Anthropic Pays Elon Musk $1.25 Billion Per Month to Run Its AI. It Was Founded to Get Away From Him.

On May 20, 2026, SpaceX filed its S-1 with the SEC in preparation for the IPO it completed on June 12. Buried in the filing: Anthropic had agreed to pay xAI $1.25 billion per month for access to Colossus 1 — 300 megawatts of compute across 220,000-plus NVIDIA GPUs in Memphis, Tennessee — through May 2029. Total potential deal value: $40 to $45 billion.

Anthropic did not announce this. The deal became public because SpaceX had to disclose its largest enterprise customer in its IPO filing.

This is the context: Anthropic was co-founded in 2021 by Dario Amodei, Daniela Amodei, and other researchers who left OpenAI partly due to concerns about Elon Musk's influence over that organization. Musk was an early OpenAI backer and board member who departed in 2018. Anthropic was built, in part, as an alternative — a safety-focused lab independent of both Musk and the Microsoft-backed entity they left. Elon Musk is now Anthropic's largest external compute supplier.

How the deal came to exist

Anthropic needed compute. The company has signed letters of intent for 1 gigawatt-plus of data center capacity, but non-binding LOIs are not operational infrastructure. The Claude demand surge — comparable deployments reporting 180 trillion-plus daily token calls — outpaced the timeline for Anthropic's own buildout. Colossus was built, available, and underutilized. xAI overbuilt relative to Grok's commercial traction: $7.7 billion in Q1 2026 capex, $818 million in Q1 revenue. The company lost $6.36 billion in all of 2025 on $3.2 billion in revenue.

At $1.25 billion per month, the Anthropic deal is approximately 4.7× xAI's entire 2025 revenue. xAI needed a customer. Anthropic needed the compute. The arrangement was expedient for both parties and enthusiastically chosen by neither.

Musk was publicly attacking Anthropic on X — "misanthropic," "evil," retweeting calls for Anthropic employees to leave the company — approximately three months before the deal was signed.

What the 90-day clause means

Either party can terminate the arrangement with 90 days' notice.

The standard framing of this clause is symmetrical: Anthropic can walk, xAI can walk. The actual asymmetry is significant.

If Anthropic terminates: xAI loses $15 billion per year — nearly 5× its 2025 revenue — and is left with 300MW of idle compute it cannot currently fill with its own Grok workloads. The financial pressure on xAI would be severe. xAI cannot afford Anthropic walking away.

If xAI terminates: Anthropic loses 300MW of compute. The question — which nobody covering this deal has answered — is whether that 300MW is being used for training (building future Fable/Mythos models) or for inference (serving current Claude users). These are very different risk profiles.

If it's training compute, a termination notice gives Anthropic 90 days to find alternative infrastructure for future model development. That's painful and would delay model releases, but it doesn't affect users today.

If any portion of Colossus is serving current Claude inference requests, a Musk termination notice doesn't just pause future training. It degrades or interrupts current production services with 90 days' warning.

On June 23, Claude experienced an outage generating over 8,000 reports — agentic pipeline failures reported across user threads. The outage was not attributed to Colossus. But the infrastructure dependency question is now urgent in a way it wasn't before. If Anthropic's production inference runs on a Musk-controlled data center that can be switched off on a quarter's notice, that is a materially different risk than a training delay.

Anthropic has not disclosed which workloads run on Colossus.

The governance paradox

Anthropic's founding documents position it as a public benefit corporation: AI development controlled by researchers focused on safety and alignment, independent of commercial actors with competing interests. The Responsible Scaling Policy, the Constitutional AI framework, the published interpretability research — all of it rests on the premise that Anthropic's research direction is not subject to external commercial leverage.

Musk said in May: the deal is "short-term because we might need the compute back."

This statement is either a pricing signal for renewal negotiations — Musk creating leverage ahead of the May 2029 expiration — or a genuine statement that xAI plans to scale its own Grok training to fill the 300MW as its revenue grows. If the latter, Anthropic needs its own infrastructure operational before that happens, or needs to find alternative capacity before Musk decides he wants his GPUs back.

The governance question is not abstract. An AI safety lab conducting research on alignment and model behavior has its primary external compute infrastructure controlled by an individual who has publicly characterized the company as "evil" and who runs a competing AI lab. The 90-day termination clause means this leverage is not theoretical.

Two IPO filings, one financial dependency

SpaceX completed its IPO on June 12 at approximately $1.75 trillion valuation. Its S-1 disclosed Anthropic as the largest enterprise compute customer — the Anthropic deal improves SpaceX/xAI's revenue picture for investors evaluating the filing.

Anthropic filed a confidential S-1 in June, targeting an October 2026 Nasdaq listing at approximately $1 trillion valuation. Its public S-1 must disclose xAI/SpaceX as its largest external compute supplier.

The two IPO filings are financially linked. SpaceX's revenue projections include Anthropic's $15 billion per year. Anthropic's cost projections include xAI's $15 billion per year. Institutional investors evaluating Anthropic in October will be looking at a company that is its largest competitor's largest customer and whose primary external compute infrastructure is held on 90-day termination notice.

The test: find the risk factor

When Anthropic's public S-1 is filed, two sections will reveal how the company internally weighs its risks.

The first: the Colossus/xAI supplier risk disclosure. How does Anthropic characterize the 90-day termination clause? Does it name the termination risk explicitly, including Musk's public statement that xAI may need the compute back? Or does it frame the deal as a routine infrastructure agreement with a standard exit provision?

The second: the BIS/export control risk disclosure. Anthropic's Fable 5 model was suspended under EAR section 744.22(b) export controls in June. The company must disclose this as a material business risk.

Count the words. If the Musk compute dependency gets fewer words — less materiality language, smaller type in the risk hierarchy — than the government export control risk, Anthropic has made a public judgment: BIS is more dangerous to the business than Elon Musk with a 90-day notice window.

That editorial judgment, coming from the company that was built to get away from Musk's influence, is the most revealing sentence Anthropic will write in 2026.

What the deal actually reveals

I think the Colossus arrangement tells you more about the state of frontier AI infrastructure than about Anthropic's values. The compute shortage is real. The buildout timelines are measured in years; the demand curves are measured in months. When you need 300MW of frontier GPU compute now, your options are: Amazon (committed to OpenAI through various arrangements), Google (your investor, may have conflict of interest concerns at scale), Microsoft (OpenAI's primary partner), or xAI, which built too much and needs the cash.

Anthropic chose the available option. The fact that the available option was Musk's company is a feature of the infrastructure landscape, not a values failure.

The governance risk is real anyway. The 90-day clause exists. Musk's public hostility to Anthropic is documented. The training-vs-inference question is unanswered. These facts don't require a values judgment to be material — they're operational risks that belong in the S-1 and in any honest assessment of Anthropic's infrastructure position.

The founding paradox makes a good headline. The 90-day clause is the actual story.