How much compute has Anthropic committed to buy?

Up to $517 billion across cloud, chip and data-center contracts signed in the eleven months from October 2025 through August 2026, covering at least 14.8 gigawatts of new capacity on top of the 1-2 GW it already had, according to The Information (September 7, 2026). That is nearly three times the roughly $180 billion in server-leasing spend through 2029 that Anthropic had previously told investors to expect. The $517 billion is an upper bound on potential spending, not cash out the door, and many contracts run well past 2030.

Who is supplying Anthropic's 14.8 GW?

Amazon and Google are the bulk: about 11 GW in agreements valued above $300 billion. Microsoft Azure has a $30 billion contract for roughly 1 GW. SpaceX deals are reported near $45 billion. Lambda and nScale account for about $80 billion for 460 megawatts (Lambda ~$35 billion, nScale ~$45 billion). Fluidstack is also a partner, and Anthropic is planning data centers of its own.

How does that compare with OpenAI?

OpenAI's public target is about 30 GW of capacity by 2030, roughly double Anthropic's 14.8 GW plus existing 1-2 GW, though The Information notes many of Anthropic's contracts extend past 2030 so the comparison is imprecise. On revenue, Anthropic's annualized run rate topped $65 billion (Bloomberg) versus OpenAI above $40 billion as of July 2026, so Anthropic is buying compute at a lower multiple of revenue. Neither company can fund its commitments from revenue alone yet.

Why is Anthropic's spending controversial?

Because of timing. Dario Amodei published 'We must pace the frontier' in September 2026, calling for labs to slow frontier development, days before the $517 billion figure surfaced; China's foreign ministry called the essay fearmongering, and critics from Axios to Reddit argue that a coordinated slowdown would conveniently conserve cash for labs burning enormous sums on compute. Earlier in 2026 Amodei had warned competitors 'don't really understand the risks they're taking' by investing too fast, while Sam Altman has warned of 'unsustainable silliness' in neocloud buildouts. Both labs are now racing on infrastructure while publicly urging caution.

Quick Answer

Anthropic's $517B Compute vs OpenAI's 30 GW Plan (Sep 2026)

Published: September 16, 2026

The short answer

Anthropic has committed up to $517 billion for at least 14.8 GW of compute in eleven months; OpenAI is targeting about 30 GW by 2030. Anthropic is buying faster relative to its size and revenue, OpenAI is targeting a larger absolute footprint over a longer horizon, and both are doing it while their CEOs publicly argue for restraint. Figures below are from The Information (September 7, 2026), Bloomberg and company statements, verified September 16, 2026.

	Anthropic	OpenAI
Headline commitment	Up to $517B (Oct 2025-Aug 2026 contracts)	~30 GW by 2030 (public target); Stargate program
Capacity	14.8 GW new + 1-2 GW existing	~30 GW target
Prior guidance	~$180B server leasing through 2029 (told to investors earlier)	Multi-hundred-billion Stargate commitments announced 2025-2026
Main suppliers	Amazon + Google (~11 GW, >$300B); Microsoft Azure ($30B, ~~1 GW); SpaceX (~~$45B); Lambda + nScale ($80B, 460 MW); Fluidstack; own data centers planned	Microsoft, Oracle, SoftBank/Stargate, CoreWeave, AWS, plus own sites and the Jalapeño custom chip
Annualized revenue	>$65B (Bloomberg)	>$40B as of July 2026
Compute-commitment to revenue	~8x	Not directly comparable; commitments spread through 2030+
Public posture	Amodei: “We must pace the frontier” (Sep 2026)	Altman: no 2026 IPO, safety concerns; warns of “unsustainable silliness” in neocloud spending
Financing status	IPO preparation; $2T valuation talk	IPO ruled out for 2026

What Anthropic signed

The Information’s tally covers contracts struck between October 2025 and August 2026:

Amazon and Google, about 11 GW, more than $300 billion. Anthropic’s two largest investors are also its two largest suppliers, across AWS Trainium capacity and Google TPUs plus GPU cloud.
Microsoft Azure, $30 billion for about 1 GW, the deal that made Claude available on Azure and Foundry.
SpaceX, about $45 billion, the most surprising line; details of what SpaceX provides (data-center power, orbital compute, or capacity via xAI-adjacent infrastructure) have not been confirmed by either company.
Lambda and nScale, about $80 billion for 460 MW, the neocloud tier. Our breakdown: Anthropic’s $35B Lambda vs $45B nScale deals.
Fluidstack and Anthropic’s own planned data centers round it out.

The energy footprint is comparable to more than ten large nuclear reactors, and The Information stresses that $517 billion is a ceiling on potential spend across the life of the contracts, not a committed cash figure.

What OpenAI is doing

OpenAI’s target is roughly 30 GW by 2030 through the Stargate program with SoftBank and Oracle, continuing Microsoft capacity, CoreWeave, AWS, and its own custom silicon; the Jalapeño accelerator benchmarked against GB300 in August 2026. Its published dollar figures have been announced deal by deal rather than as one total, which is why headlines quote gigawatts for OpenAI and dollars for Anthropic. On September 12, 2026, Sam Altman told Fortune that OpenAI will not go public in 2026, citing safety concerns and calling the timing ill-advised.

Why the comparison is imprecise

Horizons differ. Many Anthropic contracts run past 2030; OpenAI’s 30 GW is a 2030 snapshot.
Dollars vs watts. Price per gigawatt varies 2-3x depending on chip generation and whether the deal includes power and buildings.
Upper bounds. Both companies’ figures are maximums under take-or-pay style agreements with ramp schedules.
Vendor-financed. Amazon and Google are investors in Anthropic; Microsoft and SoftBank in OpenAI. Some “commitments” are partly circular.

The credibility problem both labs now have

The timing of the $517 billion report, days after Amodei’s “We must pace the frontier” essay, gave critics an opening:

Axios (September 15, 2026): pacing could reduce Anthropic’s compute spend, and no frontier lab can sit out the capital markets; a cynical reading is that a coordinated slowdown protects incumbents from cheaper or open rivals.
The Decoder: Amodei said in early 2026 that competitors “don’t really understand the risks they’re taking” by investing too fast; now Anthropic is the one racing.
Beijing: China’s Foreign Ministry called Amodei’s chip-restriction argument “fearmongering.”
Washington: President Trump rejected the slowdown call on September 14, saying his administration has “tremendous criminal and regulatory power” over AI firms and framing data-center opposition as helping China.

Meanwhile OpenAI confirmed on September 15 that it has been in weeks of safety talks with Anthropic and Google DeepMind, even as it builds toward 30 GW. Read together: the labs want a coordinated pace, not a smaller footprint. See Who is slowing down AI: OpenAI vs Anthropic vs Google.

What it means for customers

Capacity constraints should ease in 2027, when the first tranches of these deals land; the September 2026 ChatGPT Pro $200 sign-up pause and Claude Code’s 17% weekly-limit cut are symptoms of the current squeeze.
Prices may not fall as fast as hardware improves. Both labs must service enormous fixed commitments; per-token cuts in 2026 came from efficiency and competition, not from spare capacity.
Supplier concentration is a risk on both sides. Anthropic’s ~11 GW from two investor-suppliers and OpenAI’s Stargate dependence on SoftBank and Oracle financing mean a single partner’s balance sheet can gate model availability.