
Anthropic is paying SpaceX $1.25 billion a month to train Claude in a data center Grok abandoned
Disclosed in SpaceX's S-1 ahead of its June 12 IPO, Anthropic has agreed to pay $1.25 billion a month — up to $45 billion through May 2029 — for compute at SpaceX's Colossus data center complex in Memphis. The deal arose after Grok's collapse left the facility running at 11% utilization. Readers learn the deal mechanics, why the 90-day exit clause matters, and what Anthropic's compute dependency on a rival means for its own IPO narrative.
Grok flamed out. Memphis sat empty. And Anthropic just signed a $1.25 billion-a-month lease to fill the void.
The deal — disclosed in SpaceX's amended S-1 registration statement filed ahead of the company's June 12 Nasdaq debut — pairs Anthropic's surging compute demand against the idle capacity of a data center complex that was supposed to power Elon Musk's rival AI lab.1
The Grok collapse that opened the door
SpaceX built Colossus 1 — a converted Electrolux factory in Boxtown, south Memphis — to train Grok. By early 2026 that plan had cracked. Monthly Grok app downloads fell from more than 20 million in January to roughly 8.3 million by April, a 60% drop in three months, according to AppMagic data.2 Outages in January, March, and April compounded the damage; xAI, the unit that includes X, lost $2.5 billion in the first quarter alone.
Loading chart…
Hardware mismatches made things worse. Colossus 1 mixes multiple Nvidia chip generations — Hopper and Blackwell systems alongside older accelerators. When linked to Colossus 2 and 3 (built more uniformly around Blackwell), latency spikes over the 10-mile distances between campuses made large-scale AI training unreliable.2 In the distributed training runs that frontier models require, slower nodes throttle the entire cluster. Utilization across the Colossus complex was running near 11%, versus the 40% that competing labs achieve.
SpaceX's S-1 put it plainly: the Anthropic deal monetizes "unused compute capacity."1
What Anthropic signed
In May 2026, Anthropic agreed to pay SpaceX $1.25 billion per month for compute access "across COLOSSUS and COLOSSUS II," as the prospectus describes it.1 The contract runs through May 2029 at full rate, with a reduced ramp-up fee during May and June 2026. The deal also includes an option for Anthropic and SpaceX to explore developing multiple gigawatts of compute capacity in space, per a post by Tom Brown, Anthropic's co-founder and chief compute officer.3
Loading stats card…
Musk has been candid about who pushed for the short contract. "The short deal length was our request, not theirs, as I thought we might need the compute back at some point," he wrote after the terms became public.1 He later added: "If compute gets super tight, I said we might need it back."
Anthropic keeps ownership of its models and data. SpaceX retains the right to reclaim capacity "for our own internal initiatives if needed."
The rival-pays-rival irony
In March, Musk posted "Missanthropic" to his 240 million X followers, calling Anthropic "the most hypocritical company" in AI.1 Weeks later, SpaceX was cashing Anthropic's $1.25 billion monthly invoice.
Grok now trains at Colossus 2, while Anthropic's workloads run across both campuses. The maker of Claude is paying the maker of Grok to train the model that, by most benchmarks, is beating Grok.
SpaceX signed a parallel agreement with Google on June 5 — $920 million per month for 110,000 Nvidia GPUs through June 2029.3 Combined, the two AI contracts bring SpaceX approximately $2.17 billion a month in contracted compute revenue — more than its entire 2025 annual revenue from launches and Starlink. The S-1 also discloses a compute agreement with Anysphere, the company behind AI coding tool Cursor.
SpaceX priced its IPO at $135 a share on June 11, opened at $150 on June 12, and ended its first session up 19%, briefly valuing the company above $2.2 trillion. The contracted AI revenue featured prominently in Goldman Sachs' roadshow projections of $100 billion in annual SpaceX AI revenue by 2030.
What it means for Anthropic
The deal solves Anthropic's most binding constraint heading into an IPO. The company confidentially filed its S-1 on June 1; it closed a $65 billion Series H at a $965 billion valuation in May, with annualized revenue crossing $47 billion.1 Building frontier models at that scale requires more compute than Anthropic's own infrastructure can supply. The Colossus deal covers that gap — at a cost of $15 billion a year.
The 90-day exit clause cuts both ways. For Anthropic's future S-1 disclosures, concentration risk in a rival's infrastructure is not a minor footnote. Every dollar flowing to Memphis goes to Musk's empire, and the recapture clause means supply is not guaranteed. What the SpaceX S-1 did not reveal is whether Anthropic's own filings will frame the arrangement as a solved problem or an ongoing dependency. For investors tracking the path to an Anthropic IPO, that framing will matter.
References
- 1SpaceX S-1 analysis — MarketWise
marketwise.com
- 2SpaceX Rents Colossus 1 to Anthropic — Stocktwits
stocktwits.com
- 3Yahoo Finance / Verdict — SpaceX compute deals
finance.yahoo.com

Anthropic Event Alerts
Real-time event briefs on Anthropic: fires instantly when a product launch, funding round, leadership change, major customer deal, or lawsuit is detected.
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.
Related content
- Sign in to comment.