// Article · May 9, 2026
Anthropic's $1 trillion week
How one company bought its way out of a compute crisis — and committed $200B+ in deals to do it
Six weeks ago, Claude Code was the punchline of every AI engineering Slack. Performance was visibly degraded. The Pro tier briefly lost Claude Code entirely before being restored 24 hours later. Fortune ran a piece where Anthropic admitted to "engineering missteps." Reddit threads piled up with cancelled subscriptions. The diagnosis everyone landed on, including Anthropic eventually: the company had run out of compute.
This week, Anthropic crossed $1 trillion in valuation, reportedly raised at $900 billion, signed Elon Musk's data centre, doubled rate limits, and announced it has committed roughly $200 billion to Google over the next five years. None of those things happened in a vacuum. They are all the same story.
The Core Claim
Anthropic's compute crisis was the visible side-effect of a less-visible fact: revenue grew 80x in a single quarter, per Dario Amodei. The company's engineering scaffolding broke under that load. The fix is not better engineering — it's signing every cloud, chip, and data centre deal it can find, all at once, regardless of whose data centre it is or whose Pentagon ally calls them a "supply-chain risk."
What Anthropic Has Now Bought
| Partner | Deal | Status |
|---|---|---|
| Google Cloud | ~$200B over 5 years; Anthropic now ~40% of GCP's revenue backlog | Reported May 6 |
| Amazon (AWS) | 5 GW capacity | Earlier 2026 |
| Google + Broadcom | 5 GW for custom silicon | Earlier 2026 |
| Microsoft + NVIDIA | $30B partnership | Earlier 2026 |
| Fluidstack | $50B investment in US infrastructure | Earlier 2026 |
| SpaceX (Colossus 1, Memphis) | 300+ MW, 220K+ NVIDIA GPUs, all of one data centre | Announced May 6 |
| SpaceX (orbital) | "Expressed interest" in multi-gigawatt space datacenters | Conceptual |
For context: 300 MW is roughly the power consumption of a small city of 250,000 people. Anthropic just signed up for that, plus 5 + 5 GW elsewhere — and that's only the deals announced. The $200B Google commitment alone is more than the entire 2024 GDP of New Zealand.
Why Memphis Specifically Is the Beat
The Colossus 1 site started life as xAI's flagship Memphis data centre — Elon Musk's personal AI project. SpaceX, which Musk also runs, ended up holding the asset. This week's deal hands its full capacity to Anthropic, the company a Pentagon faction allied with Musk had recently designated a "supply-chain risk." The pricing at this scale has a single buyer who'll show up for 220K+ GPUs of capacity within a month, and that buyer is Anthropic. So Musk's company is now powering Musk's competitor. Politics is downstream of gigawatts.
The other half of the story: Code with Claude SF was happening live during the announcement. About an hour into the keynote, Claude itself briefly went down for thousands of users (DownDetector spike at 11:16 AM ET). Truly nothing tells the audience "we're spending billions to fix our compute problem" like the compute problem briefly breaking the app mid-keynote.
What This Means for Anyone Using Claude
- Claude Code 5-hour rate limits doubled for Pro, Max, Team, and Enterprise.
- Peak-hour throttle removed for Pro and Max.
- Opus API rate limits "considerably" increased.
- Three new agent capabilities — multi-agent orchestration, outcomes-based completion, overnight "dreaming."
- Claude for Microsoft 365 (Excel/Word/PowerPoint add-ins).
- Claude for Creative Work (Adobe / Blender / Autodesk / Ableton).
- Claude on Amazon Bedrock self-serve in 27 regions.
- 10 finance agents (pitch builder, KYC screener, month-end closer).
The Honest Take
The bull case: Anthropic is the fastest-growing company in American business history. $1B → $30B annualised revenue in 16 months (Axios). Claude Code alone is a $2.5B run-rate product line with 300,000+ business customers. Every Claude Code-replacing-developer-headcount story is a brick in this trillion-dollar wall.
The bear case: $700B in 2026 capex is a bet that AI cash flow will catch up to AI infrastructure spend. Big Tech's Q1 earnings showed only Google convinced markets the bill is being paid. Meta's stock fell 6% after announcing a $10B increase in capex. The market's bar is now Google Cloud's 63% growth — anything less and investors wait for proof. If AI revenue plateaus before infrastructure debt is repaid, this is a $700B mistake.
The Anthropic-specific bet: the $50B raise looks less like vanity pricing if Jupiter (the new model family in red-teaming) launches into the existing Claude Code distribution. The $200B Google commitment looks less crazy if it's just compute prepayment for a model class that 300K business customers are already lined up to use. But Anthropic's IPO advisors are reportedly targeting a $400-500B price — half what the secondary market is paying right now. That gap is where the truth gets priced.
What This Means
If you cancelled Claude Code in April, the rate-limit reset means it's worth revisiting this week. Multi-agent orchestration in particular is the first commercial implementation of "spawn fleets of agents that hand off subtasks" — a year ago you needed CrewAI or LangGraph and a weekend.
For your strategic thinking: every time you read "AI is overhyped," check whether the company you're reading about has signed compute deals in the trillions. The infrastructure cycle has now committed enough capital that even a modest revenue downside requires the cycle to keep building. Anthropic specifically is too big to fail in the most literal sense.
Alternatives Worth Naming
- OpenAI — hit 10 GW secured compute this week, years ahead of its 2029 target. Same story, different lab.
- xAI / Grok 4.3 — 1M context, agentic API, but the company that owned Colossus 1 just leased it to Anthropic.
- DeepSeek V4 — 98% KV-cache reduction; the open-weight competitive pressure that nobody is talking about loud enough.
- Subquadratic / SubQ — the architectural alternative to "buy 5 GW." See SubQ and the End of the Transformer's Memory Tax.