same earnings across more of the company.

What Corning has delivered, and what history can tell us

Start with what Corning has delivered so far. Its fourth-quarter release reports core earnings up 29 percent in 2025, and says adjusted free cash flow nearly doubled between 2023 and 2025. Its second-quarter release reports core earnings up 30 percent in the quarter and Optical Communications net income up 77 percent.

The build is visible too. Corning has committed to lifting its American fiber capacity by more than half and its American connectivity capacity tenfold, while Prysmian intends to take its own US fiber capacity past double. On the demand side, Verizon’s agreement alone runs to over 80 million fiber miles between 2027 and 2032, counting every strand. One figure is missing. The Corning filings I read report fiber in dollars, not miles, so how much it has produced is NOT PROVEN.

History is how I check what I am looking at. I do not use it as a forecast. I use it to see whether today resembles an earlier period, and whether the signs that once warned investors are showing up again.

On demand, 2000 turns out to be a weak guide. Corning’s 2001 annual report blamed that slump on long-haul telecom carriers buying less. Today’s buyers are hyperscalers wiring data centers, and one of them paid $1.0 billion up front. Between them, Meta, Amazon, Alphabet and Microsoft expect to spend about $730 billion on capital projects this year, against roughly $115 billion spent by US telecom carriers in 2000, and Friday’s piece sets those figures out in full.

On supply, 2000 still carries a warning. Corning’s 2002 annual report named pricing pressure alongside the fall in volume. Corning and Prysmian are both adding US capacity now. That is the part of the comparison worth keeping.

The door premium, and why it cannot be measured yet

Monday asked about a premium for Corning’s position as the easiest US way into fiber. On today’s evidence, the answer is NOT PROVEN.

A premium is a comparison. It means Corning’s price is higher than a similar business would command without the scarcity. Showing it needs the other doors measured the same way: Amphenol, Lumentum and the rest, on the same earnings, cash flow and growth arithmetic. That is next week’s work, and it runs on exactly the method used here.

What today’s numbers can say is narrower. At 58.9 times trailing core earnings and 24.2 times its own 2030 plan, Corning is priced as a company expected to keep growing quickly for years. Whether some of that price is scarcity rather than growth cannot be separated from one company alone.

Margin of safety

A margin of safety is the gap between the price you pay and a conservative estimate of what a business is worth. It is what protects an investor when the forecast is wrong.

At today’s price, Corning’s margin of safety rests on its plan arriving on time. The plan does not have to be beaten for the price to hold up. It does have to be met. A slower year, a pause in data center building, or rivals’ new capacity arriving at once would each push the 2030 multiple up from 24.2, not down.

My reading of the numbers, and it is a reading, is that the price leaves little room for error.

The hidden assumption, priced

The hidden assumption this week is that Corning’s size, contracts and technology turn into AI fiber revenue at margins its competitors cannot match, at a price that has not already paid for all of it.

Monday found the size is real and the crown is unproven. Tuesday found the agreements are real, with one customer paying a $1.0 billion deposit. Wednesday found the margin is shared. Today’s arithmetic says the price has paid for most of the plan already. On the numbers in Corning’s own filings, the price sits above what its delivered earnings support and close to what its plan supports.

That is a classification, not advice. A buyer who expects Corning to beat its plan can see room here. A buyer who wants a margin of safety will find it thin.

Tomorrow

Tomorrow brings the week together: the scorecard on every test, what Google’s AI and ChatGPT got right and wrong, and the roads beyond Corning that an American investor can take, from Lumen’s network to Amphenol, Lumentum, Coherent and Dycom.

The question to carry until then: at 24 times its own 2030 plan, how much room for error would you want before paying for a fiber company, knowing what happened after 2000?

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in GLW.

Tuesday’s piece ended on the question the contracts could not answer: what can Corning charge? Today Corning’s rivals testify, from their own reports, on price and on capacity. Then Corning’s annual report for 2002 testifies on what happened the last time the whole industry built at once.

What pricing power would look like

Pricing power is the ability to raise prices without losing customers. It is the thing a value investor most wants to find, because it is what lets a company keep its margin when costs rise or competitors arrive.

Filings can show it, but only in a particular shape. The company names price as a reason its profit grew, and its margin rises while its competitors’ margins do not. A margin that rises alongside everyone else’s shows a good market. It does not show that the good market belongs to one company.

A wide margin is evidence of good times. Pricing power is evidence that the good times belong to you.

Companies do name price in their filings when it moves. Corning’s own annual report for 2002 named pricing pressure when prices were falling, as the history below shows. So a filing that explains a rising margin without naming price is telling you something, even if it is not telling you everything.

So the test today has two parts: what each company says drove its profit, and whether Corning’s margin moved differently from its rivals’.

Corning’s own explanation

Corning’s results release for the quarter ended 30 June 2026 reports Optical Communications net income of $438 million on sales of $2.07 billion, against $247 million on $1.57 billion a year earlier. That is 21.1 cents per dollar of sales, up from 15.8, computed from the release.

Corning’s 10-Q for that quarter says the segment’s net income rose by $191 million, the main driver of the company’s growth in core net income. Its 10-Q for the quarter ended 31 March 2026 explains the segment’s sales growth by demand: its Enterprise business, driven by generative AI products, and its Carrier business, driven by data center interconnect and fiber-to-the-home products.

In the passages I read, Corning explains its growth by demand and does not name price. That leaves Corning’s pricing power NOT PROVEN from its filings. It does not mean Corning lacks it.

Sumitomo testifies

Sumitomo Electric’s financial section for the fiscal year ended 31 March 2026 is the clearest witness, because it explains each business separately.

Its info communications segment, which includes optical fiber cables and the devices that connect them, grew sales by 46.3 percent to ¥326.6 billion, on data center demand. The line for optical fiber cables, telecommunication cables and fusion splicers rose to ¥162.5 billion from ¥89.0 billion, an increase of 83 percent. Operating profit rose to ¥77.4 billion from ¥19.9 billion, and the operating margin improved by 14.8 points to 23.7 percent.

The report credits that profit to higher sales, an improved product mix and productivity. For the group as a whole it lists improved selling prices among several reasons, but the segment explanation does not. On the business that competes with Corning, Sumitomo’s own report credits selling more and selling better products. It does not list price.

Prysmian and Fujikura

Prysmian’s results release of 30 July 2026 reports an adjusted margin of 23.8 percent for its Digital Solutions business in the second quarter, up from 16.8 percent a year earlier, on sales of €540 million against €387 million. Adjusted EBITDA, the measure Prysmian uses, is profit before interest, tax, depreciation and some one-off items. I have Prysmian’s margin. I have not found its explanation, so what drove it is NOT PROVEN.

Fujikura is the gap I could not close. I was not able to open Fujikura’s own forecast revision or results summary. The summaries of them that I found disagree with each other about price, so I am printing neither. Fujikura’s pricing is NOT PROVEN.

Every margin went up

Put the three together and the shape is plain. Corning’s optical margin rose by 5.3 points. Sumitomo’s info communications margin rose by 14.8. Prysmian’s Digital Solutions margin rose by 7.0.

These are three different measures: Corning’s is segment profit after tax, Sumitomo’s is operating profit before tax, and Prysmian’s is adjusted EBITDA. The levels cannot be ranked against each other. The direction can be compared, and everyone points the same way.

When every producer’s margin rises at once, the first suspect is the market, not the moat.

That is my reading of three filings side by side. It does not prove Corning lacks pricing power. It means the margin alone cannot prove that it has it, because its rivals’ margins show the same thing.

Everyone is building

The second witness question is capacity, and every document I opened says the same thing: more.

Corning’s second-quarter results release shows capital spending of $754 million in the first half of 2026, against $516 million in the first half of 2025, an increase of 46 percent. Its 6 May 2026 release with Nvidia commits Corning to a tenfold increase in US connectivity capacity, more than half again as much US fiber capacity, and three new plants. Its Meta release includes a new cable plant in Hickory, North Carolina.

Sumitomo’s financial section shows additions to plant in its info communications segment of ¥25.0 billion in its last fiscal year, against ¥19.1 billion the year before. The segment’s assets rose to ¥367.4 billion from ¥284.6 billion, an increase of 29 percent. Its president’s message says the company will strengthen production capacity for optical cables, connectors and optical devices for data centers.

Prysmian is building in the United States. Its release of 20 July 2026 announces an agreement of up to €5.5 billion with Molex, part of the large private American company Koch, for optical cables used inside data centers. It runs for up to ten years and includes a €550 million upfront payment. Prysmian says it will more than double its US fiber capacity, including the stage where the glass is made, and Molex’s own release puts the new US jobs at about 600.

What Corning wrote in 2000

Corning has written this paragraph before.

Its annual report for 2000, filed with the SEC, says sales in its Telecommunications segment grew 73 percent that year, driven by demand for premium fiber, cable and photonic products and by acquisitions. It says Corning continued to produce at maximum manufacturing capacity and invested capital to expand further. It lists the fiber capacity expansions announced that year, one of which would increase a fiber plant’s capacity by more than 50 percent.

The margin was wide then too. Corning’s annual report for 2001 shows Telecommunications segment net income at 13.4 percent of segment sales in 2000. In 2001 the segment lost money, at 1.8 percent of sales. That segment was defined differently from today’s Optical Communications, so its levels cannot be set against the 21.1 cents above. The direction is the lesson. A wide margin in a boom year said nothing about the year that followed.

The 2000 report also notes that in early 2001 several customers announced their order rate might be lower than expected.

What Corning wrote in 2002

Corning’s annual report for 2002 gives the outcome in its own table. Optical fiber and cable sales were $2.88 billion in 2000 and $2.89 billion in 2001. In 2002 they were $859 million.

The report says those sales fell 70 percent in 2002, primarily on a decline in volume, and it names pricing pressure in 2002. The whole Telecommunications segment, which also made photonic components, fell to $1.63 billion in 2002 from $5.19 billion in 2000. The Telecommunications segment lost $1.9 billion that year, including restructuring and impairment charges.

The capacity went into reverse as well. Corning’s 2001 annual report says it would permanently close its optical fiber operations in Deeside, North Wales, and its 2002 report refers to construction projects in the optical fiber business that had been stopped in 2001.

The technology was right in 2000. By 2002 the volume and the price were both wrong.

What is different now, and what is not

Some things are different. Corning’s 2001 annual report tied the slump to falling long-haul demand from telecommunications carriers. Today’s largest buyers are hyperscale cloud companies, and Tuesday’s piece showed one of them paid a $1.0 billion deposit under an agreement running through 31 December 2029. I found nothing like that deposit in the 2000 report, though I did not read every page of it. Corning is also not the only producer being paid in advance now: Prysmian’s Molex agreement carries a €550 million upfront payment.

One question these filings cannot settle is where the value in an AI cluster ends up. Light has to be made before it travels down glass, and the laser components that make it come from companies such as Lumentum (Nasdaq: LITE). A Form 8-K exhibit Lumentum filed on 2 March 2026 says Nvidia is investing $2 billion in it, alongside a multibillion-dollar purchase commitment for laser components. That is four times the $500 million Nvidia paid Corning. It does not make one company the better investment. It does show Nvidia committed more cash to the light source than to the glass, and Friday follows that road.

Some things are the same. Producers are adding capacity at the same time, and margins are rising across the industry at the same time. Corning’s 2000 report describes both.

Neither the filings nor I can say when, or whether, demand slows. Uncertain timing is not the same as unlikely. What the filings do show is that the hidden assumption this week, that Corning earns margins its competitors cannot match, is NOT PROVEN. The margins its competitors report look much like its own.

Tomorrow

Corning’s position is strong and its margin is rising. Its rivals report the same. There is one more fact in Prysmian’s release worth carrying into tomorrow. Prysmian describes itself as one of just three US manufacturers of fiber and optical cable. Corning is another, and of those two, only Corning trades on a US stock exchange. Amphenol (NYSE: APH) is the other US-listed company on Corning’s list of rivals, through its CommScope connectivity and cable unit, but it reports no fiber segment of its own.

Tomorrow the question becomes the price: what today’s share price assumes about growth and margin, and whether being the one clean US door into AI fiber has been priced in already.

Reply to this post with the question you most want answered about Corning’s price. Replies come straight to me, and I read them all.

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in GLW.

Monday’s piece set out who Corning competes with. Today reads what its customers have agreed to, document by document, and asks the question each announcement leaves open: what does it oblige anyone to buy?

What “up to” means

Corning’s release of 27 January 2026 describes a multiyear, up to $6 billion commitment from Meta for optical fiber, cable and connectivity. Meta is the anchor customer for a capacity expansion at Corning’s cable plant in Hickory, North Carolina, and the release adds a new cable facility there.

“Up to” sets a ceiling. It tells you the most the agreement could be worth. It does not tell you the least. The release gives no minimum purchase, no price per unit and no end date, so from the release alone the floor is NOT PROVEN.

A ceiling with an anchor customer and a new plant behind it still tells you something. Corning is building capacity around this customer. What the ceiling cannot tell you is how much revenue arrives, or when.

Seven agreements, read the same way

Here are seven fiber agreements Corning has announced since September 2025, each read for what it puts in dollars.

Microsoft, 23 September 2025: a manufacturing collaboration to produce Microsoft’s hollow-core fiber, which carries light through air instead of solid glass, in Corning’s North Carolina plants. Microsoft’s blog of that date names Corning and Heraeus as its first manufacturing collaborators for that fiber. No figure is given. Meta, 27 January 2026: up to $6 billion. Two hyperscalers, disclosed in Corning’s first-quarter results release of 28 April 2026: no names and no figures. Nvidia, 6 May 2026: a partnership under which Corning will increase its US optical connectivity capacity tenfold and its US fiber capacity by more than 50 percent, plus warrants Nvidia bought for $500 million. Amazon, 8 June 2026: a multiyear, multibillion-dollar supply agreement. Verizon, 8 September 2026: a multi-billion dollar agreement for more than 80 million fiber miles from 2027 to 2032. AT&T, 29 September 2026: a multi-year agreement valued at more than $3 billion for fiber and cable.

“Multibillion” is a size word. It rules out a small deal and nothing else.

Of the seven, two put a dollar figure on what the customer may buy, and they point in opposite directions. Meta’s is a ceiling: up to $6 billion. AT&T’s is a floor: more than $3 billion. A ceiling tells you the most a deal can be worth. A floor tells you the least. Nvidia’s $500 million is a different kind of number, and it gets its own section below.

Two of the seven cannot be checked from the customer’s side. Corning’s first-quarter results release says two additional hyperscale customers entered into large, long-term agreements, similar in size and duration to the Meta agreement, and it does not name them. Their existence comes from Corning’s filing with the SEC, and that counts. Who signed them, and on what terms, is NOT PROVEN.

Nvidia: $500 million paid, $2.7 billion optional

Corning’s Form 8-K dated 6 May 2026 describes two warrants sold to Nvidia. A warrant is a right to buy shares later at a fixed price. The first lets Nvidia buy up to 15 million shares at $180.00 each. The second, called pre-funded because the price was paid up front, lets it buy up to 3 million shares for $0.0001 each.

Nvidia paid $500 million for the two together, and Corning’s second-quarter cash flow statement records that sum as proceeds from warrants. The rest is optional. Buying all 15 million shares at $180.00 would cost Nvidia $2.7 billion, and only if it chooses to. The 8-K says that warrant expires on the third anniversary of its issue, or earlier if the partnership agreement ends.

One technology news site, AI Business, reported that Nvidia “invested $3.2B in Corning.” That figure adds an option to a payment.

Nvidia has committed $500 million in cash and holds the right to spend $2.7 billion more.

Amazon and Verizon: size words, no figures

Corning’s second-quarter results release describes Amazon’s agreement as multiyear and multibillion-dollar, covering fiber, cable and connectivity for Amazon’s US data centers. AWS chief executive Matt Garman said it would create 1,000 new jobs at Corning’s facilities, according to Optics & Photonics News.

Verizon’s agreement is the only one with a volume and a calendar: more than 80 million fiber miles from 2027 to 2032. The release I read gives no dollar figure beyond “multi-billion.” Its headline pairs broadband expansion with AI infrastructure, so not all of that fiber is for data centers.

A volume and a calendar make Verizon’s agreement the easiest to track. Between 2027 and 2032, Verizon’s own reports on its fiber buildout will show whether the miles are being laid.

Where the binding money shows up

Announcements describe intentions. The 10-Q shows cash.

Corning’s Form 10-Q for the quarter ended 30 June 2026 says customer deposits let customers secure rights to products under long-term supply agreements, typically for up to 10 years. Its 10-Q for the quarter ended 31 March 2026 adds that most of these deposits were non-refundable. A customer that pays a non-refundable deposit has put money down, rather than a signature alone.

In the second quarter, the 10-Q records a $1.0 billion customer deposit under a long-term supply agreement extending through 31 December 2029. Corning’s cash flow statement shows customer deposits and government incentives adding $880 million to operating cash flow in that quarter, against $59 million in the second quarter of 2025.

Contract liabilities, the money Corning holds for products it has not yet delivered, stood at $2.7 billion on 30 June 2026, against $2.3 billion on 31 December 2025. Not all of that is fiber. Corning’s first-quarter 10-Q put customer deposits at about $1.3 billion and deferred revenue at about $737 million on 31 March 2026, and said the deferred revenue relates primarily to Hemlock Semiconductor, a business in its Solar segment. The $2.7 billion is a company-wide figure, and the fiber share of it is NOT PROVEN.

An announcement tells you what a customer might buy. A deposit tells you what it paid to hold its place.

A billion against two years of deposits

The size of that one deposit matters more when you set it beside what came before.

Corning’s 10-Q for the quarter ended 30 June 2024 put total customer deposits at about $1.2 billion. Its 10-Q for the quarter ended 30 June 2025 put them at about $1.1 billion. Its first-quarter 10-Q for 2026 put them at about $1.5 billion on 31 December 2025 and about $1.3 billion on 31 March 2026.

Then, in the quarter ended 30 June 2026, one customer deposited $1.0 billion. One agreement, in one quarter, added close to as much as Corning’s entire deposit balance at 30 June 2025.

Deposits also get used up. The second-quarter 10-Q says Corning applied $214 million of earlier customer deposits to customer bills in the first half of 2026, which is what happens as products ship. A deposit is a prepayment, and it becomes revenue only as products are delivered.

Who paid the billion

The 10-Q does not name the customer. It does say the $1.0 billion deposit was partly offset by $296 million for a warrant issued to that same customer, which Corning records as consideration payable to a customer. In plain terms, the accounting treats the warrant’s value as a discount given to the buyer.

The only warrants I found Corning issuing in that quarter are the two it sold to Nvidia, described in the 6 May 8-K. My reading is that the deposit came from Nvidia. That is a reading of two filings side by side, not a statement either filing makes.

What a deposit proves, and what it does not

A deposit proves that one customer committed real money for a fixed period. In what I read, it is the strongest evidence of commitment Corning has published this year.

It does not prove the price. The filings I read do not show a price per unit in any of these agreements, and whether the warrant bought better terms for the customer is not something they say. So whether Corning is charging more, less or the same as before is NOT PROVEN from the contracts.

That is the part the contracts cannot settle, and it is where the rivals come in.

The hidden assumption, after the contracts

The hidden assumption this week is that Corning’s size, contracts and technology turn into AI fiber revenue at margins its competitors cannot match, at a price that has not already paid for all of it.

Today tested the middle word. The contracts are real, and one of them is backed by $1.0 billion of cash. Most of what has been announced, though, is ceilings, size words and one floor, and none of it shows a margin. The contracts support the revenue half of the assumption. They say nothing about the margin half.

Tomorrow

Contracts tell you who wants Corning’s fiber. They do not tell you what Corning can charge for it. Tomorrow the rivals testify on that: Sumitomo and Fujikura on pricing, all of them on capacity, and what happened to fiber suppliers the last time the industry built at once.

Reply to this post with a company whose contracts you would like to read this way. I read every reply.

Not investment advice. The subscriber decides.

Disclosure: the author holds no position in GLW.

I asked it that way on purpose. I wanted to see whether an AI asked for a leader would hand me one. It did. I then asked who trains, installs and repairs the most fiber, and got a leader for each job: Corning for training, AFL for installation, Dycom for repairs, each described as the largest or the absolute largest.

Three questions shaped like rankings produced three rankings. That is where the filings come in.

Why fiber, from someone who has installed it

I have worked in technology for 35 years. For part of that I was a network engineer, and around 2000 I worked on large Y2K projects, installing ATM switches that ran on a great deal of fiber.

When I started researching AI hardware, the racks and the equipment inside them, I kept coming back to fiber. I have not seen a technology that replaces what it does, though something may come along, and I will say so if it does. AI agents pass more and more data between machines, and in my experience that data travels over glass.

The 2000 buildout taught me one thing that matters more than the technology.

In 2000 the fiber was the right technology. The question was who got paid for it, and at what price.

That question is still open, and this week tries to answer it for Corning.

The internet hierarchy

I call it the internet hierarchy. Ask the internet who the best company in a field is and you get a ranking with one name on top. In my experience, that name is usually the one that gets talked about most.

The company on top is often a good business. A good business can still be the wrong purchase, because the internet and a value investor are asking different questions.

The internet hierarchy tells you which company is best. A value investor asks which company is worth owning at the price.

I use the hierarchy as a starting point. I take the leader it names, open that company’s annual report, find the companies it says it competes with, and look hardest at the ones the internet skipped. That is how I found Dynatrace, by going past the name the internet put on top in that field, Datadog.

What Google and ChatGPT said

Google’s answer opened with the crown and a number: Corning, it said, holds “roughly 19.5%” of the global market.

Those two claims do not fit together. A company with about a fifth of a market is not a monopoly, by the answer’s own figure. I did not find the 19.5 percent in any Corning filing I read, so it is NOT PROVEN.

ChatGPT restated my question as which publicly traded fiber or optics company is best positioned to benefit structurally from the AI buildout. Its answer was hedged. Corning was its current pick, and it ranked Corning beside Lumentum, Coherent and Broadcom, companies that make lasers, transceivers and switching chips.

The two answers describe two different industries. Google’s question was about who supplies the glass. ChatGPT’s was about who benefits from moving light through it. Each answered the question it was given.

What they got right, and what they stretched

Start with what they got right. Amazon did sign a multiyear, multibillion-dollar agreement for Corning’s fiber, cable and connectivity, according to Corning’s results release for the second quarter of 2026. ChatGPT’s growth figures for that quarter match the release: optical sales up 32 percent, the enterprise networks business up 65 percent. And Google was right that Nvidia put $500 million into Corning. The money appears as proceeds from warrants in Corning’s second-quarter cash flow statement, and Corning’s Form 8-K dated 6 May 2026 names Nvidia as the buyer of those warrants.

Other claims went further than the documents. Google called Meta’s agreement a $6 billion deal. Corning’s release calls it up to $6 billion, which is a ceiling, not a sum. Google said Corning’s Contour cable packs double the fiber strands into the same diameter. Corning’s release of 8 September 2026 with Verizon says the cable lets Verizon deploy significantly more fiber within its conduit. It does not say double.

Google also listed CommScope as a fiber supplier. Amphenol’s Form 10-Q for the quarter ended 31 March 2026 says Amphenol completed its purchase of CommScope’s connectivity and cable business on 9 January 2026.

The list neither one used

Corning’s annual report on Form 10-K for the fiscal year ended 31 December 2025 has a section on competition. It names Corning’s principal competitors in optical communications: Amphenol, Fujikura and its subsidiary America Fujikura Ltd., Sumitomo, and Prysmian.

Google’s answer named one of those four as a fiber supplier, Prysmian. ChatGPT’s named none of them. Neither answer listed Fujikura or Sumitomo as a fiber competitor. Google mentioned Fujikura only in a later answer, as the owner of the installer AFL.

In the passage I read, Corning does not claim the top spot. It says its manufacturing scale, fiber process, technology and intellectual property give it cost advantages relative to several of its competitors. Several is a smaller word than monopoly.

The AI tells you who gets talked about. The 10-K tells you who the company says it competes with.

One door

Next I looked up where each of those four trades.

Amphenol (NYSE: APH) is listed in New York. Prysmian trades in Milan under PRY, Sumitomo Electric in Tokyo under 5802 and Fujikura in Tokyo under 5803. In the SEC filings and market listings I searched, none of those three is listed on a US stock exchange. Each has American depositary receipts, issued by US banks and registered with the SEC on Form F-6. They trade over the counter: Sumitomo’s as SMTOY, Prysmian’s as PRYMY and Fujikura’s as FJIKY.

Amphenol is easy to buy, but it does not report fiber as a segment. Its Form 10-Q splits the company into Communications Solutions, Harsh Environment Solutions, and Interconnect and Sensor Systems, with $16.38 billion of sales across them in the first half of 2026. Its results release of 29 July 2026 says it expects the CommScope connectivity and cable unit to sell about $4.6 billion this year. I did not find a separate fiber sales figure in the release or the 10-Q, so how much of Amphenol is fiber is NOT PROVEN.

Among Corning and the four rivals its 10-K names, Corning is the only one that trades on a US exchange and reports optical communications as a segment of its own, and its largest. Smaller US-listed fiber companies exist outside that list. This week stays with the list Corning wrote.

When there is only one door, everyone walks through it, and the door gets expensive.

That line is my reasoning, not a finding. Thursday tests it against the price.

Corning, on its own terms

Corning’s results release for the quarter ended 30 June 2026 reports Optical Communications sales of $2.07 billion, against $1.57 billion a year earlier, an increase of 32 percent. The segment’s net income was $438 million, against $247 million.

Optical is the largest part of the company. Corning’s five segments together reported $4.74 billion of sales in the quarter, and optical was 43.7 percent of that.

Corning reports its segments using its currency hedge rates, so the five segments add up to its $4.74 billion of core sales rather than the $4.51 billion reported under standard accounting. Every segment figure in this piece is on that basis.

The rivals are witnesses, not picks

You do not have to own a company to learn from its filings. Corning’s rivals report on the same market it sells into.

Sumitomo Electric’s annual report for the fiscal year ended 31 March 2026 says operating profit in its infocommunications segment rose by ¥57.5 billion to ¥77.4 billion, on data center demand. That is close to four times the prior year’s ¥19.9 billion. Corning is not the only company reporting that demand.

Prysmian’s results release for the six months to 30 June 2026 reports second-quarter Digital Solutions sales of €540 million, against €387 million a year earlier, in a segment Prysmian defines more broadly than fiber. That is growth of about 40 percent, faster than Corning’s 32. I have not yet read Fujikura’s own annual results summary, so its figures stay out.

Among the rivals that report in dollars or euros, Corning’s optical segment is the largest. Sumitomo reports in yen and before tax, where Corning reports segment profit after tax, so the two cannot be lined up without a dated exchange rate and the same measure. Against all four, “largest” is NOT PROVEN. On Wednesday these filings testify on a more useful question: who holds the pricing power.

The hidden assumption, and the rest of the week

An investment story rests on something that has to be true. I built the Stock Story Firewall to name that thing before anything else gets tested. For Corning, the hidden assumption is this: Corning’s size, contracts and technology turn into AI fiber revenue at margins its competitors cannot match, at a price that has not already paid for all of it.

The week tests that in order. Tuesday reads the contracts, and what “up to” means. Wednesday calls the rivals as witnesses on pricing power and capacity, and looks for what happened to fiber suppliers after the last buildout. Thursday sets the price against what the business has delivered, and asks whether Corning carries a premium for being the only door. Friday scores all of it and follows the road to companies an American investor can own that deserve a week of their own.

The week will not tell you to buy or sell anything, and it will not value Corning’s rivals. Where a document does not answer a question, the answer is NOT PROVEN, which is a verdict, not a gap.

Tomorrow

Today was about who competes. Size is not commitment, and an answer that drops two words from a contract turns a ceiling into a fact. Tomorrow reads the agreements Corning has announced and asks what each one obliges anyone to buy.

Run Corning, or any company you are weighing, through the free Stock Story Firewall at firewall.readthelongview.com before then.

Not investment advice. The subscriber decides.

Disclosure: the author holds a small position in Rigetti and no position in IonQ. This is analysis, not a recommendation.

The problem with the obvious numbers

Start with why any of this is necessary.

The figures quoted at a self-directed investor all day share one property. Each of them is capable of moving in the right direction while ownership of the company gets worse. A revenue line triples and the count of shares underneath it rises. An outlook is lifted and the thing being counted quietly widens. A performance record lands on the smaller machine. An award is announced and the stock that paid for it was issued first.

None of that is deception. It is what happens when a headline reports an event and a filing reports the cost. The event is exciting and arrives first. The cost is boring and arrives a quarter later, one line below where most people stop reading.

That is the whole edge available here, and it is not about being cleverer than anyone. The cover page, the working capital line, the revenue footnote and the utilization assumption are public, free, and take an afternoon.

Whether other people read them is not something I can know, and I am not going to pretend otherwise. What I can say is that I had not. Every number that changed my mind this week was sitting in a document I already had access to and had not opened.

The screen, published as a tool

Five characteristics that earn patience, five that end it. Everyone is markable as present, absent or not yet disclosed, and not yet disclosed is a real answer.

Earning patience: unit economics improving with scale, growth that funds itself, demand arriving before capacity, revenue that repeats, and progress costing less ownership over time.

Ending it: sales spending outrunning revenue with no leverage, one-time revenue dressed as a business, backlog that will not convert, a headline metric that changed definition, and a thesis that depends on the next funding event.

Numbers one, two and five are the load-bearing ones, because they decide whether time works for the owner or against. A company where time is an asset can survive being early. A company where time is a liability has to be right on schedule.

Rigetti, marked

Unit economics improving with scale: not yet disclosed. The company does not publish what a system costs to build.

Growth that funds itself: absent. The loss from operations was $54.0 million in the first half, and capital spending was $16.4 million on top.

Demand before capacity: partly present. Systems are sold to identified buyers rather than built into inventory, but capital spending is rising faster than the revenue those sales produce.

Revenue that repeats: not yet disclosed. No repeat purchases are reported, and an $8.4 million order from C-DAC expected later this year would be a first sale, not a second.

Progress costing less ownership: absent, and moving the wrong way. The company issued 7.74 million shares to the Commerce Department in September, about 2.3 percent, before the research was done.

On the counter-signals: the thesis does not depend on the next funding event, because $541.3 million and no debt is real independence. No metric has changed definition. That matters, and it is the strongest thing on Rigetti’s side of the page.

IonQ, marked

Unit economics improving with scale: not yet disclosed.

Growth that funds itself: absent. Operating cash outflow with a foundry now attached, and depreciation and amortisation for the half at $89.2 million against $17.2 million a year earlier.

Demand before capacity: present, with a caveat. IonQ did not build empty capacity, it bought a foundry with existing customers and revenue. The caveat is that some of those customers are competitors who now buy from a rival.

Revenue that repeats: partly present. Foundry revenue recurs. Whether quantum revenue does cannot be determined, because the two are not reported separately.

Progress costing less ownership: absent. Roughly 24.1 million new shares issued for SkyWater, alongside $741.3 million of cash.

On the counter-signals, two are live. Sales and marketing grew 242 percent year on year, which is the leverage test failing so far. And a headline metric changed definition when the September outlook became the first to include a semiconductor foundry.

The record at Rigetti, and what it shows

Look back at the marks above and notice what they have in common. Unit economics, growth that funds itself, and the cost of progress in ownership are not conditions that happen to a company. They are the residue of decisions somebody made, and the same person will be making the next ones.

So track record matters here, and written as character it is useless. Written as a record of decisions with results attached, it tells you what someone is likely to do next.

Subodh Kulkarni became chief executive of Rigetti on 12 December 2022, succeeding the founder Chad Rigetti, who also left the board that month. Before that he ran CyberOptics, a maker of precision sensors and inspection systems for chip and electronics manufacturing, from 2014 until it was sold.

What happened in between is the part worth having. When he took over, CyberOptics was losing money every quarter. The first quarter of 2014 showed a net loss of $0.8 million, the second a loss of $0.3 million, the third a loss of $0.4 million, on revenue growing at double digits from a small base. He said at the time that the goal was sustainable profitable growth.

By the first half of 2022 the company reported revenue of $51.8 million, up 21 percent, net income of $8.0 million, up 75 percent, a record backlog of $55.5 million, and it was on track for its first year of revenue above $100 million. In August 2022 Nordson agreed to buy it for $54.00 a share, about $380 million, a 47 percent premium to the average price of the previous month and 11 percent above the 52-week high.

So the answer to whether this management has moved a company from losses to profit is documented rather than asserted. It took eight years, it was done by selling more of a physical product at improving margins, and it ended with a sale above the stock’s own high.

The record at IonQ, including the part that went wrong

Niccolo de Masi became chief executive of IonQ on 26 February 2025, having joined the board in 2021 as head of the blank-cheque company that took IonQ public. Peter Chapman, who had run the company since 2019, moved to executive chair.

His main operating record is Glu Mobile, which he ran from January 2010 to November 2016. He moved it to a free-to-play model, and trailing revenue more than doubled between late 2013 and mid-2015, carried substantially by one title, Kim Kardashian: Hollywood, reported to have generated around $180 million.

Then the harder half, which matters more. The hits churned and were not replaced. Celebrity titles were a deliberate strategy and several failed. By the third quarter of 2016 revenue had fallen to $51.4 million from $63.3 million a year earlier, and the company reported a net loss of $43.7 million against roughly break-even a year before. Around a hundred jobs went. In November 2016 he moved to executive chairman and a new chief executive took over.

One correction worth making, because it is repeated everywhere. Electronic Arts bought Glu in April 2021 at an enterprise value of about $2.1 billion, and that was five years after he had stopped being chief executive. He was chairman, not the operator who delivered it.

Since then his record is capital markets rather than operations: a series of listings through blank-cheque vehicles, one of which brought IonQ public.

So on the profitability question, the honest answer for IonQ’s chief executive is not yet demonstrated. He has grown revenue sharply, handled public markets expertly, and presided over a decline that cost him the job.

The capital spending line, which is the one I will be watching hardest

This is the number that changed my mind about what to track, and it deserves its own section.

The ratio is what matters rather than the amount. Sixteen point four against nine point five means the business is putting more into hardware than it takes in from customers, and the prior year’s eight point two means that gap widened rather than closed. Add the $10.1 million still sitting unpaid in accounts payable at the end of June and the $21.0 million of equipment commitments, and the committed figure is larger than the spent one.

The direction is not a guess either. Rigetti’s own filing sets out an intention to raise capital expenditure substantially, naming an upgrade to its fabrication facility, a possible second facility, and more refrigerators, and saying plainly that this needs a significant amount of cash.

Now the honest scale check. Webvan spent $259.8 million in a single year against about $178.5 million of revenue, and before that $64.3 million against a far smaller base. Rigetti is spending sixteen million dollars. This is not Webvan, and saying so would be alarmism.

But the shape is the same shape, and the shape is what the week was about. Capital spending ahead of revenue, rising, on assets that cannot be unbuilt, funded by a cash pile rather than by customers. That road has a record, and the record is why this is now the first thing I check each quarter rather than the last.

The number neither of them discloses

Cost per system, and whether it is falling.

Tesla’s turn came when the margin on each car went from negative to positive and kept climbing with volume. That is the same question here in different clothes: does the tenth machine cost less to build than the first. It is the difference between a research program that sells its output and a manufacturing business.

Neither company publishes it. Rigetti reports revenue from system sales and never the cost of producing one. IonQ now owns a fab and does not separate what quantum costs from what the foundry earns.

So the first item on my watchlist is a disclosure that does not exist. If either company ever starts publishing it, that decision will tell me something on its own, because companies publish the numbers that flatter them.

What I am watching, with dates

Five things, each traced to the day that produced it.

Capital spending against revenue, every quarter. From Thursday. The next reading arrives with the third-quarter filings in November.

The share count on every cover page. From Tuesday. ASML’s customers funded a build and the existing holders were made whole through a buyback. Rigetti’s issuance had no offset. The count is the receipt, and it is on the first page of every 10-Q.

Whether the 99.5 percent fidelity target lands by 31 December. From Wednesday. Management named the number and the date, and about ninety days remain.

Whether IonQ separates quantum revenue from foundry revenue. From Wednesday. Until it does, the September outlook cannot be read as evidence about quantum demand.

A second system sold to the same customer. From Thursday. Amazon published repeat-purchase share when it had nothing else worth publishing, because a second order proves the product rather than the marketing.

And one item that comes straight out of the two records. At CyberOptics, Kulkarni’s evidence of progress was margin improving on a physical product sold in volume, which is exactly the disclosure Rigetti does not yet give. At IonQ, de Masi’s method is acquisition, so the thing to watch is whether the five companies bought since 2025 show up as revenue that recurs rather than as a larger cost base. Both of those are visible in the same filings as the other five.

None of those five is what I was watching a week ago. That was the fidelity headlines and the award announcements, both real news, neither of which tells you what it costs to own the company while the news happens. The new five are duller and harder to find, four of them sit in documents anybody can pull in an afternoon, and the fifth does not exist yet. That is the entire return on a week of reading old filings: a shorter list worth looking at, and a longer list to ignore.

The limit, and what happens next

This is not statistics. A handful of cases, chosen partly because we know how they ended, and the survivors get written about while the failures are forgotten. What this week produced is roads with known hazards, not probabilities. It cannot tell you whether the technology arrives either. A company can do everything on the screen correctly and fail because the physics takes another decade.

The management records have the same limit, and a sharper one. A record predicts method, not outcome. Kulkarni’s says he will build and sell product patiently. De Masi’s says he will assemble and finance. Either can work here, and the setback in de Masi’s record may be worth more than a clean one, since a chief executive who has watched a hit-driven business churn is unlikely to bet this one on a single title again, and the acquisitions suggest he has not.

And on fairness: Rigetti is a holding of mine, and the screen marks it absent or not yet disclosed on four of the five traits that earn patience. I would rather publish that than defend it.

Third-quarter filings land in November for both companies. That is when capital spending against revenue, the share count and the revenue split get their first reading under the new list, and the fidelity target has until 31 December.

Nothing in this week was a forecast. It was an argument about which four or five lines to open when the filings arrive, and about how much of what gets quoted in between can be left alone.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com.

Not investment advice. The subscriber decides.

Disclosure: the author holds a small position in Rigetti and no position in IonQ. This is analysis, not a recommendation.

What killed Iridium, and what did not

Not the engineering. The system did what it promised: a call from anywhere on earth, switched through space.

The cost structure killed it. More than $5 billion of capital spent before a single subscriber existed, then roughly $1.5 billion of debt to service, with interest running at tens of millions a month. Ten thousand customers cannot carry that. Five hundred thousand might have. The gap between those two numbers is not a rounding error in a forecast, it is the difference between a business and a monument.

Then comes the part that makes this the most useful failure in the archive. The assets were bought out of bankruptcy for $25 million, and the same satellites, in the same orbits, doing the same job, operated as a viable business afterwards. The new owner’s break-even was reported at around 60,000 subscribers instead of the million the old structure needed.

Same technology. Same customers available. Different fixed costs, opposite outcome.

Hold that against what was said at the time. The commentary in 1999 blamed handset size, call prices and a network outdated on arrival. All true, and all beside the point, because a company with a far smaller cost base sold the same service to the same kind of customer and made it work.

Webvan ran the same play on the ground

Webvan’s capital spending went from $64.3 million in 1999 to $259.8 million in 2000 against roughly $178.5 million of revenue, building automated warehouses across nine markets.

Its own break-even assumption required distribution centers running at 40 to 50 percent utilization. They ran at 25 to 30 percent while more were being built. The plan was not fantasy, it worked on paper at scale, and the company ran out of time before reaching scale.

Both companies made the same decision: commit the fixed costs first and let demand catch up. It is the single worst-performing road in the filing record, and it is worth being precise about why.

Fixed costs do not respond to a story. Depreciation, interest, rent and maintenance arrive on a schedule that has no interest in whether the customers did. A company with high fixed costs and uncertain demand has bought a deadline, and the deadline is set by the balance sheet rather than by the market.

So who is buying what Rigetti sells

This is the question I should have asked first, months ago, and did not.

Rigetti’s second-quarter revenue was $5.1 million, up from $1.8 million, and the company attributes it to on-premises sales of its Novera and Cepheus systems. The buyers, as far as the disclosures go, are national laboratories, universities, research organisations and government-linked institutions, including an $8.4 million order from C-DAC in India and work with the Pittsburgh Supercomputing Center and HPE.

That is a real customer list and it has a specific shape: institutions with research budgets, buying one machine. It is not enterprises buying computing capacity because it solves a business problem more cheaply than the alternative.

The distinction is not snobbery about who the customer is. A research budget is a real budget and a national laboratory pays its invoices. It is about what the purchase proves. A laboratory buying a machine to study it proves the machine is interesting. A company buying a second one to run production proves it is useful.

What would change the picture is a second sale to the same institution, or a commercial buyer with a profit motive. Neither appears in the disclosures yet, and Rigetti does not publish repeat purchases.

Why the Commerce Department would pay $100 million

Worth answering directly, because the reason changes what the award tells you.

The Department is not buying quantum computers. It is buying domestic capability. Rigetti owns Fab-1, a superconducting fabrication facility on American soil, and the May program spread awards across nine companies and five different technologies with the explicit purpose of keeping that capability onshore.

So the award is evidence about national policy, not about commercial demand. A customer who buys because the product earns them money is a different signal entirely from a funder who buys because the capability must exist somewhere. The first proves the product. The second proves the politics.

That distinction matters for anyone holding the stock, because the government money is real and the revenue it produces is not the kind that compounds.

What IonQ bought, and whether it was a problem

You asked the right question about SkyWater, and the pre-deal filings answer it.

SkyWater was the largest exclusively United States based pure-play semiconductor foundry, accredited by the Defense Microelectronics Activity as a Category 1A Trusted Foundry, with facilities in Minnesota, Florida and Texas. It reported revenue of $342.3 million in 2024 and $150.7 million in the third quarter of 2025. That is an operating business with real customers, not a science project.

It was also thin. SkyWater had about $31 million of cash at the end of September 2025 against IonQ’s $1.5 billion, and IonQ’s chief executive described the transaction as something of a rescue on the call. IonQ paid $35.00 a share, a 38 percent premium to the thirty-day average price as of 23 January 2026, for an equity value of about $1.8 billion.

So the honest answer is both. A proven, accredited foundry with delivered revenue, bought at a premium, with a weak balance sheet and the capital intensity and thin margins that come with owning fabrication.

One risk sits on top of that. SkyWater’s customers included other quantum companies, D-Wave and PsiQuantum among them. Those customers now buy their chips from a competitor. Whether they stay is the integration question, and it will show up in revenue rather than in a press release.

What these two have in common with the survivors

Not an equivalence, and worth stating plainly. But two similarities are real and both are visible in documents.

The first is that their funder and their customer are often the same party. Intel, TSMC and Samsung bought shares in ASML because they needed the machine. The United States government funds Rigetti and also buys quantum research. That alignment is the thing that carried ASML through years of delay, and it is present here in a weaker form.

The second is that both are selling to a small number of sophisticated institutional buyers, which is where Intuitive Surgical started before instruments and service turned installed machines into recurring revenue. Rigetti and IonQ have the installed base problem in front of them rather than behind them.

The difference from Iridium and Webvan matters most today. Rigetti holds $541.3 million and no debt. IonQ holds about $2.0 billion after SkyWater. Neither is carrying $1.5 billion of debt into a market that has not arrived.

What would move either of them onto the Iridium road

Three conditions, and all of them are checkable. Capital spending rising faster than revenue, which is the Webvan signature. Debt taken on to fund capacity, which is what removed Iridium’s ability to wait. And utilisation running below break-even while commitments continue, which is the same arithmetic in any industry.

So here is where both companies stand on the first one, from the cash flow statements rather than from the balance sheet.

Rigetti spent $16.4 million on property and equipment in the first half of 2026, against $8.2 million in the same period of 2025. Revenue for that half was $9.5 million. Capital spending is running at roughly one and a half times revenue and it doubled year on year. A further $10.1 million of equipment purchases sat in accounts payable at 30 June, and total purchase commitments were $35.0 million, of which about $21.0 million relates to property and equipment.

The company is also explicit about where this goes. Its own filing says it may seek to significantly increase capital expenditures, including upgrading its chip fabrication facility, possibly investing in a new one, and buying additional refrigerators, and that this would require a significant amount of cash.

IonQ’s version did not come through capital spending. It came as an acquisition. The company’s filing put the cash requirement for SkyWater at approximately $1.0 billion, including roughly $0.8 billion of purchase consideration and $0.2 billion of debt repayment and transaction costs. Depreciation and amortisation for the first half rose to $89.2 million from $17.2 million a year earlier.

Neither of those is Webvan. Rigetti is spending sixteen million dollars, not two hundred and sixty. IonQ bought a business with customers rather than building empty capacity. But both are now carrying fixed costs that arrive whether the customers do or not, and the direction on both is up.

What this does not tell you

Iridium and Webvan are two cases, chosen because they are famous, and famous failures are not a sample.

The comparison also flatters the two live companies in one respect. Neither has yet had to build at Iridium’s scale, because nobody is asking them to deliver a global service. The test they face is different and arguably harder: proving one unit works before anyone will fund many.

And a company can do everything on this list correctly and still fail because the technology does not arrive. The road matters. It is not the only thing that does.

Tomorrow

The whole week gets published as a screen, both companies get marked against it, and I write down what I am watching from here, with dates attached where dates exist.

Some of that list I would not have been watching a week ago. The obvious numbers, the ones quoted at everyone all day, can all improve while a position gets worse. Revenue can triple while the share count rises. Guidance can be raised while what it measures changes. A fidelity record can be set while the machine that matters falls further from its target.

Tomorrow is about the handful of lines that cannot do that.

Not investment advice. The subscriber decides.

Disclosure: the author holds a small position in Rigetti and no position in IonQ. This is analysis, not a recommendation.

What Amazon chose to put in front of people

In the 1997 letter to shareholders, Jeff Bezos did something unusual. He named the measures he wanted the company judged by, before anyone could judge it on them.

Customer and revenue growth, the degree to which customers kept buying on a repeat basis, and the strength of the brand. Not earnings, which did not exist. Not the multiple, which everyone else was arguing about.

Then he published the number. In the 1998 letter: over 58 percent of orders in the fourth quarter of 1997 came from repeat customers, rising to over 64 percent a year later. Customer accounts went from 1.5 million to 6.2 million across the same year.

Notice what that pairing does. Accounts rising on their own could be marketing spend buying strangers. Accounts rising while the repeat share also rises means the new customers are staying. One number checks the other, and both were published by a company with no earnings to show instead.

That is a company handing the market a test it can be marked against, in public, with a date on it. The test happened to be the right one, and that is the part worth slowing down for.

Why repeat purchase was the right number

A first order tells you the marketing worked. A second order tells you the product worked.

That is the whole logic, and it is why repeat-purchase share is worth more than revenue growth in a young company. Revenue can be bought. Repeat business has to be earned, one customer at a time, and it compounds without further spending.

The market spent 1999 arguing about whether Amazon deserved its price. A reader watching the repeat rate climb from 58 to 64 percent had a different question in front of them, and a better one.

It is worth being clear that this was not a prediction. The repeat rate did not say Amazon would be worth a trillion dollars. It said the thing customers were buying was working well enough that they came back, which is the narrowest useful claim available about a company that loses money.

Rigetti’s number, and its date

Now the live version, and this is the cleanest scoreboard of the three.

On 7 April 2026, Rigetti announced general availability of Cepheus-1-108Q, a 108-qubit system with approximately 99.1 percent median two-qubit gate fidelity, and said it expected to reach approximately 99.5 percent later in the year.

A number, and a deadline, both chosen by the company. The 36-qubit system already ran at 99.5 percent, so the target is a return to a level the smaller machine had reached rather than a leap into the unknown. One quarter of 2026 remains, and no announcement of reaching it has appeared.

There is a detail worth pinning down before the quarter closes, because it decides what counts as a hit. The figure is a median across the machine, not the best result on the best chip, and it applies to the 108-qubit system rather than to a smaller prototype. A press release announcing 99.5 percent on something else would not be the same claim.

Rigetti has also published a longer target: roughly 99.9 percent at about a thousand qubits within about three years. That one cannot be marked yet. The 2026 number can, in about ninety days, and the result will be information about the forecaster as much as the machine.

IonQ’s four numbers

IonQ publishes a different kind of scoreboard, and it has published it four times this year.

The 2026 outlook opened at $225 million to $245 million in February. It went to $260 million to $270 million in May. It went to $280 million to $290 million on 5 August. On 8 September it went to $450 million to $460 million.

Three raises in seven months reads as a company outrunning its own expectations. Look at the last one more closely. It is the first outlook that includes SkyWater, the semiconductor foundry IonQ bought for about $1.8 billion and closed on 31 July 2026. The jump of roughly $170 million is largely a foundry business being added to the total, not quantum demand arriving faster than expected.

The number went up. What the number measures changed at the same time.

Guidance is also the one scoreboard management controls completely. A company sets the bar, then reports against the bar it set. That is not a criticism of IonQ specifically, it is the structure of the instrument, and it is why guidance is worth less as evidence than a number an outsider can verify.

When the scoreboard changes shape

This is the counter-signal, and it is the strongest warning in the whole screen.

A company that redefines its own metric is telling you the old definition stopped flattering it, or that the business underneath has changed enough that the old series no longer means what it did. Either way the history breaks, and comparisons to last year quietly stop working.

IonQ’s revenue line now contains a semiconductor foundry. That does not make it a worse company. It makes the 2026 outlook something other than a measure of quantum demand, and anyone treating the raise as evidence about quantum adoption is reading a number that has changed underneath them.

The same caution applies to algorithmic qubits, the performance measure IonQ devised and reports against. A metric a company invented and grades itself on is not fraud. It is also not a benchmark.

What neither of them publishes, which is also a scoreboard

Silence is a choice too, and both companies make it in the same place.

Neither publishes what a system costs to build. Rigetti sells 108-qubit machines to labs and universities and reports the revenue. It does not report the cost of producing one, or whether that cost is falling as more are made. That single series would tell an owner more about the future than any fidelity figure, because it is the difference between a research program and a manufacturing business.

Neither publishes repeat purchases either. A second system sold to a customer who already owns one is the strongest evidence available that the product works, and it is exactly what Amazon chose to publish when it had nothing else to show. Both companies could disclose it. Neither does.

And IonQ does not separate quantum revenue from foundry revenue, which is the disclosure that would settle what the September outlook really contains.

None of that is misconduct. Companies disclose what they are required to disclose and what they want to. The point is that the gaps are as informative as the numbers, and a reader who notices which questions cannot be answered is already ahead of one who only reads the answers that are offered.

How to grade a scoreboard

Four questions, each answerable from documents, and they work on any company in any industry.

Did the company name a number and a date, or only a direction. Is the measure one an outsider could verify, or one the company defined. Has the definition changed between reports, and does the filing say so. And what is the record: of the targets published in the last three years, how many arrived on time.

That last one is the whole value of the exercise. A single missed target means very little, because forecasting a laboratory is hard and honest people get it wrong. A pattern of them is different, and it is the most reliable evidence about management available to an outside shareholder. Keeping the record costs nothing. A note on a calendar three times a year builds it.

The reverse case matters too. A management that hits its published numbers repeatedly has earned something that does not show up in any financial statement, and it is the closest thing to a moat a pre-profit company can have.

What this does not tell you

A published target is not a promise, and missing one is not dishonesty. Engineering slips. ASML’s EUV schedule moved by years, as yesterday’s piece laid out, and the machine still ended up defining an industry.

Nor does publishing more make a company better. A business can disclose a great deal and still be worth nothing, and plenty of companies that disclosed very little turned out fine.

What a published, dated number does is narrower. It converts a story into something that can be marked, and it puts the marking on a calendar rather than leaving it to opinion.

Tomorrow

Webvan was operating in nine markets while its distribution centres ran at 25 to 30 percent of the utilisation they needed to break even. Iridium put a constellation of satellites in orbit before it had subscribers to use them.

Tomorrow is the road with the worst record in the whole archive: committing fixed costs before one unit works.

Not investment advice. The subscriber decides.

Disclosure: the author holds a small position in Rigetti and no position in IonQ. This is analysis, not a recommendation.

What ASML agreed to, in its own filings

The terms are in the filings, and they are worth reading slowly.

Announced on 9 July 2012. Intel took 15 percent of ASML for 2.51 billion euros and committed 829 million euros of research funding. TSMC followed on 5 August with 5 percent for 838 million euros and 276 million euros of research money. Samsung completed it on 27 August with 3 percent for 503 million euros and another 276 million euros.

The subscription price was 39.91 euros a share, the average of the twenty trading days to 6 July 2012. The shares went into Dutch foundations and came back to the customers as depositary receipts that were non-voting, except in exceptional circumstances.

Add it up and three customers owned 23 percent of their supplier, while the supplier held 3.85 billion euros of cash plus 1.38 billion euros of committed research funding over five years.

Two details are worth pausing on. The equity was capped: ASML said it might issue up to 25 percent to customers and no more, so the program had a ceiling written into it from the start. And the research money was separate from the share purchase, which means the customers were paying twice, once for a stake and once for the work, and only the first of those bought them anything they could sell.

The detail that changes the deal

Here is the part that matters, and it is one sentence in the press release.

The entire 3.85 billion euros raised from selling those shares was returned to ASML’s other shareholders through a synthetic buyback, structured with a reverse split so that no dilution occurred from the program.

Read that again. Customers bought almost a quarter of the company. Existing owners were not diluted, because the cash from the share sale was handed straight back to them and the share count was neutralized. The customers got exposure and a seat at the technology. The existing holders got cash and kept their proportion of the company.

That is not how these deals normally work, and it is the whole reason this one is worth studying.

It also tells you what ASML’s board thought its job was. Faced with a chance to raise almost four billion euros from three of the largest companies in its industry, it took the money and passed it through, rather than keeping it on the balance sheet. The research funding, the part it needed, came separately and did not cost its owners a thing.

What they were funding, and how long it took

The money went to two projects. Extreme ultraviolet lithography, and 450-millimetre wafer technology.

Take the second one first. Intel’s statement in July 2012 said the investment would enable 450-millimetre prototype systems as early as 2015. At the annual meeting in 2014, ASML’s chief executive was asked why the 450-millimetre programme had been suspended, and answered that the decision followed a consensus in the industry, and that the programme would restart when the industry needed it. One of the two projects the money was raised for had been put down inside two years.

The first one slipped. In its first-quarter 2013 results, ASML confirmed its expectation of a ramp of production using EUV in 2015. At its investor day in 2014 the language had moved to enabling cost-effective manufacturing from 2016 or 2017. The scanner built for volume manufacturing, the NXE:3400B, first shipped in 2017. The customers who paid in 2012 kept paying through all of it.

Sit with what that would have felt like for an ordinary shareholder through those years. A company taking money from its customers to build something repeatedly late, with one of the two funded programs suspended. Every quarter the case rested on a machine that was not yet in volume production. The difference between holding that and holding a pre-profit quantum company today is smaller than it looks from here.

Why being the buyer changes everything

A financial investor funding development wants a share price. A customer funding development wants a machine, and will keep funding through delays that would make an investor sell. That is not loyalty. It is that the customer has no alternative source for the thing it needs, so abandoning the supplier means abandoning its own roadmap.

That distinction is worth holding on to, because it is the question to ask of any co-investment. Is the money coming from someone who needs the product, or from someone who needs the stock to go up. The first kind survives a slipped schedule. The second kind does not.

ASML’s participants were the three companies that would have to buy every machine it ever built. Their incentive and the company’s incentive were the same thing.

Rigetti’s version of the same fork

Now the live case, and the terms are public.

On 8 September 2026, Rigetti issued 7,739,938 shares to the Commerce Department as an inducement to enter an award agreement worth up to $100 million. Divide one by the other and the implied price is $12.92 a share, set by a formula from May that took the lowest of three reference closes and discounted it by 15 percent.

Rigetti had 333.8 million shares outstanding at the start of August, so the issuance added about 2.3 percent to the count. No offsetting buyback was announced alongside it, and nothing was returned to existing holders. The ownership moved on day one and the money is up to $100 million, for research not yet done.

Is the funder also the buyer, the way Intel was? Partly. The government is a customer of quantum research and a major funder of the field, and it has an interest in the capability existing on home soil. It is not buying quantum computers by the hundred the way Intel buys lithography machines, so the alignment is real but weaker than ASML’s.

There is one more difference worth naming. Intel, TSMC and Samsung paid 39.91 euros a share, set on a twenty-day average before the announcement. Rigetti’s formula took the lowest of three reference prices and then cut 15 percent off it. One funder paid an average. The other received a discount.

IonQ took the third route

IonQ (IONQ) funded its build by buying rather than by issuing to a funder. It paid $741.3 million in cash and 24.1 million new shares for SkyWater, a foundry, closing on 31 July 2026.

Same fork, third answer. ASML’s owners were protected by a buyback. Rigetti’s owners absorbed the issuance. IonQ’s owners paid with both cash the company had raised and a slice of the company itself, and got an operating business in return.

Which of the three is best depends on what the money buys, and that is the honest answer rather than a dodge. Paying with stock for an asset that produces revenue is a different transaction from paying with stock for a research award. One is buying a business. The other is buying time.

What to check when a company does this

Four questions, all answerable from documents, and they work on any company announcing outside funding.

What did the funder get, and at what price against the market at the time. What came back to existing holders, if anything. Is the funder also the buyer of the product. And what happened to the share count on the next 10-Q cover page.

The last one is the discipline. Announcements describe the money coming in. The cover page records what it cost.

None of these four requires an opinion about the technology, which is the point. You can run them on a company whose product you do not understand and still learn whether the deal was struck on terms that treat you as an owner.

What this does not tell you

ASML worked, and that makes this a dangerous case to reason from. Most companies with customer co-investment do not become the only supplier of a critical machine on earth.

The structure did not cause the outcome either. The machine worked because the engineering worked. What the structure did was keep the existing owners whole while the engineering took years longer than promised, and that is a separate thing from success.

Nor is it fair to read Rigetti’s terms as evidence of a bad outcome. They are evidence of a cost, paid up front, by people who owned the stock in August. A company with $541.3 million of cash and no debt was not forced into those terms by desperation, which means somebody decided the award was worth 2.3 percent of the company. That decision can be right. It is still a decision, and it is now on the record where it can be marked in a year.

One more limit. ASML in 2012 was an established supplier whose customers were the largest chipmakers on earth, which is why three of them could be persuaded to fund it. Rigetti is not that, and the structure available to a company depends on what it already is. Comparing the terms is fair. Expecting identical terms is not.

Tomorrow

Amazon published its repeat-purchase rate while the market argued about its multiple. Rigetti published a dated target for fidelity this year, with one quarter left to hit it.

Tomorrow is about the scoreboard a company chooses for itself, what it means when management publishes a number with a date attached, and what it means when a company publishes nothing at all.

Not investment advice. The subscriber decides.

Disclosure: the author holds a small position in Rigetti and no position in IonQ. He also holds IBM. This is analysis, not a recommendation.

The number everyone quotes, and the number that mattered

Amazon’s 1999 annual report is worth reading now, because nobody reading it then knew how the story ended.

The net loss was $720 million. The cash consumed by operations was $90.9 million. The company explains the difference itself: $358.9 million of the loss was non-cash, things like depreciation, stock compensation and goodwill amortization, and a further $230.1 million came from changes in operating assets and liabilities, mostly accounts payable rising faster than inventory.

Read that last part slowly, because it is the whole thing. Suppliers shipped Amazon books and waited to be paid, while customers paid immediately. Growth was financed by the people Amazon bought from, and the faster it grew, the more of their money it held.

That is not a story about vision or patience. It is a structural feature of the business model, and it sat in the liquidity section of the filing.

The same document shows where the rest came from. Investing consumed $922.3 million in 1999 and financing brought in $1.104 billion, mostly from a $1.25 billion convertible note sale in February. Amazon was raising money hand over fist. What matters is what the operating business was doing while that happened, and it was close to paying for itself.

The same years, a company that looked identical from the outside

Webvan went public in November 1999, two weeks before Amazon’s quarter closed, selling 28,750,000 shares at $15.00 each. Same market, same story, same kind of investor letter.

Then look at what it did with the money. Capital expenditure was $64.3 million in 1999 and $259.8 million in 2000, building automated warehouses and a delivery fleet. Revenue in 2000 was about $178.5 million. The company spent more on building capacity in one year than customers spent with it.

Amazon’s growth pulled cash in from suppliers. Webvan’s growth pushed cash out into concrete. Both were unprofitable internet companies with famous backers and a large addressable market. Only one of them had a model where scale helped.

What makes Webvan worth studying rather than mocking is that its unit numbers were not fantasy. Management expected a distribution center to break even at 40 to 50 percent utilization, and utilization ran at 25 to 30 percent while more centers were being built. The plan worked at scale and the company ran out of time before reaching scale, which is a different failure from a bad idea and much easier to miss.

Why the comparison people reach for does not work

Here is the trap, and I have fallen into it myself with Rigetti.

When someone questions a pre-profit holding, the answer that comes out is that Amazon lost money for years. True, and useless, because it compares outcomes. In 1999 nobody knew Amazon was Amazon. They knew it was a company with a large loss and a big idea, which is exactly what Webvan was.

Comparing outcomes tells you some companies in this position went on to be enormous. It cannot tell you which, because it starts from the answer and works backwards. Survivors get books written about them and failures get forgotten, so the pattern looks stronger than it was.

The same error gets used against young companies. Pointing at Webvan to dismiss anything unprofitable substitutes a famous outcome for the evidence in front of you, exactly as the Amazon argument does.

What can be compared is conditions. Not what happened next, but what was in the documents at the time, on both sides.

What this is, and what it is not

Let me be plain here, because this is the part that gets misread.

I am not saying Rigetti is the next Amazon. I am not saying IonQ is the next Tesla, or the next Webvan. I do not know, and nor does anybody else. No company in this week is being predicted, and any piece that tells you a young company is the next anything is selling a forecast dressed as analysis.

What this does is narrower. Companies that later became great left evidence behind while they were still unproven, in ordinary filings anybody could read at the time. So did the companies that failed. Learn what that evidence looks like and you have something to check a quarterly report against, instead of a story to agree or disagree with.

The reason any of it transfers is that Amazon was not always Amazon. In 1999 it was an unprofitable retailer with a big idea, a funding requirement and no earnings to be valued against, which is the same ground Rigetti and IonQ are crossing right now. The terrain repeats even when the destination does not. Companies crossing it leave the same kinds of marks in their filings, and those marks are what this week collects.

That is a lens. It does not tell you what Rigetti will be in ten years. It tells you which lines of its filing are worth more than its press releases, and what a good answer and a bad answer look like on each one.

Used properly it runs in both directions. It can build a case for patience and it can take one away, and this week it does both to a company I own.

The rule that keeps it usable

Everything in this week has to be visible in a filing while the outcome is still unknown, and markable as present, absent or not yet disclosed. Not yet disclosed is a real answer, because an absence of disclosure is itself information about a company.

Amazon’s payables funding its inventory qualifies. Bezos being a visionary does not, because that only becomes obvious once the share price has already done the work. Management quality, addressable market and technological lead all matter, and none of them can be checked by a reader on a Tuesday afternoon. This week uses only what can.

The mistake that cost people both ways

On 31 May 1999, Barron’s ran a cover story headlined Amazon.bomb. It called Bezos “just another middleman” and cited analyses valuing the shares below $10.

On the stock, it was right. Amazon fell from $221.25 in April 1999 to around $6 by 2001, and the market value bottomed near $2.9 billion in January 2002. On the business, it was wrong. Sales went from $609 million in 1998 to $3.1 billion in 2001, straight through the collapse.

Look at what happened there, because it is the most useful thing in this week. Two separate claims got fused into one verdict. This stock is overpriced. This business will fail. They need different evidence and they can both be true, or either one alone. A reader who took the article as a single judgement got both halves wrong, and sold a business that was compounding while its price fell.

That is the first thing to carry into Rigetti and IonQ. You are allowed to think the prices are unsupportable and the technology will work. Those are not contradictory positions, and keeping them apart is most of the discipline.

What this week does, day by day

Five lessons, one a day, each one a road that has been travelled before.

Tuesday: grade a company on the scoreboard it chose for itself. Amazon published its repeat-purchase rate while everyone argued about the multiple. Rigetti has published a dated fidelity target for this year.

Wednesday: who funds the growth decides who survives a closed window. Amazon’s suppliers funded its inventory. Webvan’s investors funded its concrete. The test is what happens if no new money is available for eight quarters.

Thursday: committing fixed costs before one unit works is the road with the worst record in the filing archive. Webvan is one case. Iridium is the harder one, because its satellites still fly today and its shareholders got nothing.

Friday: dilution, the quiet road, and then the whole thing published as a screen with both companies marked against it.

Why this matters for the two quantum companies specifically

Rigetti and IonQ have no earnings, so there is no price to be called low against. That removes the tool I normally use and leaves the traits as the only instrument available.

It also makes the stakes concrete, and I own one of them, so the reasoning has to survive contact with my own position. Rigetti issued 7.74 million new shares to the U.S. government in September, about 2.3 percent of the company, for an award of up to $100 million, before the research was done. That is the fifth trait moving the wrong way, checkable every quarter on the cover page of the 10-Q. IonQ’s sales and marketing spending grew 242 percent year on year, which is the first counter-signal, and the next two filings answer whether it persists.

Neither fact predicts anything. Both are the kind of thing that was visible in 1999, on both sides of the Amazon and Webvan comparison, and mostly ignored.

The work this week is to build the rest of that list out of history rather than out of opinion, then run both companies through it on Friday and publish the result, including the parts that do not flatter a position I hold.

What this week cannot do

Three limits, said now rather than discovered on Friday.

It cannot tell you whether either company makes it. The traits improve the odds of asking the right question, nothing more.

Survivorship runs in both directions. Some companies failed every test and survived anyway, because a buyer appeared or a market turned in their favour. The screen narrows the field, it does not pick.

And quantum has no revenue history to pattern-match against. What transfers is the shape of the evidence, not the economics.

Tomorrow

In 2012, Intel, TSMC and Samsung bought shares in one of their own suppliers to pay for a technology that did not work yet. The supplier was ASML, the machine was extreme ultraviolet lithography, and the customers were funding the thing they intended to buy. It took years longer than planned, the schedule slipped repeatedly, and it ended in something close to a monopoly.

That is the closest historical parallel to the U.S. government taking shares in Rigetti this month, and it is where the week starts properly.

Not investment advice. The subscriber decides.

Disclosure: the author has held IBM since March 2020, is up about 100 percent on the position, and reinvests the dividend. He also holds a small position in Rigetti. No position in IonQ. This is analysis, not a recommendation.

The Firewall run, and the assumption underneath the price

Run IBM through the Firewall and the hidden assumption it surfaces is not about quantum at all.

The bet a buyer makes at today’s price is that IBM’s software and mainframe franchise keeps producing the cash that funds a decade-long research program, and that the shortfall of July 2026 was a timing event rather than the front edge of a decline.

Classification: needs proof. Not because the evidence is thin, but because the specific thing that would settle it has a date on it and the date has not arrived.

The framework the Firewall prescribes here is the same one it prescribed for Rigetti in May and IonQ this week: narrative against evidence. What makes IBM different is that the narrative and the evidence sit in different parts of the company. The quantum story is narrative with almost no disclosed economics. The engine is evidence with almost no story attached. Most coverage this year has valued the first and ignored the second.

What the week proved for that assumption, and against it

For it: the two segment misses were one event, because the software shortfall was in Transaction Processing, the stack that runs on the mainframe. The z17 program is running at nearly 130 percent of z16, which was the strongest on record. Clients representing 85 percent of installed capacity are maintaining or growing it. Distributed Infrastructure had its best quarter on record, up 37 percent, so the money that skipped the mainframe partly went to other IBM hardware. Operating pre-tax margin expanded 30 basis points in the worst quarter IBM has had in years, which is not what a company discounting to hold volume looks like. And remaining performance obligations stood at about $68 billion.

Against it: management has guided Transaction Processing down low to mid single digits for the second half and called recovery a 2027 opportunity, so the mainframe software multiplier sits idle for roughly eighteen months. First-half free cash flow was flat, which leaves the second half carrying the promise of about a billion dollars of growth on its own. And the measures that look most reassuring, program-to-program and installed capacity, are defined and reported by IBM, audited by nobody.

That is a case with a hole in it, and the hole closes or widens on one date.

What nobody will tell you, which is its own finding

Four days of reading produced a tally worth publishing on its own.

The mainframe has no revenue line of its own in the accounts. That 42 percent everyone repeated is management’s description of a product the filing never isolates, and the nearest audited figure fell about 10 percent.

Quantum has no segment, no revenue line and no dated order book anywhere in IBM’s reporting. One number exists, quoted from a slide rather than a filing: signings approaching a billion dollars accumulated since 2017.

Widen the lens and the sector has no shared yardstick either. Each of the three publishes a different measure of its own devising, on hardware of its own choosing, verified by nobody outside. Thursday laid out all three side by side.

So the company with the most quantum revenue discloses the least about it, and the two with almost none disclose everything, because for them it is the whole business. That is not a scandal. It is a reason to value IBM on the parts it does disclose and to treat the rest as a free option.

What quantum costs an owner, at each of the three

Here is the sentence this week was built to earn, and it is an ownership sentence rather than a technology one.

At IBM I am not paying for quantum. The program costs about one billion dollars a year against free cash flow the company expects near $15.7 billion, the dividend is paid either way, and no outside equity was disclosed as the price of the government’s billion. If quantum fails entirely, I own a business that produced $4.76 billion of free cash flow in half a year and still pays me to wait.

At Rigetti and IonQ, quantum is the entire price. Rigetti paid for its award with 7.74 million new shares handed over in full before the research was done, about 2.3 percent of the company. IonQ paid $741.3 million in cash and 24.1 million new shares for its foundry, using roughly a third of its cash pile. If quantum fails at either, there is nothing underneath.

That is the difference, and it is not a judgement about which technology wins. It is arithmetic about who carries the cost of being wrong.

The bar for 21 October, in writing

IBM reports third-quarter results on 21 October 2026. Four things, set now so I cannot move them later.

One. Transaction Processing falls no faster than the low to mid single digits management itself forecast. Demanding growth from that line would be demanding something IBM has already said waits for 2027.

Two. Hybrid Infrastructure, the line the filing does disclose, prints above $2.57 billion. I use it rather than the mainframe figure everyone quotes, for reasons the disclosure tally above makes obvious.

Three. Software lands inside the band the cut guidance set, 6 to 8 percent.

Four. Cash generation in the quarter is heavy enough that the full-year commitment of about a billion dollars of extra free cash flow is arithmetic rather than hope.

[AUGUST BAR] Part two of the August piece set a forward bar for this same quarter. Where it matches the four above, say so. Where it has moved, say what moved it and why, in one paragraph. A bar that quietly changes between August and September is worth nothing.

Four out of four and the July shortfall was a bad three months. Two misses or more and it was something structural, in which case the market got the price right in July and I was the one arguing with it.

My rule for adding, written before the numbers

Not advice, and not a recommendation. A rule, with the reasoning shown, so you can disagree with the reasoning rather than the conclusion.

At $229.55 across 942.1 million shares, the market values IBM near $216 billion. Set that against the roughly $15.7 billion of free cash flow the company expects this year and the yield is about seven percent, with a dividend covered around two and a half times and raised every year for three decades.

Fair, then. Not cheap, and the gap between those two words is where the discipline lives. Six years ago the market had written this company off and priced it accordingly. What is on offer today is a company the market liked a great deal until July and then punished for a single quarter. Those two setups earn different prices from me.

So the rule is this. I add nothing before 21 October. If the four bar items land, the risk in the position falls and I will add at whatever the price is then, because a fair price on a proven engine is a trade I am willing to make. If two or more miss, I do not add at any price until Transaction Processing stops falling, because at that point the question is no longer timing.

The dividend pays me to wait either way, which is what makes waiting cheap.

Rigetti is a sizing question, not a value question

The company I hold, on the only terms available.

Rigetti has no earnings, so nothing here can be called cheap or dear. What can be decided is size and conditions, and the conditions come from this week rather than from the story.

Fidelity: the 108-qubit system runs at 99.1 percent median two-qubit fidelity against 99.5 percent on its own smaller machine and its own target for 2026. One quarter remains. Dilution: the share count is now the thing to watch each quarter, because every milestone bought with equity splits the eventual payoff further. Cash: $541.3 million with no debt, falling about $27.7 million in the quarter, which is years of runway rather than quarters.

My rule for it: the position stays small enough that a total loss changes nothing about my year, and it does not grow on good news about awards or partnerships. It grows only on evidence that each system costs less to build than the last. That is the number that turns a research program into a business.

One more thing, said plainly because it is circulating. There is speculation about Rigetti as an acquisition target, and I can find no filing, no named source and no approach behind it. The government now holding about 2.3 percent makes a deal harder rather than easier. A takeover thesis is a prediction about someone else’s behavior on a timetable you do not control, and this publication does not run those.

IonQ, the one I do not own, which is why it is the cleanest test

No position, nothing to protect, and the most interesting machine of the three.

IonQ builds trapped-ion systems, which is a different physics from IBM and Rigetti: ions held in a vacuum and manipulated by lasers, at room temperature, with every qubit able to talk directly to every other one. Its published two-qubit fidelity is 99.99 percent, the highest of the three by a distance. Its distribution is the broadest, across three clouds. And its $1.87 billion quarterly loss is mostly a non-cash warrant mark against $120.3 million of actual burn.

What would have to become true for it to be ownable on my terms is narrow and checkable. Sales and marketing growing slower than revenue rather than 242 percent faster. Revenue that recurs rather than arriving as milestones, which now needs reading carefully with a foundry inside the company. And the foundry producing a disclosed cost advantage rather than an announced one.

Until then it is the control case for this week’s method, and it earned that role by being the company where I had nothing at stake.

The mistake this week was built to avoid is treating the three as versions of one bet. IBM is a value question, with earnings to measure a price against. Rigetti is an option, where the only decisions are size and conditions. IonQ is a third category again, further along commercially and priced for a future neither has demonstrated. Getting that classification right comes before everything else. Price an option like a business and you overpay for a story. Price a business like an option and you miss the only one of the three that can be valued at all.

What happens now, with dates on it

A week of reading is worth nothing if it ends in a view. So here is the machinery it leaves behind, and every item has a date or a trigger.

21 October 2026. IBM reports. The four bars above get graded in public that week, item by item, including the ones I get wrong. The rule decides what I do, not how I feel about the result.

Every quarter, on Rigetti. Two numbers, tracked in the same place each time: the share count on the 10-Q cover page, because dilution is the cost of every milestone, and any disclosure that bears on what a system costs to build. The fidelity target of 99.5 percent has one quarter left to arrive.

Every quarter, on IonQ. Sales and marketing against revenue, which is the leverage test it is currently failing, and the revenue recognition footnote now that a foundry sits inside the company.

The disclosure tally becomes permanent. Every week from here, when a company stops an answer from being found, it gets recorded and whose wall it was gets named. This week produced three entries and one of them, the mainframe line that does not exist, changed how I read the whole quarter.

Next week. The comparison every investor in young technology reaches for is Amazon and Tesla, and it is incomplete rather than wrong, because nobody writes the patient-holder essay about the companies that did not make it. What the survivors had in their filings before anyone knew they would survive is a checkable question, and the answer is a screen you can run in an afternoon. That is the next framework piece, and it is the one that would have told me more about Rigetti in May than anything I wrote at the time.

What the week changed

I came into Monday thinking the question was whether quantum works. It is not, and that is the thing I will carry out of this week.

Whether quantum works is a question for physicists, and on the evidence it will be years before anyone can answer it with confidence. The question an owner can answer today is who pays while we find out. IBM pays out of surplus and hands its shareholders a dividend throughout. Rigetti pays in ownership, handed over in full before the research is done. IonQ pays with capital it raised and stock it issued. Same technology, three completely different bills, and all three of those facts sat in filings anybody could have read.

That is also why the week ends without a purchase. The price is fair, the engine looks intact, and fair is not the same as cheap. I have the bar written down, the date is three weeks out, and the dividend pays me to wait for it.

Five days, three companies, eleven filings, one decision deferred on purpose and written down in public so it cannot be quietly changed later. That is the job.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com.

Not investment advice. The subscriber decides.