How Anthropic Built a $200B+ Compute Empire in Under 12 Months: A Timeline
From Amazon to Google to SpaceX in under a year. Here's every Anthropic compute deal, what it costs, and when it comes online.
Anthropic Signed $200B+ in Compute Deals in Under a Year — Here’s the Full Timeline
Twelve months ago, Anthropic was rationing tokens. Today, the company controls access to one of the largest GPU fleets on the planet. If you’ve been watching Claude’s rate limits swing wildly — tighter in March, looser overnight in May — this is the story behind those swings.
The Anthropic compute portfolio now spans four major deals: Amazon (up to 5 GW), Google/Broadcom ($200B over five years, 5 GW beginning 2027), Microsoft/Nvidia ($30B in Azure capacity), and SpaceX Colossus 1 (220,000 Nvidia GPUs, 300 MW, live now). Understanding how those deals came together — and in what order — tells you a lot about where Claude’s capacity is actually coming from when you hit “send.”
This post is a timeline. It’s also an explanation of why Anthropic ended up in this position in the first place.
Why Anthropic Ran Out of Compute (and Why That’s Surprising)
Dario Amodei made a deliberate bet a few years ago: don’t over-acquire GPUs. The reasoning was sound. If AI demand didn’t accelerate at a perfect rate, a massive capex commitment could sink the company. OpenAI took the opposite approach — raise everything, buy everything, leverage the company to the hilt. At the time, Anthropic’s conservatism looked prudent.
Then demand didn’t just accelerate. It exploded.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
At the Code with Claude developer event in May 2026, Dario said this out loud: “We planned for a world of 10x growth per year. In the first quarter of this year, we saw 80x annualized growth in revenue and usage.” That’s not a rounding error. That’s a planning assumption that was off by 8x in a single quarter.
The ARR numbers make this concrete. In January 2026, Anthropic was at $14B annualized. Within months, that trajectory ran $14B → $19B → $24B → $30B → $44B. Each step happened faster than the last. The demand was real. The compute wasn’t there to serve it.
If you were using Claude Code in early 2026 and hitting your 5-hour limit before lunch, this is why. Anthropic wasn’t being arbitrary. They genuinely didn’t have the GPUs. You can read more about how that shortage played out for users in Anthropic’s compute shortage and why Claude limits were tightening.
The Deals, in Order
Deal 1: Amazon AWS — Up to 5 GW (Announced Early 2026)
The Amazon deal came first and is the largest by raw capacity: up to 5 gigawatts of new compute, with nearly 1 GW expected online by end of 2026. For context, a single gigawatt can power roughly 750,000 homes. This is not a small arrangement.
The catch is timeline. Most of this capacity won’t be available until late 2026 at the earliest, and the full 5 GW buildout extends well beyond that. Amazon’s infrastructure here runs on AWS Trainium chips — Anthropic’s own blog confirms they train and run Claude across AWS Trainium, Google TPUs, and Nvidia GPUs. The multi-hardware flexibility is intentional. It means Anthropic isn’t locked to any single silicon vendor.
The Amazon deal is the long game. It doesn’t fix a rate limit today. It’s the foundation for 2027 and beyond.
Deal 2: Google/Broadcom — $200B, 5 GW (Beginning 2027)
This is the one that briefly moved markets. The reported figure is $200B over five years — a 5 GW agreement with Google and Broadcom that begins coming online in 2027. The scale of this deal is large enough that it reportedly accounts for over 40% of Google’s $462B reported backlog, which is part of why Google briefly surpassed Nvidia as the world’s most valuable company in overnight trading when the deal became public.
Google Cloud CEO Thomas Curran had previously claimed Google had enough TPU capacity to serve the entire market — including external inference customers like Anthropic. Sundar Pichai said the opposite: more compute would mean more revenue. The Anthropic deal suggests Pichai was right.
The Google/Broadcom deal is also the most strategically interesting. Anthropic is one of Google’s primary competitors in the foundation model space. Google is also one of Anthropic’s largest investors. The fact that Anthropic is buying $200B in compute from a company that both funds it and competes with it is the kind of arrangement that would be strange in any other industry.
Deal 3: Microsoft/Nvidia — $30B Azure Capacity (Strategic Partnership)
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Less has been reported on the mechanics of this one, but the number is significant: $30B in Azure capacity as part of a strategic partnership with Microsoft and Nvidia. This gives Anthropic access to Nvidia GPU infrastructure through Azure’s cloud, which matters because Azure has been aggressively building out its own AI infrastructure following the OpenAI partnership.
The Microsoft deal is notable for what it signals: Anthropic is now a customer of every major hyperscaler simultaneously. AWS, Google Cloud, and Azure are all in the portfolio. That’s unusual. Most large AI companies anchor to one cloud provider. Anthropic is deliberately avoiding that dependency.
Deal 4: Fluid Stack — $50B in American AI Infrastructure
Anthropic also announced a $50B investment in American AI infrastructure through Fluid Stack. This is less a compute-access deal and more a capacity-building play — funding the physical infrastructure (data centers, power, cooling) that future compute will run on. The timeline here is the longest of any deal in the portfolio.
Deal 5: SpaceX Colossus 1 — 220,000 GPUs, 300 MW (Live Now)
This is the one that actually changed your rate limits overnight.
Anthropic’s blog post — titled “Higher usage limits for Claude and a compute deal with SpaceX” — announced that Anthropic now has full use of XAI’s Colossus 1 data center in Memphis, Tennessee. That’s 220,000 Nvidia GPUs (mostly H100s), operating at 300 MW capacity. The inference was available within the month of announcement, and the rate limit changes were effective immediately.
Why was Colossus 1 available? XAI had already moved its own model training to Colossus 2, their newer Blackwell-based cluster containing around 550,000 GPUs. Elon Musk’s tweet explained it plainly: “I was okay leasing Colossus 1 to Anthropic, as SpaceX had already moved training to Colossus 2.” The less flattering version, which multiple sources confirmed: XAI is rebuilding its model from scratch because “it wasn’t built right the first time,” leaving Colossus 1 sitting idle and burning money.
Idle GPUs are expensive. The total cost of ownership doesn’t pause when utilization drops to zero. Leasing to Anthropic was the rational move.
What Changed for Builders the Day the SpaceX Deal Closed
The Colossus 1 deal is the only one in the portfolio that produced immediate, measurable changes. Here’s what actually shifted:
Claude Code rate limits doubled. The 5-hour rate limit for Pro, Max, Team, and seat-based Enterprise plans was doubled, effective the day of announcement. Anthropic’s head of growth, Amal Avisari, explained the prioritization: “Only a very small percentage hit weekly limits, while a much larger portion of users hit the 5-hour limit. So we fixed that first.”
Peak-hours throttling removed. For Pro and Max accounts, the reduction in Claude Code limits during peak hours was eliminated entirely. This was a significant quality-of-life change — the throttling had made Claude Code unreliable during the hours most developers actually work.
API rate limits increased substantially. The numbers here are large. For Tier 1 accounts, input tokens per minute went from 30,000 to 500,000 — a 16x increase. Output tokens per minute went from 8,000 to 80,000 — a 10x increase. Tier 2 input went from 450,000 to 2,000,000. Tier 3 from 800,000 to 5,000,000. Tier 4 from 2,000,000 to 10,000,000.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
If you were building production agents with Opus and hitting rate limits constantly, the output token change is the one that matters most. 8,000 output tokens per minute is a hard ceiling for parallel agent workflows. 80,000 is a different category of capability.
For builders thinking about what this means architecturally: multi-agent workflows that were previously impractical on Opus are now viable. Five sub-agents each reading 50,000 tokens of context, running in parallel, no longer immediately exhausts your per-minute budget. Platforms like MindStudio — which supports 200+ models and provides a visual builder for chaining agents and workflows — are the kind of environment where these new limits actually change what you can build, since orchestrating parallel Opus calls was previously a rate-limit management problem more than a design problem.
The Elon Factor (and Why It’s Weirder Than It Looks)
Elon Musk spent most of 2025 and early 2026 publicly attacking Anthropic. “Is there a more hypocritical company than Anthropic?” was one tweet. “Anthropic will be misanthropic” was another. He called for Anthropic employees to quit. He had Grok write a “vulgar roast” of Dario Amodei.
Then he spent a week with senior Anthropic team members and tweeted: “No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good.”
The cynical read — and it’s probably mostly correct — is that this is a business decision dressed up in diplomatic language. XAI had idle compute. Anthropic needed compute. The enemy of my enemy is my friend, and Elon’s real enemy is Sam Altman, not Dario Amodei. Chimath Palihapitiya had called this shot on the All-In podcast weeks earlier: “He and Dario should do a deal tomorrow.”
The more interesting read is what it suggests about XAI’s trajectory. The company has no meaningful agentic product to compete with Claude Code or Codex. Grok 4.2 landed in February with minimal impact. Co-founders left one after another. The Cursor deal — SpaceX has an option to acquire Cursor for $60B in late 2026 or pay a $10B breakup fee — was supposed to be the reclamation project, but reporting suggested Cursor was keeping its distance from XAI’s model development. Leasing Colossus 1 to Anthropic while rebuilding from scratch on Colossus 2 is the move of a company that knows its model isn’t competitive yet and needs to monetize its infrastructure in the meantime.
What the Portfolio Looks Like Now (and What’s Still Missing)
Here’s the current state of Anthropic’s compute portfolio, organized by when capacity actually comes online:
Live now: SpaceX Colossus 1 — 220,000 H100s, 300 MW. This is the only deal producing immediate capacity.
By end of 2026: Amazon AWS — nearly 1 GW of the 5 GW agreement. This is the next meaningful capacity addition.
Beginning 2027: Google/Broadcom — 5 GW, $200B over five years. The largest deal by dollar value, but the furthest out.
Ongoing: Microsoft/Nvidia ($30B Azure), Fluid Stack ($50B infrastructure investment). Both contribute to the longer-term picture.
How Remy works. You talk. Remy ships.
The gap between “live now” and “end of 2026” is real. Colossus 1 is 300 MW. The Amazon tranche coming online by year-end is roughly 1,000 MW. That’s a meaningful step up, but it’s still months away. If Anthropic’s demand continues growing at anything close to the Q1 2026 rate, the compute crunch isn’t fully solved — it’s just less acute.
This is also why the rate limit changes announced with the SpaceX deal were targeted specifically at the 5-hour Claude Code limit rather than the weekly limit. Amal Avisari said it directly: they fixed the 5-hour limit first because that’s where most users were hitting walls. The weekly limit is still there. As more Amazon capacity comes online, expect that to be the next thing that moves.
What This Means If You’re Building on Claude
The practical implications depend on what you’re building.
If you’re using Claude Code for development work, the doubled 5-hour limit and removed peak-hours throttling are the changes that matter. You can read about how to run Claude Code more cost-effectively using Open Router if you’re also watching your API spend.
If you’re building production agents on the API, the 10x output token increase for Tier 1 is the number to pay attention to. The Claude Code source code leak surfaced some architectural details about how Claude Code manages context internally — worth reading if you’re designing agents that need to stay within limits efficiently.
If you’re building full-stack applications that call Claude on the backend, the higher throughput means you can design for more concurrent users without hitting rate limits as a primary constraint. Tools like Remy approach this from a different angle — you write an annotated spec in markdown, and it compiles into a complete TypeScript backend, SQLite database, auth, and deployment. The spec is the source of truth; the generated code is derived output. For teams that want to move from Claude-powered prototype to deployed application without managing the full stack manually, that abstraction matters.
For anyone building AI agents or multi-agent workflows, the new API limits make Opus viable as a backbone model in ways it wasn’t before. Five parallel sub-agents on Opus was a rate-limit problem in April. In May, it’s a design problem — which is the right kind of problem to have.
The Orbital Compute Footnote
One line in the SpaceX announcement is easy to skip past: “As part of this agreement, we have also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.”
GPUs in space. Anthropic’s stated reasoning is that terrestrial compute has a real long-term ceiling — power availability, water for cooling, community opposition to data center construction. Orbital compute sidesteps most of those constraints.
Jensen Huang has called it interesting. Sam Altman has called it impractical. Elon Musk is obviously bullish. Whether it happens or not, the fact that Anthropic included it in a compute deal announcement signals something about how they’re thinking about the 5-to-10-year horizon. The terrestrial deals — Amazon, Google, Microsoft, SpaceX — are the next three years. Orbital is the bet beyond that.
Not a coding agent. A product manager.
Remy doesn't type the next file. Remy runs the project — manages the agents, coordinates the layers, ships the app.
The compute race in AI has always been about who can build infrastructure faster than demand grows. For most of 2025 and early 2026, Anthropic was losing that race. The SpaceX deal bought them time. The Amazon and Google deals are the actual answer — if they come online on schedule, and if demand doesn’t outpace them again.
Given that Q1 2026 came in at 80x against a 10x plan, “if demand doesn’t outpace them again” is doing a lot of work in that sentence.
