Anthropic Locks Down SpaceX's Colossus 1: 300+ MW and 220K NVIDIA GPUs in a Month
At Code with Claude SF on May 6, Anthropic announced an exclusive deal for SpaceX's entire Colossus 1 data center — adding 300+ MW of new capacity and 220,000+ NVIDIA GPUs within the month. What it signals about Mythos compute demand.
TL;DR: On May 6, 2026, at Code with Claude San Francisco, Anthropic announced an exclusive partnership for SpaceX’s Colossus 1 data center. Anthropic gets the entire facility — adding more than 300 MW of new capacity and over 220,000 NVIDIA GPUs within the month. The deal lands less than four weeks after the official Mythos Preview launch and signals just how compute-hungry the Capybara tier really is.
What was announced
Anthropic and SpaceX confirmed an exclusive use agreement for SpaceX’s Colossus 1 facility. Per the announcement at Code with Claude SF:
- 🟢 Capacity addition: more than 300 MW of new compute capacity
- 🟢 GPU count: over 220,000 NVIDIA GPUs operational within the month
- 🟢 Exclusivity: Anthropic gets the entire Colossus 1 capacity, not a slice
Compared to typical hyperscaler builds, this is an unusual move. Most large AI labs split capacity across multiple cloud regions and partners. Locking down a single new-build facility, exclusively, suggests Anthropic anticipates sustained compute demand at a scale that doesn’t fit neatly into existing AWS / Azure / GCP allocations.
Why this matters for Mythos
Mythos Preview — the model behind the Capybara tier — was officially released April 8 with restricted access. Anthropic’s own materials described it as “far ahead of any other AI model” but noted the model is “very expensive to serve.” That phrase started making more sense once pricing was confirmed at roughly 5x Opus 4.6 rates.
A 300 MW exclusive buildout suggests three things:
- 🟡 Mythos is even more compute-hungry than Anthropic let on publicly. Pricing alone hinted at it, but committing to an entire data center within four weeks of launch is a different magnitude of signal.
- 🟡 Limited access is going to widen, but slowly. With this much capacity coming online, Anthropic has the runway to extend Mythos beyond the initial 11 Glasswing founding partners and 40+ extended-access organizations — but the rollout is being rationed against compute, not gated by policy.
- 🟢 Code with Claude is now a major announcement venue. This is the same conference series Anthropic used to launch Claude 4 in May 2025. Putting the SpaceX deal here, instead of a standalone press release, reads as a deliberate signal that future model and infrastructure milestones land at these events.
What’s still unconfirmed
🔴 The annual run-rate cost of leasing Colossus 1 has not been disclosed. Industry analysts estimate $1.5B–$2.5B/year for a 300 MW exclusive facility at current pricing, but neither party confirmed.
🔴 Whether SpaceX is also using compute on the same facility for Starlink AI workloads is unclear from the public announcement.
🔴 The split between training and inference on this capacity hasn’t been broken out.
Context: the IPO question
Anthropic is reportedly preparing for a public listing later in 2026. A 300 MW exclusive compute lease is the kind of capital commitment that has to be disclosed in S-1 filings, and the structure of the deal — exclusive use, presumably long-term — will be analyzed by underwriters and analysts when the prospectus drops.
For Mythos users, the practical takeaway is shorter wait times for Mythos API capacity expansion. For competitors and the broader AI infrastructure market, this is one of the largest single announced compute deals of 2026.