Anthropic leased SpaceX's Colossus 1 — and doubled Claude Code limits

On May 6, Anthropic made a short but heavy announcement: all compute capacity at SpaceX's Colossus 1 data center — 300+ megawatts, 220,000+ NVIDIA GPUs — is going to Anthropic and coming online within the month. In the upper half of the same blog post, something far more practical was announced: Claude Code's five-hour rate limits doubled for Pro, Max, Team, and Enterprise; peak hours throttling removed for Pro and Max; Claude Opus API rate limits substantially raised.

The first story is huge on the infrastructure side. The second is huge on the developer's daily workflow side. Both for the same reason: capacity bottleneck.

What was announced

Anthropic's announcement has three parts.

Compute side. Full access to capacity at SpaceX's Colossus 1 facility. That alone is 300+ MW and 220,000+ NVIDIA GPUs, active within the month. This figure sits next to other ongoing deals: up to 5 GW with Amazon (~1 GW by end of 2026), 5 GW with Google + Broadcom (2027), $30B Azure capacity with Microsoft + NVIDIA, $50B US AI infrastructure investment with Fluidstack. The scale is well past what a single company realistically needs from one tenant — Anthropic is investing less in its own growth curve and more in a future that demands multiple data centers for regulatory and geographic separation.

Product side. Claude Code's five-hour usage limits doubled across all paid plans. For a Pro user, this is the most concrete change: the amount of work you can do in a single session is now 2x. The "peak hours" throttling on Pro and Max was removed — that familiar window between roughly 14:00–22:00 UTC where the API tightens is gone. Claude Opus models got substantially raised API rate limits.

Target side. The post explicitly cites regulated sectors (financial services, healthcare, government) and international data residency (Asia, Europe). That answers the "why" of multi-region compute: letting an enterprise customer choose which country physically hosts their data is the table-stakes requirement for regulated markets.

Bonus: Anthropic said it will cover consumer electricity price increases caused by its US data centers, and is exploring how to extend internationally. That's an open admission that the AI infrastructure boom is now large enough to deserve a macro conversation.

What changed

The big debate of 2024–2025 in frontier AI was "will scaling continue." Mid-2026 the debate has shifted: not whether scaling continues, but who can secure enough power and GPUs to do it. The Microsoft–OpenAI axis enjoyed years of pre-paid data center buildout. Anthropic's last six months of moves are sequential signatures aimed at closing — maybe overtaking — that gap:

• Amazon (long-standing deep partnership, up to 5 GW)

• Google + Broadcom (custom TPU/ASIC designs, 5 GW)

• Microsoft Azure ($30B capacity from the rival's backbone)

• Fluidstack ($50B US infrastructure investment)

• SpaceX Colossus 1 (300+ MW in one move)

The interesting part is reading these five together. Anthropic doesn't want to depend on a single hyperscaler — the Amazon, Google, Microsoft, Fluidstack, SpaceX spread is a deliberate move to multiplex both capacity and geography. Combined with the "tiered frontier model world" I covered in the Mythos piece, the picture becomes: a lab that opens its strongest models only to verified customers is simultaneously building one of the broadest multi-region compute portfolios in the world. Not a coincidence — one requires the other.

The SpaceX side is interesting on its own. SpaceX's quiet entry into the data center market has been one of the past year's underreported stories; allocating all of Colossus 1 to Anthropic positions SpaceX as a serious AI infrastructure supplier. Starlink + on-orbit compute fantasies aside, the real deal is still megawatts on the ground.

What it means for solo developers

If you use Claude Code Pro (I do), the daily impact is immediate.

Old constraint: a fixed prompt/token budget per 5-hour window, narrowing during peak hours — that "you're using too much right now" message in the Turkish evening was a familiar friend. New state: that budget is doubled and peak throttling is gone. Practically, on my side, this doubles the amount of work I can ship in a single session on Cubitz or Konnex. The "I'll have to split this into tomorrow" moment will be much rarer.

The wider read: Anthropic's compute abundance lets Claude Code stay aggressive on price-performance. When the GPT-5.5 + Cursor 3 Composer axis strengthened last month, the "is Anthropic Pro still worth it" question grew louder. Today's announcement is the answer: same money, twice the work.

The invisible side of the limit raise is that Anthropic can hold its pricing strategy for now. The Mythos / "verified defender" logic from a month ago pointed to a tiered frontier — bad news for solo developers there. This announcement points the other way: mainline products (Claude Code, Opus API) are getting more accessible. Reading the two together clarifies Anthropic's two-tier strategy: narrow-access top tier (Mythos, Glasswing) + wide-access mainline (Code, Opus, Sonnet).

Limits and concerns

A few things remain unclear.

Is "all capacity" really all capacity? What's Colossus 1's total design capacity, and what share does Anthropic actually take? Is it "all new capacity" or "all existing"? Anthropic's wording is "all compute capacity," but there's no parallel confirmation from SpaceX. Whether SpaceX retains a slice for its own internal AI work is unknown.

Is the deployment timeline realistic? "Within the month" — i.e., 220K GPUs, 300 MW active by end of May. That's aggressive even by hyperscaler standards. Even grid interconnection alone takes months for most hyperscale data centers. The announcement says "within the month," not "in the coming months" — a tempo that needs verification.

Consumer electricity commitment. Anthropic said it will cover electricity price increases from its US data centers. How that's actually applied (rebate? offset fund?) isn't in the post. Direction is positive; implementation detail is missing.

Geographic question marks. Asia/Europe data residency is mentioned, but no concrete locations. For a developer trying to sell enterprise AI in markets like Turkey, "which region, when" remains open.

Bottom line

Two stories, one move. On the capacity side, Anthropic just made its supply problem nearly invisible for the rest of 2026; on the Claude Code side, it put that into the pocket of solo and team users. Read alongside the Mythos line, Anthropic's 2026 strategy clarifies: selective frontier capability, generous mainline, multiplexed infrastructure.

The practical impact today, for a Pro subscriber, is positive and immediate. The medium-term question is how fast this compute abundance turns into new models. Colossus 1's 220K GPUs aren't only for inference — clearly also for training. The first big signal of what comes after Opus 4.7 will probably be the next few months' model announcement.