Created Oct 30, 2026 · Latest reported quarter, ended Sep 30, 2026: GOOGL, AMZN, META (Q3) · MSFT (fiscal Q1 FY27, same quarter end). Lab revenue figures are press-reported run rates, dated where used.
You asked three connected questions after a week of selling in the hyperscalers: if nobody you know has trouble getting AI tools, where is the shortage; why do the builders say they cannot build fast enough; and how does that square with the market punishing the capex? This page answers them in that order, then turns to what would tell you which side of the return-on-capital argument is winning. It is a theme page, so it sits outside the module spine; the companies return in Modules 0016, 0021 and 0025.
When an airline is full you do not see “no flights”; you see a higher fare and a worse seat. Compute is rationed the same way. The provider decides who gets how much, and at what price, before a request ever fails:
Each step trades a visible outage for a less visible cost: a smaller model, a cap, a wait or a higher price. A casual user touches almost none of them.
The second reason you feel no pinch is that chat is a small slice of the load. What changed in 2025–26 is compute per task. A reasoning model “thinks” in tokens before it answers, and an agent that works for an hour on a coding job can use more tokens than thousands of chat replies. So usage can explode while the number of users grows modestly. Alphabet's numbers this quarter show the shape: the Gemini app reached about 1 billion users, but its first-party model APIs were processing nearly 32 billion tokens a minute, up 40% in one quarter (Q3 2026 call).
Read left to right as “is supplied by”. The newest feature (Q3 2026) is the right-hand box: the biggest clouds are short enough to rent capacity from smaller specialists.
Backlog (remaining performance obligations, RPO) is contracted revenue not yet recognized. It is the cleanest public proof that demand exists, because a customer signed for it. It is not proof the customer can pay, which comes later on this page.
| Company | Backlog measure, Sep 30, 2026 | Growth / timing | What management said about supply (Q3 call) |
|---|---|---|---|
| MSFT | Commercial RPO $677B | +73% y/y | “Customer demand continues to exceed available supply.” Some rented (third-party) GPU capacity slipped to H2 FY27. |
| AMZN | RPO about $578B, “primarily related to AWS” | $496B at Jun 30; weighted life 6.1 years | AWS growth accelerated to 39%; capacity reservations stretch into 2027–29 (call, via summaries). |
| GOOGL | Google Cloud backlog $561B | Just over 55% expected within 24 months | “Supply constrained”; began using third-party compute at the end of Q3, with near-term margin headwinds. |
The three measures are not defined identically and should not be summed into one precise number; together they are roughly $1.8 trillion. Concentration matters: Microsoft said in January 2026 that about 45% of its then-$625B RPO came from OpenAI, so a large part of “cloud backlog” is really the labs' commitments passed through.
| Lab | Reported annualized revenue run rate | Basis and caveat |
|---|---|---|
| OpenAI | About $50B (end of September 2026); targets $70B+ by year-end | Its own share of partner sales; an earlier $68B figure included partners' gross revenue (Bloomberg, CNBC, Oct 8) |
| Anthropic | About $65B (end of July 2026) | Gross basis; press reports (Bloomberg via Axios, Aug 17). See the conflict note at the top. |
Two labs at roughly $115B of run-rate revenue, growing several-fold a year, is a remarkable number. It is also measured on different bases, and it sits against roughly $720–740B of 2026 capex at just four hyperscalers (next section). The labs pay their cloud bills partly from revenue and partly from capital raised from investors, which is the heart of the return debate.
A data center needs chips, memory, a building, cooling, and power delivered to the fence. Each has a different lead time. The bottleneck is always whichever is slowest to add right now, and as the fast ones catch up, it moves down the ladder to the slow ones.
Red bars are the constraints binding now. Ranges are rough and drawn from trade press and consultancies (sources at the bottom); treat them as an order-of-magnitude picture, not data.
That is why the builders say they cannot build fast enough while chip supply has eased. Satya Nadella's line from a podcast interview in late 2025, widely reported, was that his problem was not chips but “warm shells” to plug them into. Q3 2026 adds two refinements:
On power, developers have announced around 100 GW of on-site gas generation in the U.S., but only a few GW were running by mid-2026 (RBC, trade press), which tells you how far announcements run ahead of delivery. The grid equipment tangent has the transformer and turbine makers, their backlogs and how much customer cash backs them.
Demand and return on capital are different questions. The market is no longer asking whether demand exists. It is asking whether the spending earns an adequate return before the hardware wears out, and in Q3 the cash flow statements made that question hard to ignore.
Shared scale, 0–80. Combined: about $176B generated, about $202B spent on property and equipment in one quarter. *Microsoft's quarterly operating cash flow is derived as cash capex plus the $18.6B free cash flow reported in secondary summaries of the release; not checked against the 10-Q. **Meta's reported FCF also deducts $1.0B of finance-lease principal. Amazon's gross capex is shown; Amazon's own FCF nets out proceeds and incentives and is reported only on a trailing-twelve-month basis. Alphabet's operating cash flow fell from $48.4B a year earlier; the reason is not covered here.
| Company | Capex definition used | Q3 2026 capex | Year ago | 2026 capex outlook (as of Q3 call) |
|---|---|---|---|---|
| MSFT | Incl. finance leases / cash PP&E | $45.2B / $39.4B | — | ~$170B for calendar 2026 (CFO); over $50B in the Dec quarter |
| GOOGL | Purchases of PP&E | $51.4B | $24.0B | $195–205B, unchanged; “increase significantly” in 2027 |
| AMZN | Gross purchases of PP&E | $77.3B | $35.1B | ~$220B cash capex; 2027 “meaningfully more” (call, via summaries) |
| META | Incl. finance-lease principal | $34.9B | $19.4B | $135–145B (low end raised from $130B); 2027 “continue investing aggressively” |
The four outlooks add to roughly $720–740B, but on mixed definitions (calendar vs fiscal, with and without leases), so treat the sum as a scale marker only.
Terms are as announced in 2025 and may have been revised since. The loop does not make the demand fake: the end customers using the labs' products are real. It does mean that one lab's funding round shows up as revenue at several listed companies at once, and that a funding problem would travel the same circle in reverse.
| Fiber build, 1996–2001 | AI build, 2023– | |
|---|---|---|
| Was demand real? | Yes, internet traffic did grow enormously | Yes: backlog, token volumes, lab revenue |
| Who built? | Leveraged new entrants (Global Crossing, WorldCom, Level 3) | Mostly cash-rich incumbents, plus leveraged neoclouds and Oracle |
| Utilization during the build | Very low: only a small single-digit share of fiber was lit | High: capacity is sold out; no “dark GPUs” yet |
| Asset life | Decades: idle fiber was later lit and used | Years: idle GPUs become obsolete, not merely early |
| How it ended | Price collapse, bankruptcies; the eventual users won | Unknown |
The lesson from fiber is that being right about demand and earning a return are separate bets: the forecasts came true, and most of the builders' shareholders lost anyway. Two differences cut in opposite directions. Strong balance sheets and full utilization make a fiber-style collapse less likely now. Short asset lives mean that if an overbuild does come, it would show up quickly as write-downs and falling rental prices rather than as slowly lit capacity.
Pricing power goes to whoever controls the scarce input at the time. As the bottleneck moves, so does the margin, while the return-on-capital risk stays with whoever owns the buildings and the chips.
| Layer | Where the constraint is | Listed examples | In this curriculum |
|---|---|---|---|
| Accelerators and networking | Easing; still premium-priced for the newest generation | NVDA, AMD, AVGO | Modules 0016, 0017 |
| Memory and chip equipment | Binding (HBM, DRAM) | MU, SK hynix, Samsung; LRCX, AMAT, ASML | Module 0017; HBM tangent |
| Servers and racks | Volume, low margin | CLS, DELL | Modules 0009, 0018 |
| Clouds and neoclouds (capital at risk) | Short of shells and power; carry the depreciation | MSFT, AMZN, GOOGL, ORCL; CoreWeave, Nebius | Modules 0016, 0021, 0025 |
| Power equipment | Binding (turbines, transformers) | GEV, Siemens Energy, ETN, Hitachi | Modules 0027, 0028; grid tangent |
| Generation, fuel and grid | Binding (connections, firm power) | CEG, NEE; gas pipelines; uranium | Module 0033; Lesson 0005 (gas to power); Lesson 0006 (Cameco, Westinghouse) |
| Infrastructure capital | Funds the buildings and power off the hyperscalers' balance sheets | Brookfield (BN, BAM, BIP) | Lesson 0003; Module 0013 |
A map of where each company sits, not a list of picks. The same name can be a winner in one scenario and the most exposed in another: equipment makers with customer deposits are better protected than builders with long leases if spending slows.
| Signal | Bull reading | Bear reading | Where to find it |
|---|---|---|---|
| Cloud backlog vs capex growth | Backlog grows faster than capex; less of it from one lab | Backlog stalls, or concentrates further in OpenAI/Anthropic | 10-Q RPO notes; call commentary |
| GPU rental prices (per GPU-hour) | Newest chips hold price; older ones decline slowly | Fast price drops on recent generations | Rental indexes; currently contested (one tracker flat over 30 days, others disagree on direction) |
| Token pricing | Holding or rising (Microsoft's Q3 comment) | Price wars outpace volume growth | Provider price lists; call commentary |
| Lab revenue | Run rates keep compounding and convert to cash | Growth slows while commitments stay fixed | Press reports; any IPO filings |
| Capex ÷ operating cash flow | Peaks and falls as revenue catches up | Stays above 100%, funded with debt | Cash flow statements |
| Depreciation as % of revenue | Rises but margins hold | Useful lives shortened; write-downs | Property notes in 10-K/10-Q |
| Rented-capacity delays | Delays clear in H2 FY27 as Microsoft guided | More slippage; neocloud financing strain | Calls; neocloud filings |
| Grid and turbine queues | Connections accelerate; on-site generation delivers | Announcements keep outrunning delivery | ERCOT/PJM queue data; GEV and Siemens Energy orders |