AI capex, the economics of compute, and circular financing — the biggest narrative of this cycle, and the cash flows underneath it.
- 01
The FT Says OpenAI's Annualized Revenue Is About $50 Billion, Not the $70 Billion Everyone Was Quoting, the Nasdaq Fell 1.25% Thursday, and the Market Says "AI Demand Just Got Marked Down" — But the Missing $20 Billion Is Not Demand, It's Bookkeeping: Revenue Is Not a Number, It's a Ruler!
On Thursday, October 8 (US Eastern), the FT reported that OpenAI's annualized revenue is approaching $50 billion, well below the roughly $70 billion in circulation, and the Nasdaq fell 1.25% with Nvidia and Micron leading the slide. This piece argues that the gap is mostly gross versus net accounting, that the market read a change of ruler as a cut in demand, and that the real question is whether we trust the ruler at all.
Oct 09, 2026 - 02
Foxconn Posted Its First Trillion-Taiwan-Dollar Month and a 47% Jump in Q3 Revenue, Nvidia Hit a Record Alongside the Nasdaq on Monday, and the Market Says "the Supply Chain Just Confirmed AI Demand" — But Foxconn's Ledger and Nvidia's Ledger Count the Same Rack: Two Counts Are Not Two Pieces of Evidence!
Foxconn reported on October 5 that Q3 revenue was about NT$3.03 trillion, up 47%, with September the first month above NT$1 trillion; Nvidia rose about 2% to a record that day. The market read it as the supply chain confirming AI demand. This piece argues that two numbers from the same chain cannot validate each other, and that the real test of demand sits at the other end.
Oct 06, 2026 - 03
Meta's New AI Agent 'Muse' Hit No. 1 on the US App Store in Ten Days, and Chip Stocks Roared — Everyone Assumed It Was Nvidia's Stage Again, Yet the Two Names That Actually Ripped Were Intel, Up More Than 12%, and AMD, Crossing a $1 Trillion Market Cap for the First Time, While Nvidia Was Left Far Behind — What Gets Repriced Is Never the Best Chip Story, It Is Whichever Node in the Network Has the Deepest Shortage!
Monday, September 21 (US Eastern) — early Tuesday, September 22 in Beijing — US stocks closed broadly higher, the Nasdaq set a fresh record close, and the Philadelphia Semiconductor Index led the way, up more than three percent on the day. The spark was Meta's new AI agent, Muse, which climbed to No. 1 on the US App Store's overall chart in just ten days. Financial headlines almost unanimously read this as "AI is back," quietly assuming it was Nvidia's stage once again. But the two names that actually topped the day's gainers were not Nvidia: Intel jumped more than 12%, and AMD rose nearly 10% to cross a $1 trillion market cap for the first time — the fourth US chipmaker to do so, after Nvidia, Broadcom, and Micron. Nvidia rose too, but was left far behind. This piece argues that what actually got repriced that day was never whose chip story sounded best — it was whichever node in the compute network had dug the deepest shortage. Earlier in the week, Intel's CEO had already put that shortage on the table: CPU orders were running so hot the company could "only supply about half of what customers want."
Sep 22, 2026 - 04
Chip Stocks Fell 5.9% While Google, Microsoft and Meta All Rose the Same Day — What the Market Repriced Was Never AI, but Who Pays the Bill!
Monday, September 14 (US Eastern) delivered what the financial press immediately christened the AI doomsday trade. Over the weekend, Anthropic's CEO published an essay arguing that AI companies should slow the pace at which they push model capability and deal with safety first; OpenAI's founder publicly agreed. Monday's open: the Philadelphia Semiconductor Index fell 5.9%, Nvidia dropped 3.4%, Intel and AMD fell further. Yet on that same day, the biggest spenders in the entire buildout closed green — Alphabet rose nearly 2%, with Microsoft and Meta also higher, while the S&P 500 lost just 0.48% on the day, less than half a percent. Less prominently reported was what happened in another room at the same hour: the 10-year Treasury yield touched 5.014% intraday, its first visit above 5% since October 2023, with the federal funds rate still pinned at 3.50%-3.75% and market-implied odds of a hike at that week's meeting already above 80%. The popular reading was that the AI bubble had begun to crack. But if the market had truly taken that warning as a verdict on AI, the buyers of shovels should not have rallied. What follows is about what was actually repriced that day — not whether AI slows down, but who picks up the check; and about where the real brake lives, which is nowhere near anybody's conscience.
Sep 15, 2026 - 05
TSMC's August Revenue Just Jumped 53.3% Year Over Year to an All-Time Record, and Nearly Every Outlet Reported It the Same Way — "AI Chip Demand Is So Strong Supply Can't Keep Up"; But When Fab Construction Is Already Running at Four to Five Times Its Historical Pace and Still Can't Catch Up, What's Really Climbing Inside That 53.3% Was Never the Chips. It Was the Price!
On Thursday, September 10 in New York, TSMC reported August revenue of NT$514.8 billion, up 53.3% year over year and up 10.1% from July — the company's highest month on record, and its fourth straight month of new highs. Nearly every outlet, from Bloomberg to CNBC, ran the same line: AI chip demand is so strong that supply can't keep up. That sentence isn't wrong, exactly, but it quietly collapses two different things into one — how much revenue grew, and how many more chips got built. TSMC's own management, asked recently about capacity, said the company is now building fabs at four to five times its historical pace and still falling behind orders. Meanwhile, contract prices for ordinary server memory rose nearly 100% quarter over quarter in Q1 2026, and forecasts still call for another 10-18% in Q3 — decelerating, but never reversing. And on the very day TSMC posted that record number, its own stock closed down 0.61%. This piece is about what's actually climbing inside that 53.3%: the chips, or the price.
Sep 11, 2026 - 06
Broadcom's AI Chip Revenue Just Grew 221% Year Over Year and the Stock Still Closed Down Over 3% — What Gets Rewarded Was Never the Earnings, It's Whether the Earnings Beat the Number Price Had Already Written In!
The US trading session behind this morning's (Sept 4) Beijing close was Thursday, September 3 in New York: Broadcom reported fiscal Q3 revenue of $29.6 billion, up 86% year over year, with AI chip revenue of $16.7 billion, up 221% year over year and 54% quarter over quarter — a record-shattering quarter by almost any measure. Yet the stock closed down more than 3% that day, because its $34.8 billion fourth-quarter revenue guide came in under $1 billion short of the $35.03 billion consensus, a gap of well under 1%. The same day, the S&P 500 and Nasdaq both rose more than 1%, snapping their own losing streaks. The popular reads say either the market overreacted and this is a dip to buy, or that the AI capex story is cooling. Both are watching the move and missing the ruler that moved it.
Sep 04, 2026 - 07
Nvidia Says Demand Is Higher Than Seventy Percent — That's Not a Verdict, It's the Defendant's Own Confession!
Beijing time's pre-dawn close today tracks Wall Street's Wednesday (Aug 26) session: Nvidia reported after the bell — revenue of $96.2 billion, up 106% year over year — and for the first time ever issued forward guidance spanning into the next fiscal year: roughly 70% revenue growth for fiscal 2028. Jensen Huang was explicit on the call that supply, not demand, is the binding constraint, and that real demand runs higher than that number. Headlines everywhere read this as the verdict that finally settles the AI-bubble argument. But a verdict can never be read out by the defendant. A supply ceiling proves nothing about how many outside buyers want in — it only proves how much one factory can build in a year. The same day, AWS announced an incremental order for 2 million more Nvidia GPUs — an outsider doing its own math and signing its own budget. That is what real demand evidence looks like.
Aug 27, 2026 - 08
Chip Stocks Fall for Seven Days Straight — Not on Fading Confidence, But on Nvidia's Own Check Coming Due!
The US trading session in question is Monday, August 24 (US Eastern) — the close that landed in Beijing in the small hours of today. Nvidia closed lower for a seventh straight session, its longest losing streak since September 2022, down 7.5% cumulatively over the stretch, with the rest of the chip complex sliding alongside it. The popular read is that everyone is simply waiting for Wednesday's earnings (August 26 US Eastern, Thursday morning in Beijing) to settle the question. But the real suspense in that report was never how big the number is — in 2026 alone, Nvidia has announced more than $540 billion in what the market calls "circular" financing deals, more than double its entire prior fiscal year's revenue. Seven days of selling is not the market waiting for an answer. It is the market marking down a check Nvidia wrote to itself.
Aug 25, 2026 - 09
Nvidia Didn't Fall on Demand — It Fell Because the Credit Pipe Finally Showed!
U.S. stocks closed Monday, August 10 — the session that wrapped up here in the small hours. Nvidia announced it was teaming up with six Wall Street giants — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to mobilize over $500 billion in third-party capital for AI infrastructure. The stock fell anyway, down roughly 2 to 3 percent, market value down by the tens of billions on the day. One reading is everywhere: concentration risk — Nvidia fronting money so its own customers can buy its own chips has finally spooked investors. Yet the same day, TSMC posted July revenue up nearly 45 percent year over year — no sign demand is cooling. What the market repriced was never whether the chips would sell. It was what has been growing this whole buildout: cash customers already had, or a credit pipe Nvidia built itself. Once that pipe is out in the open, it can make everyone richer on the way up — and hand everyone the same bill on the way down.
Aug 11, 2026 - 10
Apple's Cost Is Amazon's Revenue!
On the July 31 session that closed early this morning Beijing time, the market wore two faces. Apple cratered more than seven percent after earnings — its worst post-report day in over a decade — while Amazon exploded higher by roughly fifteen. The tape had one reading: Apple has fallen behind, growth has peaked, it missed this AI wave; Amazon caught the cloud-and-compute updraft. One loser, one winner, two separate stories. But that reading mistakes the two ends of a single thing for two things. Apple did not fall because it is weak — on the contrary, it just posted record quarterly revenue. It fell because Tim Cook called it: memory is in a "100-year flood on pricing," and for the next several quarters it will raise prices and ration supply. And where did that flood rise from? From Amazon's own AI cloud, devouring chips and memory day and night. The same scarce chips that book as revenue on Amazon's ledger book as cost on Apple's. Scarcity is one account; revenue and cost are its two sides.
Aug 02, 2026 - 11
Memory Isn't AI's Thermometer, It's the Tip of the Whip!
On Monday July 27 (ET), memory chips led the whole semiconductor complex lower — the flash leader lost a big chunk in a single session, and the three memory giants were dragged into a bear market together. The strange part: days earlier they had posted record revenue and said the high-bandwidth memory that feeds AI accelerators was sold out, booked well into future quarters. Everyone is stuck on one question — if AI hasn't cooled, why did memory crack first? That question treats memory as a thermometer of AI demand, as if its fall measures AI going cold. It never was a thermometer. Memory is a long whip that the AI supply chain cracks, and what broke is the tip — the segment farthest from the hand. The hand moves an inch and the tip travels a yard: lift the hand and the tip shoots to the sky (rising prices, sold out, record revenue); pause it, and the tip is what slams down hardest. That record revenue is not a margin of safety — it is the glut being poured, live. Every plant breaking ground today is a brick in the oversupply of the year after next.
Jul 28, 2026 - 12
Nvidia Isn't Backstopping OpenAI. It's Backstopping Its Own Order Book!
Nvidia fell nearly 5% on Monday, July 27, after reports that it is in talks to provide up to $250 billion in guarantees so OpenAI can lease compute from an Ohio data center project. The reassurance arrived the same day: a guarantee isn't an investment, costs no cash, touches no income statement. Every word of that is true, and it misses the point — it is useful precisely because it costs nothing today. A guarantee does not create demand. It moves the credit risk of demand off the buyer's book and onto the seller's. The upside is a finite gross margin already priced in; the downside is lease principal, payable in cash, and the day it triggers is the day Nvidia can least afford it. The rarer signal that day wasn't in equities: the cost of insuring Nvidia's debt jumped about 20% in a single session. Equity markets ask what a company is worth. Credit markets ask whether it can pay.
Jul 28, 2026 - 13
Capex Isn't a Cost — It's an IOU Waiting to Be Claimed!
This week the market looked like it flipped — Alphabet took a hit after lifting its 2026 capex ceiling toward $205 billion, Tesla fell more than 14% on July 23, the Nasdaq slid 2.15%, and every recap said the same thing: the market has finally had enough of AI spending. Then, hours after that same July 23 close, Intel — which is burying fabs in the ground and spending no less than anyone — reported and rose instead of fell. Same market, same day, opposite verdicts on two nearly identical piles of spending. The difference isn't how much you spent; it's whether a real payer stands at the other end of the chain: Tesla booked record revenue near $28 billion while its operating margin collapsed to 1.4%, the money gone into robots, robotaxi and compute not yet built — a promise at the far end; Intel named its first external 18A customer — an order at the far end. The market never punishes spending. It punishes spending no one has agreed to pay for.
Jul 24, 2026 - 14
Memory Led the Tape — That Is Not the Rally Broadening, It Is One Bet in a New Mask! Paid · $3
At the July 21 US close the Nasdaq rose 1.3%, and the leaders were not the usual AI mega-caps but memory chips — Micron, SanDisk and Western Digital jumped together, and Wall Street exhaled: the rally is finally broadening. But open any bullish note and the reason is the same single word — AI demand. Memory rose because AI wants memory; the mega-caps rose because of AI. The whole "broadening" runs on one engine. This is not the rally finally broadening. It is the same bet, wearing a new mask, handed back to you. The memory squeeze does not live on the demand side — it lives on supply: makers are shaving wafer capacity off commodity DRAM and pouring it into HBM, a share climbing from under a fifth toward a third. You think you bought a cyclical independent of AI. You bought one more layer of leverage on AI capex.
Jul 22, 2026 - 15
300 Global Debuts Launch Not Demand, but Capacity!
The World AI Conference opened in Shanghai today: exhibition space past 100,000 square meters for the first time, over 1,100 companies, more than 3,000 exhibits, over 300 products making their global debut. Every write-up uses the same sentence — unprecedented scale, proof of roaring demand. But a trade show never measures demand. A trade show measures supply. A debut is a decision made by a vendor, not by a customer.
Jul 17, 2026 - 16
Profit Up 711% — You're Not Buying Growth, You're Buying the Worst Risk-Reward There Is! Paid · $3
Earnings pre-announcements are stacking up today, and optical communications is the brightest beam of all — Yangtze Optical guided first-half net profit up 711% to 914% year on year, Dongshan Precision up nearly 300%, PCB leaders locked limit-up in the afternoon. Every screen chants "compute demand is exploding, optical is a secular bull." But the moment growth is fastest is usually the moment risk-reward is worst. When everyone charges in holding that staggering number, what you buy isn't growth — it's a position whose upside is already spent and whose downside has only just opened.
Jul 15, 2026 - 17
AI Capex Repays Depreciation, Not Demand! Paid · $3
The most reassuring number in the market this year is the "confirmed order": the upstream discounts two years of compute demand into a near-trillion-dollar backlog and keeps telling the market that demand is locked, not imagined. Meanwhile another number is spoken of lightly — the industry pours nearly half a trillion dollars a year of capex into AI, and the revenue that actually monetizes on the enterprise side is a fraction of it. Most read the two as one sentence: demand this certain, spending this justified. But pull the time axis one notch longer and something surfaces that every backlog covers up — and it never stops for a day.
Jul 11, 2026 - 18
An IPO Does Not Sell Equity, It Sells Certainty! Paid · $3
SK Hynix landed in New York as ADRs and took away more than $28 billion in a single raise. Chips ripped, the Nasdaq surged, and the market read it as one sentence: AI demand is so hot that even the most upstream link is being fought over. But every raise has two directions — what the buyer buys, and what the seller sells. When the link closest to physics, the one that knows the capacity ramp schedule best, chooses to hand equity to the public at this price, what it sells is not stock.
Jul 10, 2026 - 19
Debt-Funded Expansion Buys You Time, Not Compute! Paid · $3
The five largest cloud and AI giants issued $159B of bonds in five months, 47% more than last year, while their credit spreads quietly widen. The market reads it as "smart use of cheap capital." But expanding on cash you earned is investment; expanding on borrowed money whose interest depends on revenue that has not appeared is discounting a future not yet arrived to plug a hole in the present. Debt cannot buy the return on compute — it can only buy the time you spend waiting for it.
Jul 08, 2026 - 20
Selling Your Compute Is the First Crack in the Compute Faith!
Meta starts renting out its AI compute, and the market reads it as "monetizing idle capacity," a maturing platform business. But a hoarder who suddenly agrees to rent out his compute is not doing it because he has surplus — he is doing it because his own calculator told him the marginal return of holding it has fallen below the market rent. The water-maker has started selling water, and the crack opens from his own hands.
Jul 08, 2026