中文EN
← Deep Research
Deep Research · Plain-language

Is the AI Build-Out Another 1999? What Rhymes, What Doesn't, What Not to Believe (Plain-Language Edition)

This is the plain-language edition · read the deep dive (full arguments & sources) →
TL;DR
$710 billion a year of AI build-out against tens of billions of visible AI revenue — ten to one, though how far apart depends on the arithmetic. 1999's real disease was fake demand (the traffic myth inflated reality tenfold, spread entirely by sellers), fake accounting (WorldCom), and debt — not "building infrastructure"; and the infrastructure surviving didn't save that era's shareholders. What rhymes now: the demand story told only by sellers, suppliers financing their own customers, money shifting from profits to debt and off-balance-sheet structures. What doesn't: the funders are the most profitable companies in history, revenue is genuinely growing, and the scale is about half of dot-com. Three gauges: where the money comes from (debt), how long the chips live (depreciation — Amazon has officially conceded AI shortens equipment life), and demand quality (wait for independent data).
traffic myth: 100 days vs a yearcapex eats 90% of cash flow$100B circular deal: unsignedAmazon concedes AI shortens lives

This is the condensed version of the deep dive of the same name. Every key number was independently checked; for the full argument and sources, read the deep dive. Figures as of July 2026.

An Arithmetic Problem That Grew Sevenfold in Three Years

In 2023, David Cahn of the venture firm Sequoia posed an arithmetic problem: the money the industry spent that year on AI chips and data centers would need $200 billion of AI revenue to pay back — and visible AI revenue came nowhere close. In 2024, the same problem's answer became $600 billion. By July 2026: $1.5 trillion.

So everyone thought of the same thing: the telecom bubble of 1999 — also a frenzy of infrastructure building (fiber optics, that time), also "the demand is just about to arrive," and then a crash: bankruptcies, shareholders wiped out. Is the AI build-out a rerun?

This article first gets 1999 right (it is misremembered more than you'd think), then lays out this cycle's books, then compares line by line: what rhymes, what doesn't, and which popular talking points neither side should trust.

Cahn's three questions: gross requirement and net hole are two different series Algorithm: Nvidia data center run-rate revenue ×2 (GPUs ≈ half of TCO) ×2 (50% end-user margin) 2023-09 "$200B Question"$200Bnet hole $125B2024-06 "$600B Question"$600Bnet hole ~$500B2026-07 "$1.5T Question"$1.5T (~$3T cumulative)no net-hole figure lifetime revenue required by one year's capex net hole after big tech's AI revenue
Schematic: $200B→$600B→$1.5T is the gross-requirement series; $125B→$500B is the net hole — reading $1.5T as "the hole" is a common error. Cahn calls no bubble, and in 2026 says his GPU-only math increasingly underestimates the denominator (TPUs/ASICs, memory)

What Actually Happened in 1999

First: the demand was made up. In 1999 everyone "knew" internet traffic doubled every 100 days, so fiber could never be overbuilt. A researcher at AT&T Labs, Andrew Odlyzko, debunked it with real data at the time: traffic actually doubled about once a year — the myth inflated demand more than tenfold, and the people spreading it were the ones selling fiber and bandwidth. The result was massive overbuild: only a single-digit share of the fiber laid ever lit up.

Second: where demand fell short, accounting stepped in. WorldCom dressed up day-to-day expenses as capital investment (about $3.5 billion of that trick, over $11 billion of total overstatement) and went bankrupt in 2002 — then the largest bankruptcy in US history. Qwest inflated revenue by over $3.8 billion. Equipment makers lent customers money to buy their own gear (Lucent committed an $8.1 billion credit line, actually disbursed $2.1 billion, and much of it went bad) — manufacturing their own "demand" with their own money.

Third: the infrastructure survived, but the people who paid for it lost their shirts. The fiber glut did later make Netflix and YouTube cheap to run, just as the railway network left by Britain's 1840s railway mania eventually turned a profit — but railway shares fell about 60%, telecom stocks lost about $2 trillion, and roughly half a million people lost their jobs. "The stuff survived" and "the investors got paid" are two different things — the hardest lesson of every infrastructure bubble in history.

Two 1999 gauges: demand myth vs measurement, and the capex round trip Myth: "doubling every 100 days" = 700–1500%/yrsellers' yardstickMeasured: doubling ~annually = 70–150%/yr (Odlyzko)independent dataSWITCH link 1996-2001: capacity +144%/yr vs traffic +87%/yr → utilization fell structurallyUS communications equipment investment (1996 $B/yr, Richmond Fed)$62B1996Q1$135B2000Q4$93B2001Q4
Schematic: the demand myth inflated real growth ~10x, spread by those selling fiber and bandwidth; equipment investment ran $62B→$135B→$93B (69% of a year earlier). Today's parallel: token counts and order visibility are likewise supply-side-produced, with independent measurement absent

This Cycle's Books: $710 Billion a Year vs Visible AI Revenue

The four big cloud companies (Microsoft, Google, Amazon, Meta) plan combined capital spending of about $710 billion for 2026 — and every one of them keeps raising the number.

How much visible AI revenue is there? Microsoft says its "AI business" runs above $37 billion a year; Amazon says $15 billion; OpenAI runs at about $25 billion; Anthropic says it reached $47 billion. Sounds like a lot, but three problems: each company defines its own unaudited metric, and they can't be added up; the ~$17 billion OpenAI paid Microsoft for computing also shows up inside Microsoft's AI revenue — the same dollars counted twice; and even summed, it's tens of billions against $710 billion of spending — an order of magnitude apart.

But "how far apart" depends on the arithmetic. The spending is paid up front, while the equipment lasts years — like comparing one year's salary to a house's full price. Measure revenue against annual depreciation instead and the gap shrinks a lot. The bubble callers and their critics often aren't arguing about facts. They're arguing about denominators.

One macro number worth memorizing: in the first half of 2025, 92% of US GDP growth came from information equipment and software investment — just 4% of the economy. The AI build-out is carrying American growth right now. Which also means: the moment it decelerates, growth is exposed.

The 2026 ledger: capex vs visible "AI revenue" (incompatible yardsticks) Big-four 2026 capex guidance, company midpoints$710BAnthropic run rate (May 2026, self-reported)$47BMicrosoft "AI business ARR" (FY26Q3)$37BOpenAI run rate (end of Jan 2026)$25BAWS "AI revenue run rate" (first 3 years)$15B The four revenue bars are vendor-defined, unaudited, and non-summable; OpenAI's ~$17B paid to Microsoft is also inside Microsoft's AI revenue (double counting)
Schematic: even on the most aggressive yardsticks, visible AI revenue sums to tens of billions against $710B of single-year capex — roughly an order of magnitude, the factual floor under every gap calculation; how big the "gap" is depends on the denominator (full capex vs annual depreciation), see text

Dispute One: Where Is the Money Coming From?

Three years ago these companies built data centers out of their own profits. That no longer covers it: in 2026 the big four's capital spending is on track to eat about 90% of their operating cash flow (65% in 2025). So the borrowing and structuring began. Meta put a $27 billion data center into a joint venture so the debt stays off its own books (while quietly guaranteeing its resale value for 16 years). Oracle's cash flow has gone negative; it borrowed $43 billion in a year, and the price of insuring its debt hit an all-time high in March 2026 — higher than during the 2008 financial crisis. Loans collateralized by GPUs went from a 15%-interest fringe product to investment grade.

The Bank for International Settlements (the central bank of central banks) summed up the trend in January 2026 in one phrase: AI financing is moving "from cash flows to debt." It also flagged something strange: lenders charge AI companies almost the same interest as everyone else — so either the lenders are underestimating the risk, or the stock market is overestimating AI's future. At least one of them is wrong.

The most 1999-like piece is the "circular deals": NVIDIA announced up to $100 billion of investment in OpenAI (which turns around and buys NVIDIA chips) — and nearly three months later, NVIDIA's own CFO admitted no definitive agreement had been signed. AMD went further: it granted OpenAI stock warrants at one cent a share — the more chips OpenAI buys, the more AMD stock it earns. Suppliers paying to manufacture their own demand: a modern remake of exactly the game the equipment makers played in 1999, before it blew up.

The circular-deal map: announced amounts vs status (as of 2026-07) up to $100B investment · no definitive agreement$300B cloud contract · signed (2027-31)up to $5Bup to $10B$30B Azure purchase commitment160M warrants at $0.01 ↔ 6GW purchasesNVIDIAOpenAIOracleMicrosoftAnthropicAMD dashed = announced/LOI (no definitive agreement) solid = signed
Schematic: supplier balance sheets subsidizing customer demand — the modern variant of Lucent's 1999 "committed $8.1B, disbursed $2.1B"; dashed announcements should not count as demand evidence. Oracle's contract is signed, but the counterparty is a single loss-making customer

Dispute Two: How Many Years Does a Chip Last?

Buried fiber lasts 20 years. GPUs? In recent years the big clouds stretched server depreciation from 3-4 years to 5-6 — the equipment became "more durable" on paper, and profits looked better for it (about $3.7 billion a year extra for Microsoft, $3.9 billion for Google). The short seller Michael Burry calls this window dressing: GPUs really become obsolete in two or three years, and the giants are hiding the bad news.

Who's right? Neither side's evidence is clean. Burry's specific numbers sit behind a paywall and nobody has independently verified them; the bulls' favorite counter — "old chips still rent out at nearly full price" — comes from a company whose valuation depends on GPU resale values, and our audit step struck it down. But two facts are hard: the A100 chip, launched in 2020, was still actively renting in 2026 at roughly $1-2 an hour (multiple independent sources) — old cards haven't gone to zero; and in 2025 Amazon, in its official filings, moved some servers back from six-year to five-year depreciation, citing in writing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning" — a giant conceding, on the record, that AI shortens equipment life. Worth watching: whether the others follow.

Server depreciation: the collective stretch, and Amazon's reversal Amazon: 3→4→5→6 years (2020-2024)6Amazon 2025: subset back to 5 yrs, citing "pace of AI"5its own 10-K/10-QMicrosoft: 4→6 years (from FY2023)6profit +$3.7BAlphabet: 4→6 years (from Jan 2023)6depreciation −$3.9BMeta: stepwise to 5.5 years (from Jan 2025)5.5est. −$2.9B (2025) depreciation life (years)
Schematic: longer lives materially lift reported profits (factual layer, primary 10-Ks); whether they should is the disputed layer — Burry claims 2-3-year GPU economic life (accusation-grade, math paywalled), while Amazon's AI-citing reversal is the only audit-grade internal confirmation. Watch whether the others follow

Dispute Three: Is the Demand Real?

This is where the analogy breaks hardest. 1999 was "built and nobody came." Now Microsoft's CFO says, in her own words, "I have been short now for many quarters. I thought we were going to catch up. We are not." Google says cloud revenue would be higher if it could meet demand. Chips sit in inventory because of a shortage of powered buildings, not a shortage of buyers. And revenue really is growing: Google Cloud up 63% in a quarter; Anthropic's annualized revenue more than tripled in three months.

But apply the discount: nearly all of this evidence comes out of the sellers' own mouths. "Sold out" comes from CFOs defending their spending; "tokens processed grew 7x in a year" comes from Google, and counts the AI it pushes into every search whether you asked or not. The lesson of 1999 is exactly this: "doubling every 100 days" was also unanimous — and everyone saying it was selling something. The one person with independent data was right. This cycle's independent measurer hasn't shown up yet.

So Is It 1999? The Scorecard

What rhymes: the demand story is told entirely by the sellers, with no independent audit; suppliers financing their own customers (circular deals); funding shifting from profits to debt and off-balance-sheet structures; accounting conventions (depreciation) becoming a major profit lever; a whole country's growth riding on one trade.

What doesn't: the money comes from the most profitable companies in corporate history, not 1999's never-profitable telecom upstarts; revenue is materializing fast, with genuine shortages; the total is about 1% of GDP — roughly half the dot-com tech-investment peak (the central bankers' math); and there is no proven fraud — the depreciation fight is a dispute over yardsticks, not a conviction.

What not to believe: "the $1.5 trillion hole" (that's gross required revenue, not a net gap — and its own author doesn't call it a bubble); "the divergence is 46%, worse than 2001's 32%" (that 32% has no source anywhere in the original report); and "the infrastructure will survive anyway, so the bubble is worth it" (surviving ≠ you getting paid — and nobody has ever seriously quantified the bubble's leftover benefits).

One sentence: not yet 1999 in size, better than 1999 in demand quality — but the money's origins are turning 1999-like at visible speed, and "the demand story is told only by the sellers" is the deepest rhyme of all.

Scorecard: what rhymes with 1999, what doesn't, what both sides skip Rhymes (structural)demand narrative supply-side-produced, unauditedmodern vendor financing (announced ≫ deployed)financing: cash flow → debt/off-balance-sheet (BIS)accounting as a profit lever (depreciation fight)macro concentration: 4% of investment, 92% of growth Doesn't (substantive)funders = strongest balance sheets ever, not CLECsrevenue materializing fast, supply-constrainedscale ~1% GDP, about half of dot-com (BIS)GPUs are 3-6-yr assets vs fiber's 20no proven fraud (depreciation is a yardstick fight) What both sides skip Bulls: infrastructure surviving ≠ investors paid; the rigidity of 90% OCFBears: the gap shrinks under a depreciation yardstick; BIS's ~1%-of-GDP cooling noteBoth misquote: Allianz's 32% comparator has zero methodological support
Schematic: the verdict is not a yes/no but a set of monitorable sliding variables — not 1999 in scale, better in demand quality, converging on late-1999 in financing structure

How to Check This Article's Judgments

The deep dive's twelve claims, rewritten in plain words, ordered from most to least confident:

  1. Spending exceeds visible AI revenue by roughly ten to one — check any quarter's earnings reports.
  2. The money has shifted from profits to borrowing — watch the big four's cash-flow coverage and the central banks' follow-up reports.
  3. Lenders aren't charging AI extra for risk — watch the interest spread: when it widens, the market has started to worry.
  4. Circular deals are announced big, delivered small — watch whether NVIDIA-OpenAI's $100 billion ever gets a signed definitive agreement.
  5. Stretched depreciation flattered profits, and Amazon has already reversed course — watch whether the others shorten equipment lives within a year (if they do, the skeptics' direction is confirmed).
  6. The five giants have signed about $662 billion of data center leases that haven't started yet — we rebuilt this number ourselves from five official filings; watch how fast it grows.
  7. 1999's disease was fake demand + fake accounting + debt, not "building infrastructure" — a historical conclusion; no further test needed.
  8. This cycle's demand evidence all comes from the sellers — watch for independent third-party utilization data; if it appears and contradicts them, that's the bubble case's strongest signal yet.
  9. Oracle is this cycle's stress point (negative cash flow + record debt-insurance prices + one giant customer) — watch its CDS and OpenAI's payments.
  10. US GDP growth is riding on the AI build-out — watch the GDP breakdown in the first quarter capex slows.
  11. "The infrastructure will survive" won't save investors — an iron law of history; don't use it as a reason to buy.
  12. If the 1999 script runs, the middle layer (data center builders, GPU landlords) fails first, not the app companies — a theoretical inference, our least confident claim, but watch where the first defaults land.

The Things That Matter Most