News · AI Market
AI compute is paid in cash: Google Cloud doubles, Alphabet and Amazon burn through free cash flow
On October 28 and 29, the three big clouds reported growth of 39% to 103% and capital spending that has roughly doubled in a year. Alphabet and Amazon show negative free cash flow, Anthropic put its commitments at $895 billion on November 2, OpenAI is reportedly cutting its round from $30 billion to $20 billion, and base model prices have not moved since October 28: the token stays stable, while the cost of the capacity that produces it and the capital that funds it are what change.
Meotis Research·2 November 2026·10 min read

The three big clouds published their third-quarter results on October 28 and 29, 2026, and demand for compute is not the issue: Google Cloud revenue grew 103% to $30.7 billion, AWS 39% to $46.0 billion, and Azure 48%. What caught the markets’ attention was the bill. Alphabet closed the quarter with negative free cash flow of $9.1 billion, Amazon invested $77.3 billion in three months, and Alphabet’s shares lost about 10% in extended trading (CNBC, October 29). On October 30, the Financial Times reported that OpenAI was cutting its round from $30 billion to $20 billion. On November 2, Anthropic’s amended prospectus put its commitments, almost entirely for compute, at $895.4 billion. Meanwhile, base model prices have not moved since our October 28 survey. The price of a token stays stable; what changes is the cost of the capacity that produces it and the patience of the capital that funds it.
Three clouds, three growth rates, high margins
The figures below come from the filings with the SEC, which we read.
Alphabet (quarter ended September 30, published October 28): revenue of $127.6 billion, up 25%, of which $30.7 billion for Google Cloud, up 103%, with operating income of $11.1 billion for that division alone. The revenue backlog (remaining performance obligations) stands at $567.7 billion, of which $560.8 billion relates to Google Cloud; Alphabet expects just over 55% of it to be recognised as revenue within 24 months (10-Q quarterly report). A portion of $16.8 billion comes from an agreement with Crux AI, a joint venture that is a related party: that amount is therefore less independent than the rest. We did not find the backlog broken down by customer.
Microsoft (first quarter of its fiscal 2027, ended September 30, published October 28): revenue of $92.3 billion, up 19%, net income of $36.6 billion, Azure revenue of $33.2 billion, up 48% ($22.4 billion a year earlier).
Amazon (published October 29): AWS posted $46.0 billion, up 39% ($33.0 billion a year earlier), with operating income of $17.8 billion against $11.4 billion.
A calculation from these releases, which is ours: Google Cloud’s operating margin is about 36% ($11.1 billion on $30.7 billion) and AWS’s about 39% ($17.8 billion on $46.0 billion). These margins cover each division as a whole, not compute alone. For a buyer, they indicate that the price of capacity does not depend only on token prices at the model makers.
Investing twice as much, with cash flow turning red
At all three groups, the investment line has roughly doubled in a year.
Amazon bought $77.3 billion of property and equipment in the third quarter, against $35.1 billion a year earlier, before proceeds from disposals. Its free cash flow over the last twelve months is negative $42.4 billion, against an inflow of $14.8 billion a year earlier. The group maintains a forecast of about $220 billion of investment in 2026 (release of October 29).
Alphabet invested $51.4 billion in the quarter, $132 billion over nine months against $63.6 billion a year earlier (10-Q). With $42.3 billion of operating cash flow, free cash flow for the quarter is negative $9.1 billion. CNBC reports that the group maintains its range of $195 to $205 billion for 2026, short of analysts’ expectations, and that the stock fell about 10% after the close.
Microsoft recorded $39.4 billion of additions to property and equipment (cash flow), against $19.4 billion a year earlier. CNBC reports an annual forecast lowered from $175 billion to $170 billion, with the finance team citing delays at third-party GPU suppliers that push some capacity into the first half of 2027; we did not find that forecast in the release itself, nor the basis (calendar year or fiscal year) to which it refers.
CNBC also writes that hyperscaler spending is expected to exceed $1’000 billion next year, without detailing the estimate in the article; that CoreWeave, Nebius, IREN and Oracle fell after these results; and that Gartner predicts 80% of specialised compute companies (“neoclouds”) will be absorbed or disappear by 2030. For a buyer, this last item concerns the lifespan of the supplier more than its price.
Meta, already cited on October 28, raised the floor of its 2026 investment range, now $135 to $145 billion (CNBC, October 29).
Anthropic: the amended prospectus puts its commitments at $895 billion
On October 26, Anthropic announced the public filing of its Form S-1 with the SEC for a Nasdaq listing under the symbol “ANT”. Goldman Sachs and Morgan Stanley are named as contacts for copies of the preliminary prospectus (Anthropic announcement, October 26). On November 2, an amended version (S-1/A) added the third-quarter accounts and filled in the commitments total, left blank in the first filing. The number of shares and the price range are still blank, and the registration is not yet effective.
The figures that follow come from that document, which we read, and rest on unaudited interim accounts.
- Revenue: $33.2 billion over nine months ($2.5 billion a year earlier), of which $17.0 billion in the third quarter, up 1’127% in a year. About 79% comes from consumption-based fees.
- Cost of revenue: $4.8 billion over nine months, about 14% of revenue (that ratio is our calculation). The prospectus describes it as substantially all compute and infrastructure serving customers, plus some personnel and overhead costs.
- Compute and infrastructure, all uses: $18.8 billion over nine months, 57% of revenue, of which $9.9 billion in the third quarter, up 382% in a year. About three quarters of that compute is therefore booked outside cost of revenue (our calculation), mainly in research and development, a $19.2 billion line that carries training, research and internal usage.
- Results: an operating loss of $1.26 billion over nine months. After operating income of $974 million in the second quarter, the third shows an operating loss of $567 million. The prospectus attributes it mainly to $5.0 billion of compute bought under short-term commitments, which currently costs significantly more than comparable compute under multi-year commitments, and to $2.0 billion of non-cash charges linked to its donation matching program. The net loss of $104.0 billion includes $94.3 billion of fair value changes (convertible notes, derivatives, SAFEs).
- Customers: none exceeds 10% of revenue over nine months, and the top five account for about 25%. The prospectus adds that many top customers have no long-term contract or one close to expiry, and that many contracts allow usage to be cut or terminated with limited or no notice.
- Commitments: $895.4 billion as of and after September 30, including leases not yet commenced and short-term commitments, almost entirely for compute and infrastructure. $447.5 billion is expected to be spent before the end of 2030 and almost all the rest from 2031 to 2045; $69.8 billion falls due within twelve months. Of this total, $213.8 billion depends on termination rights or on conditions to be met before commencement (financing, facilities, approvals). This figure replaces the press totals ($430 billion to $750 billion), worked out while the prospectus left the amount blank, and the $518 billion we used on October 28.
For an API buyer, two lessons. Serving customers costs a fraction of what they pay, but compute as a whole absorbs 57% of revenue: the token price also pays for training the next model, which leaves room for price cuts without guaranteeing them. And the supplier commits over many years while its revenue, it writes, is primarily consumption-based and largely without long-term commitments: the mirror image of the risk a buyer runs by depending on a single supplier.
OpenAI: $20 billion instead of $30 billion, according to the press
On October 22, CNBC described a round of up to $30 billion at a valuation of $1’250 billion before the new money, led by investors from the United Arab Emirates and Coatue, as bridge financing ahead of a stock market listing. OpenAI reportedly filed its prospectus confidentially in June, and its chief financial officer has said it will be listed in 2027.
On October 30, the Financial Times reported that the target had been lowered to $20 billion, and that the first close would be below that amount. A later offer to wealthy individuals could complete the total. The Wall Street Journal adds that OpenAI has secured $11 billion so far and that some large investors reacted with lukewarmness (TechCrunch, October 31, relaying both newspapers). TechCrunch links this difficulty to bad market timing, between investor wariness over AI infrastructure spending and the decline in technology stocks. OpenAI has confirmed none of these figures, which remain press reports.
Reading for a buyer: the floor prices we described on October 12 and 28 are market positions financed by capital. When capital hesitates, one can expect fewer price cuts, or more conditions around prices. That is not a forecast: no base price has changed since October 28.
List prices have not moved; the lines around them have
On November 2 we re-read the official pages of OpenAI (developers.openai.com), Anthropic (platform.claude.com), Google (aistudio.google.com), xAI (docs.x.ai) and DeepSeek (api-docs.deepseek.com). The base prices are identical to our October 28 survey: GPT-6.1 Terra at $1 / $5, Claude Opus 5.5 at $4 / $20, Gemini 4 Argon at $2 / $10 until November 30 then $4 / $20, Grok 4.7 at $2 / $6. What has changed or will change, by date:
- November 1, DeepSeek: web search and page fetching are billed at $0.75 per 1’000 successful requests (Beijing time). The page also mentions that server-side calls made through certain agent environments are free “for now”; we could not establish whether that holds after November 1.
- November 1, OpenAI: a 10% uplift applies to FedRAMP endpoints (official page). The 10% uplift for regional processing of models published since March 5, 2026 remains on the same page.
- November 21, OpenAI: the page indicates promotional pricing of $4 / $20 for GPT-5.6 Sol through November 21 at least, without giving the reference price; we could not establish the gap.
- December 1, Google: requests to Gemini 2.5 Pro and Gemini 2.5 Flash, deprecated models, will be redirected to Gemini 3.8 Flash and Gemini 3.5 Flash-Lite and billed at their rates. For a customer still on 2.5 Flash ($0.30 / $2.50), the Flash-Lite 3.5 rate is identical; for 2.5 Pro ($1.25 / $10), it falls to that of 3.8 Flash ($0.75 / $3.75, which doubles on January 1, 2027).
Subscriptions. Claude (claude.com/pricing, November 2): Pro at $20 a month ($17 on annual billing), Max at $100, Ultra at $200, Team Standard seats at $25 a month ($20 annual) and Team Premium at $125 ($100 annual), Enterprise at $20 per seat with a twenty-seat minimum, plus usage billed at API rates. Mistral (mistral.ai/pricing, November 2): Le Chat Pro at $14.99 a month, Team at $24.99 per user. We could not consult ChatGPT’s subscription page: we cannot confirm OpenAI’s rates, and we did not consult Google’s.
GPU rental. The getdeploying survey, dated November 2 and updated continuously, showed that evening on-demand medians of $3.99 an hour for an H100 (46 providers list a price), $4.88 for an H200 (37) and $8.09 for a B200 (20), with minimum offers of $1.80, $2.09 (availability not verified by the page) and $3.75. The October 26 survey we cited counted 39 providers for the H100 (median $3.70): the two series do not have the same scope, and we conclude neither a rise nor a fall. The page itself points out that a provider’s own site may quote a different figure (tax, region, promotion).
Kimi K3.1 Pro (Moonshot): its open weights were promised for November 4; according to news digests of November 2 (not cross-checked), they were not yet available. Its official rate card was empty at our consultation: we do not price it.
API price comparison table
| Model | Input ($/M tokens) | Output ($/M tokens) | Note |
|---|---|---|---|
| GPT-6.1 Terra | 1.00 | 5.00 | Cache 0.05; beyond 272k tokens: 2.00 / 7.50 |
| GPT-6.1 Sol | 2.00 | 10.00 | Cache 0.10; beyond 272k: 4.00 / 15.00 |
| GPT-6 Astra | 10.00 | 50.00 | Cache 1.00; beyond 272k: 20.00 / 75.00 |
| GPT-6 Luna | 0.10 | 0.50 | Cache 0.01; beyond 272k: 0.20 / 0.75 |
| GPT-5.6 Terra | 2.00 | 12.00 | Still listed |
| Claude Fable 5.1 | 10.00 | 50.00 | Cache read 0.25 |
| Claude Opus 5.5 | 4.00 | 20.00 | Unchanged; cache read 0.20 |
| Claude Sonnet 5.5 | 2.00 | 10.00 | Unchanged; cache read 0.10 |
| Claude Haiku 5.5 | 0.10 | 0.50 | 0.50 / 2.50 beyond 100k tokens |
| Gemini 4 Argon | 2.00 | 10.00 | 4.00 / 20.00 from December 1 |
| Gemini 3.8 Flash | 0.75 | 3.75 | 1.50 / 7.50 from January 1, 2027 |
| Gemini 3.5 Flash-Lite | 0.30 | 2.50 | No dated change listed |
| Gemini 3.1 Pro (preview) | 2.00 | 12.00 | Up to 200k tokens; 4.00 / 18.00 beyond |
| Grok 4.7 | 2.00 | 6.00 | 4.00 / 12.00 from 200k tokens (whole request) |
| DeepSeek V4-Pro | 0.66 to 1.32 | 1.98 to 3.96 | Off-peak / peak |
| DeepSeek V4.1 Flash | 0.15 to 0.30 | 0.60 to 1.20 | Off-peak / peak; web search and fetch $0.75 per 1’000 requests from November 1 |
| Mistral Large | 0.50 | 1.50 | The page does not specify the model version |
Prices verified on November 2, 2026 on the official pages of OpenAI (developers.openai.com), Anthropic (platform.claude.com), Google (aistudio.google.com), xAI (docs.x.ai), DeepSeek (api-docs.deepseek.com) and Mistral (mistral.ai/pricing). Kimi K3.1 Pro is absent: its official rate card was empty at our consultation. Reasoning models bill tokens the user never reads, and the same text does not count as the same number of tokens from one supplier to another: the lines of this table cannot be compared without testing on your own workload.
What a Swiss decision-maker should take away
Four findings, with no change in base prices since October 28.
Demand for compute is not weakening, and the seller of compute is highly profitable. Google Cloud and AWS earn operating margins of about 36% and 39% (our calculations from the releases). The token price is set at the model makers; the price of the underlying capacity is set by sellers whose margins do not force them to cut it. Anthropic itself writes that compute bought short term costs it significantly more than under multi-year commitments. In an AI budget, the two lines do not move together.
Serving a token appears to cost only a fraction of its price, which guarantees no price cut. At Anthropic, cost of revenue is about 14% of revenue over nine months, but compute as a whole 57% (prospectus of November 2; the first ratio is our calculation). That says there is room on serving. It does not say it will be handed back: that room also pays for the next model, and the capital that financed floor prices is tightening, with OpenAI’s round going from $30 billion to $20 billion in eight days according to the press.
Supplier risk runs both ways. A supplier with many contracts that can be cut or terminated on little or no notice, and a specialised compute sector of which Gartner sees 80% absorbed or gone by 2030, are reminders that an AI contract reads like an infrastructure contract: term, exit conditions, location of processing, and a plan B.
The deadlines are written down. November 21 (GPT-5.6 Sol promotion), December 1 (Gemini 4 Argon at $4 / $20, redirection of the 2.5 models), January 1, 2027 (Google’s Flash models). Budgeting 2027 on the rates of November 2 means budgeting on rates several of which expire.
These findings call for the answer of previous weeks: a layer that routes each task to the model with the right cost/capability ratio, measures what it really costs, and allows you to change supplier the day a rate card, a contract or a funding round moves. When demand for compute doubles at one supplier and capital hesitates at another, the advantage is not in the listed price. It lies with whoever knows what their own workload costs.
Control your AI costs instead of enduring them. Comoto Enterprise OS orchestrates the market’s best models from a sovereign core hosted in Switzerland: every task routed to the model with the right cost/capability ratio, the location constraint applied where it is required, and the ability to switch models that stays with you.
Discover Comoto Enterprise OSBe the first to read our upcoming papers.
Some articles are reserved. Leave your email to unlock access to upcoming Meotis Research publications.
Your turn to try.
- 30 minutes, by video or in Neuchâtel
- A demo built around your own case
- With the founding team, no strings attached


Meotis
Discovery call
Choose a day
Choose a time
Neuchâtel time
Pick a day, then a time.

