What is the xAI-Anthropic compute deal?

Disclosed in SpaceX's May 20, 2026 S-1 filing: Anthropic agreed to pay xAI $1.25 billion per month through May 2029 for compute capacity on Colossus 1 and Colossus 2 — generating over $40 billion in projected revenue for xAI. The deal began ramping in May 2026 and June 2026. xAI migrated its own training workloads to Colossus 2 to free up Colossus 1 for the Anthropic contract. It's the largest single AI compute contract ever publicly disclosed.

Why is Anthropic buying compute from a competitor?

Three reasons. (1) Supply: Anthropic's training and serving demand outpaced what AWS Trainium and Azure Maia could deliver in 2026. (2) Speed: Colossus has ~220,000 NVIDIA GPUs (H100/H200/GB200) already racked and powered in Memphis — faster than building new capacity. (3) Diversification: Anthropic was over-reliant on AWS; the xAI deal hedges against any single-cloud supply shock. Counterintuitively, xAI selling compute to Anthropic is closer to a pure infrastructure play — xAI's own training has moved to newer Colossus 2 hardware, so the Colossus 1 capacity is monetizable surplus.

How big is OpenAI's Microsoft compute deal compared to xAI-Anthropic?

Different structure. OpenAI-Microsoft is an exclusive (with carve-outs) infrastructure partnership where Microsoft commits multi-tens-of-billions of capex over 5+ years and OpenAI commits to Azure as primary cloud. The xAI-Anthropic deal is a finite, time-boxed compute purchase ($1.25B/month, 36 months, ~$45B total). OpenAI-Microsoft is bigger over a longer horizon but is opaque on monthly cash flow. xAI-Anthropic is the largest cleanly-disclosed monthly compute contract.

Where does Google fit?

Google is the outlier — they built their own TPU silicon stack, so they don't buy compute from third parties at this scale. Google trains and serves Gemini on TPU v6/v7 racks in Google Cloud data centers. They sell TPU access to third parties via Google Cloud (Anthropic was historically a TPU customer before scaling to AWS Trainium + xAI Colossus). Google's strategy: full vertical integration. xAI/Anthropic/Microsoft/OpenAI all share NVIDIA GPU dependency; Google is the only frontier lab with a meaningfully different supply chain.

Quick Answer

xAI-Anthropic Compute vs OpenAI-Microsoft vs Google TPU (May 2026)

Published: May 27, 2026

xAI-Anthropic Compute vs OpenAI-Microsoft vs Google TPU (May 2026)

SpaceX’s May 20, 2026 S-1 filing revealed a stunning detail: Anthropic is paying xAI $1.25 billion per month for compute access to Colossus. It’s the largest cleanly-disclosed AI compute contract ever — and it reshapes how to think about who’s buying compute from whom in 2026. Here’s how it compares to the OpenAI-Microsoft and Google-TPU strategies.

Last verified: May 27, 2026.

TL;DR table

	xAI ↔ Anthropic	OpenAI ↔ Microsoft	Google (TPU, in-house)
Structure	Compute purchase	Strategic partnership	Vertical integration
Disclosed monthly cost	$1.25B / mo	Not broken out	N/A (in-house)
Total contract	~$45B over 36 months	Multi-tens of billions (5+ yr horizon)	N/A
Hardware	NVIDIA H100/H200/GB200 (Colossus 1)	NVIDIA + AMD MI300X	Google TPU v6 / v7
Site	Memphis, TN	Multiple US + EU Azure regions	Google Cloud (global)
Power scale	~1.4 GW (Colossus 1+2 combined)	Distributed across Azure	Distributed across Google Cloud
Customer dependency	Anthropic moderately diversified	OpenAI mostly Azure-locked	Google: zero (self-supply)
Start date	May 2026	2019 (renewed 2023, 2025)	Long-standing
End date	May 2029	Open-ended	N/A
Strategic logic	Surplus monetization + diversification	Co-investment in AI buildout	Avoid NVIDIA tax

Why each deal exists

xAI ↔ Anthropic: surplus monetization meets emergency supply

Three forces converged.

1. xAI has surplus Colossus 1 capacity. They migrated their own Grok 5 training to Colossus 2 (the newer site, on GB200 silicon). That left Colossus 1 — ~220,000 NVIDIA H100/H200 GPUs in Memphis, fully racked, fully powered — without an internal workload. Either you find a buyer or you’ve stranded $20B+ of capex.

2. Anthropic needed compute fast. Their growth from $4.8B Q1 2026 revenue to a projected $10.9B Q2 outpaced their existing AWS Trainium + Azure Maia capacity. They didn’t have 12 months to wait for new data centers to come online.

3. Both sides wanted to de-couple from existing dependencies. Anthropic was uncomfortably AWS-heavy. xAI didn’t want to be a pure single-customer Musk-entity business. The deal solves both problems simultaneously.

The result: $1.25B/month from May 2026 through May 2029, ~$45B total. Embedded in the SpaceX S-1 prospectus to signal to public-market investors that xAI is a real cash-generating asset post-IPO.

OpenAI ↔ Microsoft: co-investment in AI buildout

Different structure entirely. Microsoft has put multi-tens of billions of capex into Azure data centers specifically for OpenAI workloads. In return, OpenAI uses Azure as primary cloud (with carve-outs for SoftBank’s Stargate and recently xAI Colossus). The relationship is exclusive in spirit, ambiguous in practice.

What’s different vs xAI-Anthropic:

Time horizon: 5+ years vs 36 months
Disclosure: opaque (no monthly cash flow published) vs cleanly broken out
Counterparty risk: OpenAI heavily depends on Microsoft’s continued investment; Anthropic can walk from xAI in 2029
Strategic alignment: Microsoft is also OpenAI’s largest investor (10%+); xAI and Anthropic have no equity entanglement

Google: the only vertically integrated frontier lab

Google’s TPU v6 (“Trillium”) and the upcoming TPU v7 power Gemini training and serving. Google doesn’t buy compute from third parties at scale — they sell it. Anthropic was historically a TPU customer before scaling to AWS Trainium and now xAI Colossus.

Why this matters: every other frontier lab has NVIDIA dependency in their cost structure. Google pays Broadcom for TPU silicon fab and has 5-10x lower per-FLOP serving costs at scale because they’ve cut out the NVIDIA margin. This is the structural reason Gemini 3.5 Flash can be priced at $1.50 / $9 per 1M tokens while Claude Opus 4.7 is $15 / $75.

The Anthropic strategic question

Why is Anthropic — a Claude shop, an Anthropic-API-protocol vendor, a direct competitor to Grok — buying compute from xAI?

Answer: at scale, compute is fungible. A GPU running an Anthropic training job is just a GPU. xAI doesn’t see Anthropic’s training data, model weights, or inference workloads — they sell raw capacity. From Anthropic’s perspective, the moral commitments they have around safety apply to their training and serving — what hardware vendor they buy from is a procurement decision, not a values decision.

Counterpoint: there are reputational risks (Anthropic publicly differentiates from xAI on safety norms). But $45B of urgently-needed compute trumps reputational softness when growth is going from $4.8B to $10.9B quarterly.

Power and infrastructure

Site	Power capacity	Lab
Colossus 1 (Memphis)	~700 MW	xAI → leased to Anthropic
Colossus 2 (Memphis)	~700 MW	xAI (own training)
Stargate (Texas)	1+ GW announced	OpenAI / Microsoft / SoftBank
Azure regions (multiple)	~ multi-GW combined	OpenAI
Google Cloud (multiple)	~ multi-GW combined	Google (self-supply)

The Memphis grid is now ~1.4 GW of AI compute on its own. This is the kind of infrastructure scale where US grid capacity becomes a national-security bottleneck — and it’s why the May 2026 data-center power crisis is the dominant infrastructure story alongside the IPO calendar.

What this means for AI economics

1. NVIDIA’s pricing power is now fully realized. Even Anthropic, the lab founded by ex-OpenAI safety researchers, is sending $1.25B/month to a Musk-controlled entity to get NVIDIA GPU time. NVIDIA captures margin upstream of that.

2. Google’s TPU bet is paying off massively. Self-supply at scale is the only escape from the NVIDIA margin tax. Gemini 3.5 Flash being 5-10x cheaper than Claude Opus 4.7 reflects this directly.

3. xAI is now an AI infrastructure business, not just a model lab. The SpaceX-xAI merger reframed it; the Anthropic deal validates it. If $40B+ of revenue is coming from compute resale over 3 years, xAI’s valuation is no longer dependent on Grok 5 winning the model race.

4. The “compute is moat” thesis is back. Through 2024-2025, the narrative was “models are commoditizing, the moat is product.” May 2026 says: compute is the moat, and the labs that can self-supply (Google) or lock in supply (Anthropic-via-xAI) win the next 24 months.

Verdict

Cleanest disclosed deal: xAI ↔ Anthropic. $1.25B/month, 36 months, embedded in a public S-1.
Biggest aggregate commitment: OpenAI ↔ Microsoft. Multi-tens of billions over 5+ years.
Most strategically sound: Google TPU. Vertical integration is the only escape from NVIDIA.
Best for AI lab counterparty risk: Google (no counterparty). Worst: OpenAI (heavy Microsoft dependency).
What to watch next: whether Anthropic’s $1.25B/month is a one-off or the template for how compute gets allocated in 2026-2028.

Sources: SpaceX S-1 prospectus (May 20, 2026), CleanTechnica, NetworkWorld, Reuters, x.ai/news/anthropic-compute-partnership, Investing.com, Morningstar.