Anthropic's Multi-Gigawatt TPU Bet with Google and Broadcom: Inside the Compute War

Anthropic's "Compute Castle"

In 2027, Anthropic will own a "compute castle" — not a metaphor. Google and Broadcom have signed a multi-gigawatt TPU supply contract with Anthropic, with deliveries beginning in 2027. It is the latest bombshell in the AI arms race.

What does multi-gigawatt mean?

• 1 GW = 1 billion watts = 1 million kilowatts
• Three Gorges Dam total capacity: 22.5 GW
• A large nuclear plant: roughly 1–2 GW
• New York City peak demand: roughly 5–6 GW

The contract essentially locks in the entire output of a small power station, dedicated to AI training. It is not buying chips — it is buying the "future."

Why now?

Next-generation TPUs deliver in 2027, leaving a two-year gap between signing and delivery — nearly an eternity in AI. Anthropic is betting on two things:

1. Scaling laws keep working. GPT-3 to GPT-4, Claude 2 to Claude 3 — each generational leap came with exponential training-cost growth. Anthropic believes this trend holds at least through 2027. 2. Compute becomes a strategic scarce resource. NVIDIA GPUs are sold out with months-long lead times. Locking in multi-gigawatt capacity early means Anthropic can train models in 2027 that others cannot afford to train.

What does $30B annualized revenue mean?

Anthropic also disclosed annualized revenue above $30 billion — reportedly above OpenAI's ~$8–10B, dwarfing Snowflake (~$3B) and MongoDB (~$2B). But revenue is not profit: GPT-4 reportedly cost over $100M to train, and inference costs add more. That is why Anthropic needs this TPU deal — not for show, but for survival.

China's alternative path

While Anthropic signs TPU deals, Chinese labs are taking a different route. DeepSeek V4 is planned to run natively on Huawei Ascend 950PR chips while remaining compatible with NVIDIA's programming interfaces — effectively "running CUDA code on Huawei silicon," so existing PyTorch/TensorFlow projects migrate directly.

The Ascend 950PR reportedly outperforms the export-compliant H20 and trails the H200, but the key point is: it works. Alibaba, ByteDance, and Tencent have reportedly placed large orders, driving chip prices up about 20%. A closed-loop domestic AI compute stack is forming in China — some memory chips still need imports, but core compute is self-sufficient — a serious challenge to the real-world effectiveness of US export controls.

The $100 billion gamble

Industry reports suggest OpenAI and other top labs may spend over $100 billion on compute by 2028 — equivalent to 10 Ford-class aircraft carriers, NASA's entire decade budget, or all university R&D worldwide for a year. If real, AI training would become the most expensive scientific project in human history, exceeding the Manhattan Project, Apollo, and the Human Genome Project.

• Believers in scaling laws say it's worth it: each scale-up brought qualitative capability jumps.
• Skeptics note diminishing returns: GPT-3→GPT-4 was smaller than GPT-2→GPT-3. If $100B buys only 10% improvement, is it worth it?

Nobody knows. But every major player is betting: Anthropic's TPU deal, OpenAI's reported Stargate project, Google and Amazon expanding data centers. It is a prisoner's dilemma — even if you don't believe in scaling laws, you must follow, or your competitors will.

Cursor's "black magic"

Beyond the grand narrative, a small team is fighting "brute force" with "cleverness." The Cursor team implemented "warp decode" for their Composer MoE model on Blackwell GPUs, claiming ~1.84x faster token generation with better output quality — approaching hardware-upgrade gains through software alone. Relatedly, the Muon optimizer will have a fast path on consumer Blackwell cards by reusing matmul main loops. In a $100B arms race, such optimizations may be how small teams survive.

Open source vs closed source compute economics

When Gemma 4 can run locally on an iPhone and communities share trained models, the "need an API key" business model is under pressure. But the compute war gives closed-source models a defense: astronomical training costs need subscription revenue to pay for them. Meta can open-source Llama because of ad revenue; Google can open-source Gemma because it has money. If training truly costs $100B, is "fully open, fully free" sustainable — or will the future be layered, with open base models and closed, paid frontier models?

Closing thoughts

Anthropic's TPU contract marks compute shifting from "resource" to "strategic asset," and the AI race moving from algorithm competition to capital-plus-infrastructure competition. But tech history is full of seemingly invincible incumbents that were disrupted — IBM by PCs, Nokia by smartphones, Intel by mobile-first and AI-first. Will today's compute castle become tomorrow's burden? Maybe Cursor's warp decode offers an answer: sometimes clever beats strong. Multi-gigawatt power, hundred-billion-dollar bets, open vs closed — all of this is shaping the world we are about to enter, and that world may be crazier than anyone imagines.

---

Source commit: 2c47ab1