Is the AI hype
cooling down?
A composite reading of physical build-out and platform economics. Higher means the expansion is still intensifying; lower means the momentum is fading. Here's what goes into it.
What the answer is built from
Five independent signals, each weighted by how directly it reflects real, committed demand. Each is scored from 0 (contracting) to 100 (surging).
The trajectory
Where the money and the megawatts are actually going.
Combined infrastructure spend, per quarter
Billions of dollarsSpend by operator
$B / quarterNew data-center power queued
Gigawatts / yearTotal new power capacity queued
Gigawatts / yearWhere data-center load is landing
GW in queue, top regionsThe latest print
The charts above still run through the June quarter: the combined series only advances once all five operators have filed, and three of them report the September quarter this week, Alphabet and Microsoft on Oct 28 and Amazon on Oct 29. The prints already in hand do not point to a cooldown, but they do show the bill arriving. Oracle opened the season on Sep 10 with cloud-infrastructure revenue up 121% and a $664B backlog, paid for with $28.5B of capex in a single quarter, negative free cash flow and $20B of newly sold stock. SpaceX followed on Oct 20: revenue up 161%, its AI segment now the largest of the three, $19.6B of quarterly capex and a guide to $35–40B more in the fourth quarter. Then Meta on Oct 21, the clearest sign yet that the build is overtaking the cash: revenue up 26% to $64.6B, capex up 80% to $34.9B, operating margin halved to 20%, and free cash flow negative for the quarter, down from $10.6B a year earlier. The shares fell 7% the next day. Between the three, the latest quarter's capex already tops $83B before Alphabet, Microsoft and Amazon, who spent $135B between them in the June quarter, have said a word. Their June-quarter tiles stay below until the new numbers land.
Total revenue
$64.6B
+26% vs last year
Capex, this quarter
$34.9B
+80% · incl. finance-lease principal
2026 capex guidance
$135–145B
floor raised again, from $130B
Operating margin
20%
down from 40% · incl. a $9.9B legal charge
Free cash flow
−$2.1B
down from $10.6B a year ago
Shares, next day
−7%
the cash burn, not the revenue, moved the stock
Meta's September quarter is this page's thesis in a single print. Revenue grew 26% to $64.6B, but $34.9B of capex (purchases of property and equipment plus finance-lease principal) outran $32.8B of operating cash flow, so free cash flow was negative $2.1B against $10.6B a year earlier. Operating income fell 37% to $13.0B and the margin halved to 20%, though $9.9B of that is a one-time charge for a settlement with 52 state attorneys general; the year-ago net income and EPS were themselves depressed by a one-time tax charge, so the headline +264% net income and $3.83 diluted EPS are not clean growth rates. Management narrowed 2026 capex to $135–145B, lifting the floor for the second time this year, gave no 2027 figure beyond a promise to "continue investing aggressively", and said Reality Labs losses will narrow to about $15B next year as spend shifts to AI. Headcount is down 15% year over year at 66,775. Shares fell 7% the next day. Source: Meta Q3 2026 earnings release (quarter ended Sep 30, 2026), reported Oct 21, 2026.
Total revenue
$14.6B
+161% vs last year · nearly double Q2
SI (AI) segment revenue
$7.6B
+830% · now its largest segment
Capex, this quarter
$19.6B
Q4 guided to $35–40B
Income from operations
$2.8B
from a $273M loss a year ago
Adjusted EBITDA
$8.8B
+306% vs last year
Starlink subscribers
13.6M
up from 7.2M · $5.2B segment rev, +69%
SpaceX's second report as a public company dwarfed its first. Revenue nearly doubled quarter on quarter to $14.6B, the company swung to $2.8B of operating income and $4.4B of net income, and the AI segment, now called Super Intelligence (SI) and including Cursor, acquired on Aug 14 at a $60B implied equity value, booked $7.6B of revenue, more than Starlink. Nameplate compute reached 1.7 GW with 2.5–3 GW targeted by year-end, backlog stands at $63B, and the first-quarter 2027 revenue guide is $33–37B. The capex line is the one to watch: $19.6B in the quarter, $48.1B over nine months, a fourth-quarter guide of $35–40B, and within days Musk said the company had closed $40B of debt for GPU purchases. The stock rose 8% on the print and gave it all back the next day. Source: SpaceX Q3 2026 earnings release and 8-K (quarter ended Sep 30, 2026), reported Oct 20, 2026, plus the earnings call for the guidance.
Total revenue
$19.3B
+30% vs last year
Cloud infrastructure revenue
$7.4B
+121% vs last year
Backlog (RPO)
$664B
up $209B vs last year
Capex, this quarter
$28.5B
+235% · up from $8.5B
Free cash flow
−$5.4B
−$28.7B over the trailing year
FY2027 capex guidance
$90–95B
at most $70B net of customer prepayments
Oracle is the live case for the funding section below. Cloud-infrastructure revenue more than doubled to $7.4B and more than $30B of new AI contracts lifted the backlog to $664B, about half of which management expects to turn into revenue within 36 months. The price of that growth is $28.5B of capex in one quarter against $23.1B of operating cash flow, free cash flow of negative $5.4B (negative $28.7B over the trailing year), $20B of new stock sold through an at-the-market program during the quarter, and a credit rating one notch above junk after a July downgrade by S&P. The stock opened up 7% on the print and closed lower, and by late October it had fallen a further 16% as roughly $18B of loans tied to an Oracle-leased New Mexico campus traded below par. Source: Oracle fiscal Q1 2027 earnings release (quarter ended Aug 31, 2026), reported Sep 10, 2026, and the earnings call for the FY2027 capex guidance.
Total revenue
$119.8B
+24% vs last year
Google Cloud revenue
$24.8B
+82% vs last year
Cloud operating income
$8.8B
up from $2.8B a year ago
Capex, this quarter
$44.9B
~+100% vs last year
Operating income
$40.8B
+30% · 34% margin
2026 capex guidance
$195–205B
raised from $180–190B
One caveat on Alphabet's headline: reported net income leapt to $112.1B and diluted EPS to $9.11 (both up roughly 295%), but that is flattered by a one-time $99B unrealized gain on equity investments, booked below the operating line. Operating income ($40.8B, +30%) is the clean read. Source: Alphabet Q2 2026 earnings release (quarter ended Jun 30, 2026), reported Jul 22, 2026.
Total revenue
$90.0B
+18% vs last year
Intelligent Cloud revenue
$39.3B
+32% vs last year
Azure revenue growth
+43%
past $100B annual run-rate
Capex, this quarter
$41B
incl. finance leases
Commercial backlog (RPO)
$678B
+84% vs last year
Next-quarter capex guide
>$50B
still climbing
Azure crossing a $100B annual run-rate and the $678B backlog are the demand signals; the $41B of quarterly capex (cash for property and equipment $35.8B plus $5.6B of finance leases) is the spend. GAAP diluted EPS of $4.81 (+32%) includes a $3.2B gain from marking up Microsoft's Anthropic stake, worth about $0.33. Management guided FY2027 capex to roughly $175B and over $50B next quarter. Source: Microsoft FQ4 FY2026 earnings release (quarter ended Jun 30, 2026), reported Jul 29, 2026.
Total revenue
$200.6B
+20% vs last year
AWS revenue
$42.2B
+37% · fastest since 2021
Capex, this quarter
$54.2B
+68% vs last year
Operating income
$27.5B
+43% · 14% margin
2026 capex guidance
~$220B
raised from ~$200B
Free cash flow (TTM)
−$7.6B
outflow as the build outpaces cash
Amazon's $62.6B net income and $5.75 diluted EPS are flattered by a $53.4B pre-tax gain on its Anthropic stake, booked below the operating line; operating income ($27.5B, +43%) is the clean read. The $54.2B of quarterly capex is purchases of property and equipment, and AWS growth of 37% ($42.2B) was its fastest since 2021. Andy Jassy raised full-year 2026 capex to about $220B, citing surging memory prices, and said even that will not meet the demand Amazon is seeing. Source: Amazon Q2 2026 earnings release (quarter ended Jun 30, 2026), reported Jul 30, 2026.
Funding the build-out
The spending is enormous, but is it being paid for out of pocket or on credit? The cash-flow statements answer that. Between them the five operators throw off about $710B of operating cash a year and are now plowing roughly 85% of it straight back into capex. The giants still generate tens of billions in free cash flow after they spend; but at the frontier the build has overtaken the cash, and Amazon and Oracle are now spending more than they earn, filling the gap from reserves and debt. Meta, the first to report the September quarter, saw its trailing capex pass $100B and its free cash flow shrink by nearly a third in one quarter.
Operating cash flow, capex & free cash flow, combined
$B · fiscal years 2020–2025Summed across all five operators. Operating cash flow keeps hitting records, but capex is accelerating faster: combined free cash flow peaked near $246B in 2024 and slipped to about $198B in 2025 even as operating cash flow set a high, and the latest snapshot (~$109B TTM) extends that slide.
Snapshot view: trailing twelve months from the operators’ cash-flow statements. History view: annual fiscal-year figures combined across the five (2021–2025 from filings via stockanalysis.com, 2020 from company 10-Ks). Free cash flow here is simply operating cash flow minus purchases of property and equipment; companies’ own free-cash-flow measures differ and can fall on either side of the simple figure shown here: Amazon reports a smaller trailing-twelve-month outflow (about −$7.6B) because it nets proceeds from equipment sales and incentives out of capex, whereas counting finance-lease payments would push the figure the other way. Amazon’s capex also mixes AWS with retail logistics. Fiscal years are offset for Microsoft (ends June) and Oracle (ends May); snapshot periods end Jun 30, 2026 for Microsoft, Alphabet and Amazon, Aug 31, 2026 for Oracle (its fiscal Q1 2027) and Sep 30, 2026 for Meta, the first of the five to report the September quarter. Illustrative, not investment advice. Sources: company 10-Q / 10-K cash-flow statements via stockanalysis.com (2020–2026).
The microscalers
The hyperscalers are not the only ones building. A second tier has grown up beneath them: specialist GPU clouds like CoreWeave, Nebius, Crusoe and Lambda that do essentially one thing, rent out Nvidia accelerators at scale. Call them microscalers, or neoclouds. They own no consumer platform and throw off almost no cash flow of their own, so they borrow against the chips themselves and pre-sell the capacity on multiyear contracts, then plow the proceeds straight back into more chips. Their capex curve is the steepest on this page, and their customer lists read like a directory of the AI labs, plus the hyperscalers who still cannot build fast enough on their own. In October 2026 the market began to price that leverage: CoreWeave lost about a quarter of its value in a week as yields on its bonds climbed toward 15%, days after it had priced $3.7B of convertible notes.
Capex: the neoclouds and the converted miners
$B · 2023–2026 · dashed = estimatedCoreWeave alone still out-spends the entire rest of the field, and its curve is steeper than any single hyperscaler's early ramp: $14.9B of capex in 2025 and a 2026 guide of $35–39B, raised twice this year. Nebius has gone from $4.1B in 2025 to a $20–25B plan for 2026, with $8.1B spent in the first half alone. The converted miners (IREN, Applied Digital, Core Scientific, TeraWulf, Cipher) enter only in 2025, when their AI pivots began, and are drawn fully dashed because their numbers are total company capex that blends mining and AI, much of it customer financed, on fiscal years that do not line up with the calendar: IREN's 2026 point is its fiscal year to June 2026 ($4.3B), and it has since guided $25–30B for the year to June 2027. Solid lines are CoreWeave and Nebius reported figures through 2025; every dashed point is guidance or an estimate. The private neoclouds (Crusoe, Lambda, Together) still do not disclose capex, so they are not plotted.
Who they build for
This is the part that ties the neoclouds back to the frontier. The capacity is not speculative: most of it is pre-sold to named AI labs and to the hyperscalers themselves, who rent burst compute from the neoclouds rather than wait on their own construction.
The converted miners
A whole cohort of former Bitcoin miners has worked out that its real asset was never the hashrate, it was the powered land: gigawatts of grid interconnection and cheap energy that now take years for anyone else to permit. Most of them do not build a cloud of their own; they lease that capacity to the neoclouds, hyperscalers and, increasingly, the labs themselves, with Google, Nvidia or AMD backstopping the financing or taking equity. It is the same chokepoint the "shells, land and interconnection" card further down tracks, seen from the supply side. IREN is the exception that went all the way and now runs its own GPU cloud.
Two names sit behind almost everything here: OpenAI on the demand side and Nvidia on the supply-and-financing side, the same loop the funding section flags, since Nvidia invests in the neoclouds that buy its chips to serve the labs it also backs. The converted miners add a second version of that loop, with Google and Nvidia backstopping the leases and taking equity in the very hosts they rent from. Anthropic, once the conspicuous absence here, became one of the tier's biggest tenants in 2026: a multibillion-dollar CoreWeave deal, a 401 MW TeraWulf campus in Kentucky, and reported leases at Riot and a $35B commitment to Lambda, on top of its Google TPU and Amazon Trainium capacity.
Microscaler here means a specialist GPU cloud (also called a neocloud), distinct from the hyperscalers above. Capex: CoreWeave and Nebius from company filings and guidance. CoreWeave's 2025 figure is $14.9B on its own definition, which counts financed equipment ($10.3B on a cash basis), and its 2026 point is the $35–39B guide; Nebius spent $4.1B in 2025 and plans $20–25B in 2026. The 2026 points are guidance and are drawn dashed. Crusoe, Lambda and Together are private, so no capex line is shown for them; their entries reflect disclosed funding rounds and announced or reported customer deals. Converted-miner figures are from company filings and announcements: IREN's ~$9.7B Microsoft and $3.4B Nvidia cloud deals, Applied Digital's ~$11B CoreWeave leases at Ellendale, Core Scientific hosting CoreWeave and now AMD after its holders rejected CoreWeave's buyout (Oct 2025), the Google-backstopped Fluidstack leases at TeraWulf and Cipher, and the Anthropic and AWS leases signed since. The converted miners' capex points on the chart above are total company capex on each company's own fiscal year where reported (IREN to June 2026, Applied Digital to May 2026) and editorial estimates from first-half 2026 run-rates otherwise, not AI-only figures; much of that spend is customer financed. Contract and backlog figures are the parties' announced numbers, are multi-year lifetime totals rather than annual, and can include options and later phases. Illustrative, not investment advice.
The picks and shovels
The hyperscalers get the headlines, but the chipmakers, memory houses and foundries behind them are in a capex super-cycle of their own, together laying out around $175B a year on current plans to build the fabs, HBM lines and tools the AI build-out physically runs on. TSMC alone has lifted its 2026 budget to $60–64B, Micron doubled its capex in a year, and SK Hynix plans to spend half again as much as it did in 2025. Trace that capital back through their filings and you can see exactly how the build-out got here.
The multi-year climb in chip capex
Annual capex, $B · 2019–2026 · dashed = guidedOne line per filing trail (10-K / 20-F / 6-K). Memory (SK Hynix, Micron) shows the deep 2023 downturn cut and then the HBM-driven surge, with Micron's fiscal 2026 spend double the year before; TSMC keeps setting records and has budgeted $60–64B for 2026; Intel, the one line that bent down through 2025, turns back up on a budget of more than $20B funded partly by a $20B stock sale. Samsung publishes no capex guidance, so its line stops at the 2025 actual.
Annual capital expenditure compiled from company filings (10-K / 20-F / 6-K and quarterly reports); points through 2025 are reported actuals, converted to USD at period rates (won figures at roughly 1,420 to the dollar). The 2026 points are current guidance, drawn dashed: TSMC’s $60–64B budget, SK Hynix’s “high-40 trillion won” plan and Intel’s “more than $20B”. Micron’s 2026 point is a reported actual because its fiscal year ended on Sep 3, 2026. Fiscal years and treatment differ (Intel’s figures are gross additions, before the partner offsets and CHIPS grants that lower net capex materially, to $11.2B in 2025). “Samsung” is its semiconductor (DS) division. CXMT only began reporting publicly with its July 2026 Shanghai listing, so its bar is the roughly 34.9B yuan of fab and test projects it announced on Sep 29, 2026 rather than an audited capex history, and it is absent from the chart above for the same reason. Two kinds of company carry small capex by design and should not be read as small spenders: equipment makers, whose revenue is the mirror of everyone else’s capex, and fabless designers (AMD, Broadcom, Nvidia), who buy foundry capacity instead of building it, so R&D is the figure that matters and their volume lands on TSMC’s capex line. Illustrative, not investment advice. Sources: company filings & investor releases (2019–2026).
The sovereign money
It isn't only US hyperscalers writing the checks anymore. Governments and sovereign wealth funds now treat AI compute as strategic infrastructure, and have pledged hundreds of billions to build it, buy into it, or avoid being left behind. Here's who's committing what, and how each region's playbook differs.
Headline announcements, not deployed cash, a mix of state programs, sovereign wealth funds and public-private ventures, so they aren't strictly comparable (European figures overlap; * South Korea's cluster runs to 2047). Illustrative, not investment advice. Sources: government announcements, Global SWF, IEA and press reporting (2024–2026).
Where it could actually jam
Demand is only half the story. Even with the money flowing, the build-out slams into hard limits, from the accelerator itself (a market chokepoint as much as a physical one), through raw silicon and the machines that pattern it, packaging, passives, optics and the network fabric, out to water, power, the skilled trades who build it, and the export politics that can throttle any link overnight. Here's how binding each chokepoint becomes year by year through 2030, and the companies with the most leverage over it.
Company names indicate supply-chain exposure to each bottleneck, illustrative, not investment advice. Trajectory synthesized from IEA, McKinsey, SemiAnalysis, TrendForce, Yole, SK Group and industry commentary (2024–2026).
How the chokepoints connect
The same constraints, drawn as a supply chain instead of a list. At the center sit the buyers, the frontier labs and the hyperscalers that host them. They buy accelerators; the accelerators are printed by a foundry and fed by stacked memory; the foundry runs on lithography, masks and materials; the racks need boards, optics and power; the sites need grid, water and people. Each bubble is one market. Its color is how binding that link is in the year you pick below, anywhere from 2024 to 2030; its size is how concentrated the market is, the share held by the single largest supplier. Tap any bubble to see who dominates it.
A schematic of the primary buy-and-supply links, not every relationship. Color follows the selected year's column of the trajectories above; concentration is the estimated share of the largest supplier in each chokepoint, an order of magnitude rather than a precise split. Dashed links (capital, export controls) gate the whole flow rather than feeding one node.
What the bottlenecks paid
A thought experiment to close on. Take the flagship company behind each of the hardest chokepoints above, one name per bottleneck, put an equal amount of money into each a year ago and hold. Here is how that basket of picks-and-shovels would have done against the S&P 500 over the last twelve months.
Equal-weight basket · 1yr
+63%
7 bottleneck bellwethers, held 12 mo
S&P 500 · same window
+12.3%
the basket beat it by ~5×
Leader · Micron
+301%
the HBM / DRAM squeeze, cooling from +628%
Laggard · Broadcom
−9%
the one name that lost money
Hypothetical and illustrative, not investment advice. Equal-weight means an equal dollar amount in each of the seven names a year ago, held without rebalancing; figures are price return only (excluding dividends) from the Oct 24, 2025 close to the Oct 23, 2026 close, the same window for every line. The average is still pulled upward by Micron's memory-shortage run: the median name returned about +39%, well ahead of the S&P 500's +12%. The basket has come a long way down from the end of July, when it stood at +141% against +19.5% for the index: Micron has given back part of a +628% year, Vertiv has gone from +70% to roughly flat, and Broadcom, the custom-silicon and networking proxy, has turned negative. Nvidia has moved the other way, off the bottom of the table to a narrow beat of the index. Which is the point of the section above: the bottleneck keeps moving, and so does the money. Source: daily closing prices via Yahoo Finance.