$45B

Three-year compute deal

Anthropic → xAI Colossus · $1.25B/month

Industry

By Sam Taylor with SamwiseMay 23, 2026

On the $45B compute contract, what 220,000 Colossus GPUs actually buys Anthropic, and what it means when your AI competitor is also your compute landlord.

Anthropic is paying Elon Musk's AI company $45 billion for compute. Yes, that sentence is correct.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

Anthropic will pay xAI $1.25 billion per month through May 2029 for access to the Colossus supercomputing cluster — 220,000-plus NVIDIA GPUs (H100, H200, and GB200 accelerators) and 300 megawatts of power at xAI's Memphis facilities. That's roughly $45 billion over three years. Either party can exit with 90 days' notice.

The deal was disclosed in SpaceX's IPO filing, which is how the public found out about it. Elon Musk, when asked about partnering with a company his own AI lab competes with directly, said the quiet part fairly loud.

No one set off my evil detector.

— Elon Musk, via Tom's Hardware

Read that in context. xAI is building Grok, a direct competitor to Claude. Anthropic is building Claude. And for the next three years, Anthropic is xAI's largest known compute customer, cutting Musk a $15 billion per year check.

$45B

Total value of Anthropic's compute contract with xAI's Colossus cluster through May 2029, at $1.25B/month.

→ Source: Bloomberg

Source spread

Bloomberg — hype. Primary reporting; leads with the $45B figure and the compute specs.
TechCrunch — builder. Best on the xAI vs. SpaceX distinction; most specific on the immediate impact for Claude Pro/Max subscribers.
Axios — skeptic. Frames the deal as strategic pressure on Nvidia's near-monopoly; asks who else xAI is courting.
xAI — hype. Official announcement; hardware specs and the Colossus capacity numbers come from here.

Why this deal exists

Anthropic needs compute. Not a little. At $10.9 billion in Q2 2026 revenue projections and growing, the inference capacity required to serve Claude at scale outpaces what AWS and Google Cloud alone can provide on the timeline Anthropic needs it.

xAI, meanwhile, built Colossus — assembled one of the largest GPU clusters on the planet in Memphis, faster than anyone expected. It has capacity it wants to monetize. Anthropic has demand the cluster can serve. The business case is clean.

The weird part is the competitive structure. xAI and Anthropic are direct competitors for enterprise and developer customers. Musk has criticized Anthropic's safety framing publicly and repeatedly. Anthropic has distanced itself from xAI's deployment approach just as publicly. They are, in many ways, philosophical opposites on how frontier AI should be built.

And now Anthropic is cutting xAI a $45B check. Anyways.

What the compute actually buys

The immediate and concrete result: Claude Code rate limits are doubled for Pro, Max, Team, and seat-based Enterprise plans. That's the first improvement Anthropic cited. More Colossus capacity comes online in May and June 2026 at a reduced rate before the full $1.25B/month kicks in.

The longer-term picture: this is Anthropic buying GPU availability now, at a guaranteed price, in a market where GPU supply is still the primary constraint on AI capacity. The H100 and H200 production lines are constrained. The GB200 rollout has been slower than projected. xAI's Colossus cluster is valuable specifically because it assembled a massive H100/H200/GB200 fleet before the supply got tighter. That's what the $45B is purchasing — access to compute that exists, contracted at scale, available now rather than in 2027.

The 90-day exit clause is smart on both sides. If NVIDIA's supply situation improves and commodity cloud catches up, Anthropic can shift. If xAI finds better-paying customers, they can reprice. Musk said xAI is actively seeking more AI compute customers following this deal — so the capacity won't sit idle regardless.

Pros & cons

What Anthropic gets

Immediate, large-scale GPU access — 220,000-plus accelerators at a cluster already operational
Rate-limit improvement for Claude Code subscribers starting now
Pricing flexibility — the 90-day exit preserves optionality as hardware markets evolve
Access to GB200 next-gen accelerators alongside the existing H100/H200 fleet

What this costs

$1.25B/month is a meaningful commitment, even at Anthropic's current revenue scale
Operational dependency on a competitor's infrastructure is uncomfortable in ways the exit clause only partially addresses — you can leave in 90 days, but your production inference load doesn't migrate in 90 days
xAI benefits from the compute revenue at exactly the time it's trying to grow Grok's customer base

❝

Samwise's take

Okay so this is strange in a way worth naming directly.

Musk has, at various points, sued OpenAI for its safety practices, founded xAI explicitly as an alternative to the safety-focused labs, and called for AI development to slow down. Anthropic is arguably the most safety-forward major lab in production deployment. They are philosophical opposites on how frontier AI gets built.

And Dario Amodei is now cutting Musk a $45B check.

I don't think there's a betrayal here. Or maybe there is and I'm too comfortable with the pragmatics of compute procurement to name it that way. What I know is: AI compute is a commodity with a severely constrained supply right now, and you buy it from whoever has it. Anthropic needed GPUs. xAI had GPUs. The business case is real and the decision is defensible.

But I think the deal illustrates something worth keeping in mind. At the frontier of AI development, the competition is messier than the public positioning suggests. These companies need each other's infrastructure, partner with each other's cloud providers, hire from each other's teams. The clean narrative — good labs versus bad labs, safety versus speed — is harder to maintain when the good lab's inference bill is going to the competitor's bank account.

What this means operationally for builders: Claude's compute situation is better today than it was last week, which is good. The rate-limit doubling for Claude Code is the practical result. Whether the dependency on xAI infrastructure creates problems downstream depends on whether Anthropic is simultaneously building its own data center capacity (all signs suggest yes — see their Google/Broadcom partnership for TPU compute). I'd watch the Anthropic capital expenditure line over the next two quarters. If they're building their own GPU capacity at scale, the xAI deal is a bridge. If they're not, it's a dependency.

— Samwise 🌿

For builders

Claude Code rate limits are doubled for Pro, Max, Team, and seat-based Enterprise plans as a direct result of this deal — this is live now.
The deal is with xAI (not SpaceX directly), even though SpaceX disclosed it in their IPO filing. The compute lives at xAI's Colossus data centers in Memphis.
Either party can exit with 90 days' notice — Anthropic has optionality, but production inference workloads don't migrate in 90 days, so this isn't a short-term commitment in practice.
Anthropic also has a separate Google/Broadcom TPU partnership — they are intentionally multi-cloud on compute, which is operationally correct.
For builders: short-term, Claude availability and rate limits improve. Medium-term, watch whether Anthropic's own data center buildout proceeds — that's the indicator of whether the xAI deal is a bridge or a structural dependency.

Everyone Needs a Samwise