Anthropic officially filed its S-1 with the SEC on October 26, and buried in the compute-contracts section is a supplier this site didn’t see coming when we covered the leaked draft two days earlier: xAI, Elon Musk’s data-center arm and a SpaceX subsidiary since February 2026, good for up to $120.5 billion in GPU capacity through 2029. Most of that capacity runs through xAI’s Colossus 1 cluster in Memphis, with the rest spread across a second SpaceX data center — meaning Anthropic is paying Musk’s company tens of billions of dollars to rent the same GPUs that train Grok, one of Claude’s direct competitors.

That’s not a typo and it’s not a hedge-your-bets side deal. It’s the second-largest compute commitment in the entire filing, behind only Google.

Quick Summary: What’s In the Official S-1

Detail	Info
Filed	Oct 26, 2026, with the SEC — ticker ANT, Nasdaq Global Select Market
xAI compute commitment	Up to $120.5 billion through 2029
Run rate once fully delivered	~$3.1 billion/month ($37.2 billion/year), expected December 2026
Where it runs	Mostly xAI’s Colossus 1, Memphis (220,000+ Nvidia GPUs, 300+ MW), plus a second SpaceX data center
Exit terms	Either side can cancel with 90 days’ notice after lock-ins end Jan/Mar 2027
Other disclosed suppliers	Google ($114.5B), Amazon ($110B), Microsoft ($44.9B)
H1 2026 revenue	$16.3B — 80% consumption-based ($13.1B), 20% subscription ($3.2B)

Bottom line: The leaked draft’s scary number was Anthropic’s hyperscaler dependency. The filed version adds a stranger wrinkle — a multibillion-dollar GPU lease from the company that makes Claude’s most aggressive rival, with an exit clause either side can pull in 90 days.

What Actually Happened

Our post on the leaked draft S-1 covered three named compute suppliers — Google Cloud, AWS, and Microsoft — and a risk factor about hyperscalers treating Anthropic as a rival instead of just a customer. That reporting, sourced from Fortune, Dealroom, and others reviewing a Wall Street Journal draft, listed roughly $750 billion in aggregate future compute commitments across those three.

The filed version tightens some of those numbers and adds one the leak never mentioned. According to the S-1 and reporting from Benzinga and the Motley Fool, the officially disclosed minimums are: Google Cloud at least $114.5 billion (April 2026–July 2033), Amazon at least $110 billion (2026–2036), and Microsoft at least $44.9 billion (November 2026–May 2033) — all structured as pay-regardless-of-usage commitments. Add those three and you get roughly $269 billion, a materially different shape than the leak’s single $750 billion aggregate.

Then there’s the fourth name. Per the S-1 and corroborating coverage from Drive Tesla and Blockspace, Anthropic entered a series of agreements with xAI between May and October 2026, spanning Colossus 1 and a second SpaceX data center, and committing up to $120.5 billion in compute spend through 2029. Monthly spend is expected to hit roughly $3.1 billion — about $37.2 billion a year — once full capacity is delivered, which the filing projects for December 2026. The initial non-cancelable periods under the separate agreements run out in January and March 2027; after that, either Anthropic or SpaceX can walk away on 90 days’ written notice.

None of the four named suppliers were secret before this filing. What’s new is seeing exact dollar figures and exit terms attached to all of them in one document, filed with the SEC instead of leaked to a reporter.

Why Pay Your Rival’s Parent Company for GPUs?

Here’s where it gets genuinely odd. The compute Anthropic is renting from xAI doesn’t come from some neutral SpaceX facility built for the deal. Most of it runs through Colossus 1, the Memphis data center xAI built to train Grok, with the remainder at a second SpaceX site.

What is Colossus 1? Colossus 1 is a Memphis, Tennessee data center originally built by xAI, Elon Musk’s AI company, to train its Grok models. It houses more than 220,000 Nvidia GPUs and draws over 300 megawatts of power — one of the largest single AI training clusters in the world.

xAI became a wholly owned subsidiary of SpaceX in a $1.25 trillion all-stock merger that closed February 2, 2026 — SpaceX’s biggest deal yet, but not a dissolution. xAI kept shipping Grok under its own name after the close, the same way it has every release since, including Grok 4.7 in September. So the entity Anthropic is now paying up to $120.5 billion isn’t some arm’s-length infrastructure vendor. It’s the company that makes Grok, run by the same person, using the same hardware Grok trains on, now sitting one level down in SpaceX’s corporate structure instead of standing alone.

This relationship isn’t brand new, either. TechCrunch first reported in May 2026 that Anthropic would pay xAI $1.25 billion a month to rent all of Colossus 1’s capacity, a few months after xAI became a SpaceX subsidiary. Axios and DataCenterDynamics each confirmed the figure independently that same week, citing SpaceX’s own IPO filing. What the S-1 adds, five months later, is the full contracted ceiling — $120.5 billion through 2029 — and confirmation that the relationship survived, and grew, straight through the corporate reshuffling that put Grok’s maker and Anthropic’s GPU landlord under the same roof.

What Does Anthropic’s SpaceX Deal Actually Include?

1. Multiple agreements, not one contract — and not a free handover. The S-1 describes a series of agreements signed between May and October 2026, not a single master deal. Earlier coverage of this relationship described Anthropic getting all of Colossus 1’s capacity in terms that read like a facility transfer; the S-1 clarifies it’s a paid, contracted lease that started smaller and scaled.
2. Two facilities, not one. The contracted capacity spans Colossus 1 and a second SpaceX data center, with Colossus 1 carrying the bulk of the spend.
3. A hard capacity ceiling. Up to $120.5 billion in total spend through 2029, with monthly run-rate reaching roughly $3.1 billion once Anthropic takes full delivery.
4. A near-term delivery date. Full capacity is expected by December 2026 — weeks, not years, from the filing date.
5. Real cancellation rights. Unlike the decade-plus, largely noncancelable deals with Google, Amazon, and Microsoft, this one lets either side exit with 90 days’ notice after the initial lock-in periods lapse in January and March 2027.

That last point is the detail worth sitting with. Anthropic’s hyperscaler contracts are structured to survive a bad quarter. The SpaceX deal is built to be renegotiated or ended within a year of full delivery — which tells you something about how both sides are pricing the risk of this particular arrangement.

Why This Matters

The leaked draft’s hyperscaler risk factor was already strange: Anthropic disclosing that the companies it depends on for compute — Google, Amazon, Microsoft — increasingly compete with it through their own models. That’s a supplier-is-also-a-competitor problem, but at least those three aren’t primarily in the business of building a rival chatbot. They’re diversified giants with AI as one line of business among many.

xAI exists to build Grok. Full stop. That’s its entire reason for existing, the thing we’ve covered repeatedly as it ships faster models with fewer guardrails than most of the industry is comfortable with. Anthropic is now sending tens of billions of dollars to the company bankrolling Grok’s training runs, for GPU capacity Anthropic needs to train and run Claude. Every dollar Anthropic pays for Colossus time is a dollar that also funds the infrastructure Musk’s team uses to compete with Claude directly.

Compare that to Replit’s decision to shift spend from Anthropic to OpenAI over a pricing dispute — a reversible, single-customer call. The SpaceX arrangement isn’t reversible in the same way, but it is bounded: a 90-day exit window after early 2027 means Anthropic isn’t locked into funding a rival’s infrastructure for the full life of the IPO the way it’s locked into Google through 2033. Whoever negotiated those exit terms for Anthropic clearly understood exactly what kind of supplier they were dealing with.

It’s also worth asking the obvious question: why would Anthropic do this deal at all, when distrust between the two labs should, on paper, run both directions? The likely answer is the one that explains most of this filing — Anthropic needs GPU capacity faster than its hyperscaler partners can reliably deliver it, a constraint we flagged in the leaked-draft coverage, and its $45 billion Nscale deal from August shows the same pattern: when Google, Amazon, and Microsoft can’t or won’t prioritize Anthropic’s capacity requests, Anthropic goes shopping with whoever has idle GPUs. In October 2026, that happened to be Musk.

The Bigger Picture

Zoom out and the total compute picture gets bigger than any single number in this post. Google at $114.5 billion, Amazon at $110 billion, Microsoft at $44.9 billion, and xAI at up to $120.5 billion add up to roughly $390 billion across those four named suppliers alone. Layer in Broadcom chip leases and other contracts the S-1 discloses, and outside analysts tallying the full filing — including a Fortune analysis of the filing — put Anthropic’s total compute and infrastructure commitment at up to $567 billion.

That’s a genuinely different shape than the leaked draft’s tidy $750 billion aggregate across three hyperscalers. The real filing spreads the exposure across more counterparties, with wildly different contract terms, including one counterparty whose core business is building a product that competes with the thing all that compute is meant to produce.

The revenue side of the filing is the other number worth holding onto, and we already walked through the topline figures two days ago. What that earlier coverage didn’t have is the revenue mix buried in the official filing: 80% of H1 2026 revenue, $13.1 billion, is consumption-based — customers paying per token — while the remaining 20%, $3.2 billion, comes from subscriptions. That split matters against the SpaceX number specifically, because consumption revenue scales with usage, and usage scales with available compute. A $120.5 billion GPU lease isn’t a side bet. It’s load-bearing for the revenue line investors are being asked to price.

Anthropic is also set to list under the ticker ANT on the Nasdaq Global Select Market, a detail confirmed by CNBC’s live coverage of the filing. Compared to the back-and-forth over a possible $2 trillion valuation we covered against SpaceX’s own IPO record in August, Anthropic hasn’t yet disclosed share count or price range — the usual next step before a roadshow. And OpenAI’s competing S-1 process is still running in parallel, which means the two frontier labs are about to get priced against each other with both filings finally public instead of leaked.

Our Take

I think the real story in this filing isn’t the dollar figure, it’s the paperwork. $120.5 billion sounds like a vote of confidence. A 90-day mutual exit clause says the opposite.

Look at the shape of the contract next to the other three. Google, Amazon, and Microsoft got multi-year, largely noncancelable terms running out to 2033 and 2036 — the kind of commitment you write when you expect to still be doing business together a decade from now. xAI got lock-in periods that expire within five months of signing, then an open door either side can walk through on three months’ notice. You don’t structure a deal that way with a partner you’re betting the business on. You structure it that way with someone you need right now and don’t fully trust to still be useful, or friendly, later.

Flip it and look from Musk’s side and the logic holds just as well. SpaceX didn’t tie xAI into a decade of Colossus rent the way Anthropic tied itself to Google. It kept its own exit ramp, which only makes sense if xAI wants the option to pull that capacity back for Grok the moment it needs it more than it needs Anthropic’s money.

Short lock-ins like this show up when neither side wants to hand the other leverage it can’t take back. Anthropic needed GPUs faster than its hyperscaler partners could deliver them. xAI needed revenue for capacity it wasn’t fully using. Both sides took the deal, and both sides kept an exit in their back pocket. My read: this is the first of Anthropic’s four named compute contracts either side is likely to let lapse, and it’ll happen the moment the broader GPU shortage eases enough that neither company needs the other anymore.

Frequently Asked Questions

How much is Anthropic paying SpaceX for compute?

Up to $120.5 billion through 2029, according to Anthropic’s S-1 filed October 26, 2026. Monthly spend is expected to reach roughly $3.1 billion — about $37.2 billion annualized — once Anthropic takes full capacity, which the filing projects for December 2026.

Why does Anthropic’s SpaceX deal involve xAI and Grok?

Most of the compute Anthropic is renting runs through Colossus 1, a Memphis data center originally built by xAI to train Grok, plus a second SpaceX data center. SpaceX acquired xAI as a wholly owned subsidiary in a $1.25 trillion merger that closed February 2, 2026; xAI kept shipping Grok under its own name afterward. Anthropic’s contract is with xAI, which makes a chunk of its compute spend effectively flow to the company building a direct Claude competitor, now one level down in SpaceX’s corporate structure rather than fully independent.

Can Anthropic or SpaceX cancel the compute deal?

Yes, after an initial lock-in period. The agreements have non-cancelable periods ending in January and March 2027, respectively. After those dates, either Anthropic or SpaceX can terminate with 90 days’ written notice — a much shorter leash than Anthropic’s multi-year commitments to Google, Amazon, and Microsoft.

What ticker will Anthropic trade under?

ANT, on the Nasdaq Global Select Market, according to the S-1 and confirmed by CNBC’s coverage of the filing. Anthropic hasn’t yet disclosed share count or price range.

How does the filed S-1 compare to the leaked draft this site covered two days earlier?

The leaked draft, reported by Fortune and corroborated by Dealroom and others, cited roughly $750 billion in aggregate future compute commitments across Google, AWS, and Microsoft. The officially filed S-1 breaks that down into specific minimums — $114.5 billion with Google, $110 billion with Amazon, $44.9 billion with Microsoft — and adds a fourth supplier the leak never mentioned: xAI, at up to $120.5 billion.

What was Anthropic’s revenue in the first half of 2026?

$16.3 billion, per the S-1. Of that, 80% ($13.1 billion) was consumption-based revenue from per-token API usage, and 20% ($3.2 billion) came from subscriptions.

Does this affect Claude pricing or availability for existing users?

Not directly, and not immediately. Nothing in the S-1 changes current Claude API or subscription terms. The practical relevance is longer-term: Anthropic’s compute capacity, and therefore its ability to serve growing demand without rationing, now depends in part on a supplier relationship with a company whose own incentives point toward prioritizing its own model.

---

Last updated: October 27, 2026. Sources: Anthropic S-1 filing, SEC EDGAR · Benzinga — Anthropic’s IPO Filing Reveals Over $500 Billion AI Compute Spending · The Motley Fool — 5 Key Numbers From Anthropic’s S-1 · Drive Tesla — Anthropic Reveals $120.5 Billion SpaceX AI Compute Deal · TechCrunch — Anthropic Will Pay xAI $1.25B Per Month · CNBC — Musk’s xAI, SpaceX Combo Is the Biggest Merger Ever · CNBC — Anthropic IPO Live Updates · Fortune — Anthropic Plans to Spend Up to $567 Billion on Compute.

Related reading: Anthropic’s Leaked S-1: An $80.7B Red Flag · Grok 4.7 Ships After 5 Delays, SpaceX Data Inside · Anthropic’s $45B Nscale Deal · Anthropic’s IPO Could Top SpaceX · OpenAI Files for IPO