Anthropic has quietly signed roughly $517 billion worth of compute contracts since October 2025 — a figure that dwarfs the $180 billion in server costs the company told investors to expect through 2029 just nine months earlier. The tally, first reported by The Information on September 6, 2026 and confirmed by multiple outlets in the days that followed, covers at least 14.8 gigawatts of committed capacity spread across at least nine separate cloud and chip deals with Amazon, Google, Microsoft, SpaceX, Riot Platforms, Akamai, and others.
None of this is Anthropic's valuation or its next funding round — that's a separate story tied to its $965 billion Series H price tag and a confidential IPO filing targeting an October 2026 Nasdaq debut. This is purely the bill for the chips, servers, and data-center capacity Anthropic has locked in to keep training and serving Claude. For the millions of people paying $20 a month for Claude Pro or building on the API, the obvious question is whether that bill eventually shows up on their invoice, and what it says about how fast Anthropic can keep shipping new models.
Anthropic has committed roughly $517 billion to compute since October 2025, per The Information — up from the $180 billion in server-leasing costs it guided investors to through 2029 as recently as December 2025. The commitments cover 14.8+ gigawatts across deals with AWS (~$100B), Google/Broadcom (~$200B), Microsoft/Nvidia ($30B), SpaceX (~$45B), Fluidstack ($50B), Nscale ($45B), Lambda ($35B), Riot Platforms ($9.1B), and Akamai ($1.8B) — roughly 8x Anthropic's current $65 billion annualized revenue. Despite the spend, Anthropic just made its cheaper Sonnet 5 API pricing ($2/$10 per million tokens) permanent instead of raising it as planned. For now, the compute bill isn't hitting Claude subscribers' wallets.
What Anthropic actually signed up for
The $517 billion figure isn't one contract — it's an aggregate of at least seven multi-year, take-or-pay capacity reservations signed in less than a year, none of which currently sit on Anthropic's balance sheet as liabilities. Sources: [The Information: Anthropic Has Locked in 14.8 GW of Compute Worth Up to $517B Since October](https://aiweekly.co/alerts/the-information-anthropic-has-locked-in-148-gw-of-compute-worth-up-to-517b)
- Amazon Web Services: more than $100 billion over roughly 10 years for up to 5 gigawatts of Trainium2/Trainium3/Trainium4 and Graviton capacity, announced April 20, 2026, alongside Amazon's own pledge to invest up to $25 billion more directly into Anthropic.
- Google Cloud / Broadcom: an agreement Anthropic's aggregated compute tally puts at roughly $200 billion over 5 years for about 5 gigawatts of TPU capacity — up from the 3.5-gigawatt figure disclosed when the deal was last expanded in April 2026, which itself grew out of an October 2025 agreement for up to 1 million TPUs.
- Microsoft Azure / Nvidia: $30 billion in Azure compute credits plus up to 1 gigawatt of Nvidia Grace Blackwell and Vera Rubin capacity, bundled with Microsoft investing up to $5 billion and Nvidia up to $10 billion directly into Anthropic.
- SpaceX: about $1.25 billion a month (roughly $45 billion total through May 2029) for dedicated access to more than 200,000 Nvidia GPUs across SpaceX's Colossus I and II data centers.
- Fluidstack ($50 billion), Nscale ($45 billion), Lambda ($35 billion), Riot Platforms ($9.1 billion for a 20-year Texas data-center lease), and Akamai ($1.8 billion) round out the remaining capacity, according to The Information's tally.
From $180 billion to $517 billion in under a year
As recently as December 2025, Anthropic told investors it expected to spend about $180 billion renting servers through 2029. Nine months later, the figure reported for contracts signed since October 2025 alone is $517 billion — nearly triple that guidance, with almost the entire increase coming from deals struck in a single 11-month stretch. Sources: [Anthropic's $517B Compute Ceiling Reached in 11 Months](https://forkast.news/anthropics-517b-compute-ceiling-reached-in-11-months-even-as-its-ceo-called-to-slow-down/)
Scaled against revenue, the number is stark. Anthropic's annualized revenue run rate was about $9 billion at the end of 2025, climbed to $30 billion by April 2026, roughly $47 billion around its May Series H announcement, and $65 billion by the end of July 2026, with investors reportedly told to expect $100–120 billion by year-end. Even against that trajectory, $517 billion in compute commitments works out to roughly eight times current annualized revenue — though the spending is contracted across close to a decade rather than due all at once. Sources: [Anthropic tells investors annualized revenue run rate climbed to $65 billion in July — CNBC](https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html)
| Company | Reported Commitment | Time Frame | What It Covers |
|---|---|---|---|
| Anthropic | $517B | 11 months of deals (Oct 2025–Aug 2026), paid out through the mid-2030s | 14.8+ GW across AWS, Google/Broadcom, Microsoft/Nvidia, SpaceX, Fluidstack, Nscale, Lambda |
| OpenAI | ~$600B to ~$1.15T (estimates vary by source and date) | 2025–2030 or 2025–2035, depending on the tally | Broadcom, Oracle, Microsoft, Nvidia, AMD, AWS, CoreWeave, plus the Stargate joint venture |
| Google / Alphabet | $205B capex guidance | Single fiscal year (2026) | TPU v7 production, data centers, and networking — an annual capex figure, not a multi-year contract tally |
These figures aren't directly comparable. Anthropic's and OpenAI's numbers are sums of multi-year contracted capacity reservations stretching out to 2035, while Alphabet's $205 billion is a single year of capital expenditure covering far more than just Gemini. OpenAI's own figures have moved around, too: Sam Altman cited roughly $1.4 trillion in commitments in earlier disclosures, the company reportedly told investors in February 2026 it was budgeting closer to $600 billion in compute spend through 2030, and outside tallies of its named contracts with Broadcom, Oracle, Microsoft, Nvidia, AMD, AWS, and CoreWeave add up to about $1.15 trillion through 2035. On any of those readings, OpenAI's compute pipeline is still larger than Anthropic's, but Anthropic's has grown roughly three-fold in under a year, and closing that gap looks like the point. Sources: [OpenAI's $1 Trillion Infrastructure Spend](https://tomtunguz.com/openai-hardware-spending-2025-2035/), [OpenAI resets spending expectations, tells investors compute target is around $600 billion by 2030 — CNBC](https://www.cnbc.com/2026/02/20/openai-resets-spend-expectations-targets-around-600-billion-by-2030.html), [Google Raises 2026 CapEx to $205B — Converge Digest](https://convergedigest.com/google-raises-2026-capex-to-205b-as-ai-infrastructure-demand-outpaces-capacity/)
Will Claude subscriptions or API prices go up?
The short answer, at least so far, is no — if anything, Anthropic has moved the other way. Claude Sonnet 5 launched with "introductory" API pricing of $2 per million input tokens and $10 per million output tokens, with a step-up to $3/$15 scheduled for September 1, 2026. On August 10, 2026, weeks before the $517 billion figure became public, Anthropic cancelled that scheduled increase and made $2/$10 the standard, permanent rate — a move widely read as a response to competitive pricing from rivals like OpenAI's GPT-5.6 rather than a sign of falling compute costs. Sources: [Claude Sonnet 5 Price: Anthropic Cancels 50% API Hike, Keeps $2/$10 Rates — IBTimes](https://www.ibtimes.sg/claude-sonnet-5-price-anthropic-cancels-50-api-hike-keeps-2-10-rates-93232)
That pricing discipline is happening even as Anthropic's own numbers show compute still eating a large share of revenue. Gross margins have reportedly moved from roughly -94% in 2024 to an estimated 44–60% in 2026 depending on how training-cost amortization is counted, and the company has reportedly told investors it expects sustained free cash flow only in 2027–2028, once burn falls from about a third of revenue this year toward roughly 9% next year. In other words, Anthropic is funding this compute buildout through investor capital and its pending IPO, not by charging Claude users more — at least for now. A confidential S-1 targeting a raise of more than $60 billion, layered on top of the $965 billion valuation from its May 2026 Series H round, is the mechanism actually paying for these contracts. Sources: [Anthropic's gross margin is the most important number in tech — Yahoo Finance](https://finance.yahoo.com/markets/stocks/articles/anthropics-gross-margin-most-important-070500338.html), [Anthropic confidentially files for IPO after raising $65 billion in a funding round at a $965 billion valuation — Fortune](https://fortune.com/2026/06/01/anthropic-confidentially-files-ipo-965-billion-valuation/)
If you're budgeting API costs, Sonnet 5's $2/$10 per-million-token rate is now Anthropic's stated standard price, not a limited-time offer — it's reasonable to model it as your baseline rather than padding for a near-term hike.
| Plan | Price | Best For | Notes |
|---|---|---|---|
| Free | $0/month | Casual, occasional use | Limited daily messages, standard models only |
| Pro | $20/month ($17/mo billed annually) | Everyday individual use | Higher usage limits, includes Claude Code access |
| Max 5x | $100/month | Heavy daily users, developers | About 5x Pro's usage allowance |
| Max 20x | $200/month | Teams running long agentic sessions | About 20x Pro's usage allowance |
| API — Haiku 4.5 | $1 / $5 per million tokens (in/out) | High-volume, low-latency tasks | Cheapest current Claude model |
| API — Sonnet 5 | $2 / $10 per million tokens (in/out) | General-purpose default | Introductory price made permanent Aug 10, 2026 |
| API — Opus 5 | $5 / $25 per million tokens (in/out) | Frontier reasoning, complex agents | Anthropic's top-tier model |
The irony of "pace the frontier" week
The compute story broke during an unusually pointed week for Anthropic's public messaging. On September 12, 2026, CEO Dario Amodei published a roughly 3,800-word essay titled "We Must Pace the Frontier," arguing the industry should deliberately slow the rate at which AI capabilities improve to give safety evaluation and oversight time to keep up — not halt progress, but pace it. Sam Altman publicly agreed within hours, and Anthropic reportedly committed unilaterally to the essay's first proposed step: embedding independent evaluators within its labs. Sources: [Anthropic CEO Dario Amodei Calls For A Slowdown In Frontier AI — Forbes](https://www.forbes.com/sites/gabrielalinzainescu/2026/09/13/anthropic-ceo-dario-amodei-calls-for-a-slowdown-in-frontier-ai/), [Anthropic, OpenAI CEOs call for slowdown in AI development — Axios](https://www.axios.com/2026/09/12/anthropic-ai-amodei-pacing)
That call for restraint landed in the same week reporting surfaced that Anthropic had spent the prior 11 months locking in one of the largest industrial buildouts in corporate history. The two stories aren't necessarily contradictory — pacing capability releases and building compute capacity years in advance are different levers — but the juxtaposition was hard to miss, and it's a fair question for Claude users to ask whether "pace the frontier" shows up mainly as a slower model release cadence or mainly as more careful evaluation before models ship.
What this actually means for Claude users
- Fewer capacity crunches, probably: 14.8 GW of new capacity coming online mostly from 2026 through 2028 should ease the rate limits and "at capacity" errors that have periodically hit Claude Pro, Max, and API users during demand spikes.
- No sign of a near-term price hike: Anthropic just cancelled one scheduled increase; the more likely lever for extracting more revenue is deeper Max and Enterprise tiers, not a rewrite of Pro's $20 price point.
- Faster model cadence is plausible, not guaranteed: a large share of the new capacity, including part of the Google/Broadcom TPU expansion, doesn't come online until 2027, so any training or serving benefit shows up gradually rather than overnight — and Amodei's pacing commitments could deliberately offset some of the capability gains anyway.
- The IPO is the real funding mechanism: most of the $517 billion is take-or-pay capacity Anthropic must eventually pay for regardless of demand, and its confidentially filed S-1 (targeting more than $60B raised, an October 2026 Nasdaq listing, and a valuation bankers have reportedly floated as high as $2 trillion) is what's meant to cover it, not Claude subscription revenue.
01Is the $517 billion the same thing as Anthropic's valuation?
No. Anthropic's valuation comes from its funding rounds — a $965 billion price tag from its May 2026 Series H round, with bankers reportedly discussing figures as high as $2 trillion for its planned October 2026 IPO. The $517 billion is a separate number: the sum of multi-year compute contracts Anthropic has signed with cloud and chip providers since October 2025.
02Who reported the $517 billion figure?
The Information first reported the tally on September 6, 2026, aggregating publicly known and previously unreported deals into a total of at least 14.8 gigawatts of committed capacity. It was subsequently covered by outlets including Data Center Dynamics, Yahoo Finance, and Forkast.
03Will Claude Pro or API prices go up because of this?
There's no evidence of that so far. If anything, Anthropic moved the other way on August 10, 2026, cancelling a scheduled price increase for Claude Sonnet 5's API and making the cheaper $2/$10 per-million-token rate permanent. The compute buildout is being funded through investor capital and the upcoming IPO rather than subscriber price increases, though that could change if IPO proceeds or margins disappoint.
04How does $517 billion compare to what OpenAI and Google are spending?
Estimates of OpenAI's total compute commitments range from about $600 billion (a February 2026 investor reset reported by CNBC) to roughly $1.15 trillion tallied across its named contracts through 2035, with Sam Altman having cited figures as high as $1.4 trillion in earlier disclosures. Google/Alphabet's number isn't a multi-year commitment at all — it's a single-year 2026 capital expenditure guidance of $205 billion. On any reading, OpenAI's compute pipeline remains larger than Anthropic's, but Anthropic's has grown roughly three-fold in under a year.
05Does more compute mean Claude will get noticeably better or faster to use?
It should help with capacity — fewer rate limits and outages during peak demand — since most of the new capacity comes online through 2026–2028. Whether it translates into faster model releases is less certain, especially after Amodei's September 12 essay calling for the industry to deliberately pace capability improvements.
Whatever Anthropic's compute bill means for its balance sheet, the practical test for most people is simpler: does Claude stay worth what you're paying for it, relative to GPT-5.6, Gemini, and everything else on the market. LumiChats tracks pricing, model updates, and head-to-head comparisons across these chatbots in one place, so you can check current plans and benchmarks side by side before deciding where your subscription dollars go.
