Quoting SpaceX S-1

## SpaceX-Anthropic: $1.25B/month — What the S-1 Actually Reveals

### 1. The Raw Numbers

SpaceX's SEC S-1 filing formalizes a deal signed in May 2026: Anthropic pays **$1.25 billion per month** for compute access across COLOSSUS and COLOSSUS II through May 2029. That's a potential contractual commitment of **~$45 billion over three years**, subject to a mutual 90-day termination clause. Capacity ramps at a reduced fee in May-June 2026, implying full billing likely kicks in around July 2026.

For scale: Anthropic had raised roughly $7.3B in cumulative funding through end-2024 (Google, Amazon, Spark Capital). A single month of this agreement equals ~17% of all capital raised over several years. Even assuming Amazon and Google cover portions through existing cloud commitments, the cash flow pressure on Anthropic is structurally severe.

### 2. Why COLOSSUS and Not AWS/GCP?

Anthropic already has deep agreements with AWS ($4B investment in 2023, another $4B in 2024) and Google Cloud ($2B in 2023). Turning to COLOSSUS signals several things:

- **Real GPU capacity constraints**: AWS and GCP cannot absorb Anthropic's frontier training needs at the scale required for post-Claude 3.5 models. COLOSSUS II, built by xAI/SpaceX, is one of the densest H100/H200 clusters in existence (100,000+ GPUs in phase one, expansion ongoing). - **Partial strategic independence**: Diversifying compute suppliers reduces dependency on Amazon and Google, who are simultaneously investors, customers, and competitors (Bedrock, Vertex AI). - **Availability timing**: xAI built COLOSSUS in ~122 days in 2024. COLOSSUS II is mid-deployment. SpaceX/xAI had capacity to sell before hyperscalers caught up.

### 3. The Structural Tension: Grok 5 vs. Claude

The S-1 explicitly states that SpaceX/xAI is using **these same resources to train Grok 5** while Anthropic rents them. This is not a footnote. It means:

- Anthropic is paying a direct competitor (xAI) to train its own models on infrastructure shared with adversarial models. - The symmetric 90-day termination clause gives xAI the ability to reclaim capacity if Grok 5 or a priority training run demands it — and gives Anthropic an exit if a better deal materializes. - The public S-1 text contains no workload isolation guarantees. Questions around hyperparameter confidentiality, training data separation, and intermediate checkpoint security remain entirely open.

### 4. Potential Losers and Sector Implications

**AWS and Google Cloud** are the immediate losers: a slice of Anthropic's compute budget that could have flowed to their infrastructure is going to a third party. For AWS, whose agreement designated Anthropic's use of AWS as its "primary cloud provider," this SpaceX deal warrants close reading of exclusivity clauses.

**CoreWeave, Lambda Labs, Oracle Cloud** — independent GPU cloud providers — see confirmation that $10B+ multi-year deals are achievable, but also that SpaceX/xAI is now a direct competitor in that segment.

**Other frontier labs** (OpenAI, Google DeepMind, Meta) note that Anthropic accepted pricing that implies, back-of-envelope, roughly $12.50/GPU-hour assuming 100,000 H100s at full utilization — a significant premium over spot rates, justified by guaranteed dedicated capacity.

Finally, the fact that this deal appears in a **SpaceX S-1** — a pre-IPO document — turns Anthropic into a sales argument for valuing xAI infrastructure. SpaceX monetizes COLOSSUS at two levels simultaneously: recurring revenue from Anthropic, plus amplified public market valuation from demonstrating a customer paying $15B/year.