Anthropic's Multi-Gigawatt TPU Deal with Google: The AI Compute War Heats Up

Anthropic's "Compute Castle": The Multi-Gigawatt TPU Bet

*Source commit: 2c47ab1*

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

What Does Multi-Gigawatt Mean?

1 GW = 1 billion watts. For comparison:

• The Three Gorges Dam has a total installed capacity of 22.5 GW
• A large nuclear power plant is roughly 1–2 GW
• New York City's peak power consumption is about 5–6 GW

This contract locks in the entire output of a small power station, dedicated to AI training. It's not buying chips — it's buying the future.

Why Now?

The next-generation TPU arrives in 2027, a full two years from signing — nearly an eternity in AI. Anthropic is betting on two things:

1. Scaling laws will keep working. Every generational leap (GPT-3 to GPT-4, Claude 2 to Claude 3) has come with exponentially rising training costs, and Anthropic believes the trend will hold through at least 2027. 2. Compute will become a strategically scarce resource. NVIDIA GPUs are sold out with months-long lead times. Locking in multi-gigawatt capacity two years early means Anthropic can train models in 2027 that others cannot afford to train.

What Does $30 Billion Annualized Revenue Mean?

Alongside the contract, Anthropic revealed annualized revenue exceeding $30 billion — reportedly above OpenAI's $8–10 billion level, and dwarfing Snowflake (~$3B) or MongoDB (~$2B).

But high revenue doesn't mean high profit. GPT-4's training reportedly cost over $100 million; Claude 3.5/4-class models cost even more, plus inference costs. That's why Anthropic needs this TPU deal — not for show, but for survival.

China's Alternative Path

While Anthropic signs TPU deals in the US, Chinese AI labs are taking another route. DeepSeek V4 is planned to run natively on Huawei Ascend 950PR chips while remaining NVIDIA API-compatible — essentially running CUDA code on Huawei silicon, so existing PyTorch/TensorFlow projects can migrate directly.

The Ascend 950PR reportedly outperforms the H20 (export-compliant) but trails the H200. Crucially: 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 — though some memory chips still need imports — posing a serious challenge to the real-world effectiveness of US export controls.

The $100 Billion Gamble

Industry reports suggest that by 2028, top labs like OpenAI may spend over $100 billion on compute. For scale: that's 10 Ford-class aircraft carriers, NASA's entire ten-year budget, or the combined annual R&D spending of all the world's universities.

• Scaling-law believers say it's worth it: every scale-up has delivered qualitative capability jumps.
• Skeptics point to diminishing marginal returns: what if $100 billion buys only a 10% improvement?

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

Cursor's "Black Magic"

Amid the macro narrative, a small team is fighting brute force with cleverness. Cursor implemented "warp decode" for their Composer MoE model on Blackwell GPUs, claiming ~1.84x faster token generation with better output quality — software optimization approaching the effect of a hardware upgrade. Separately, the Muon optimizer will have a fast path on consumer Blackwell cards, since the implementation can reuse the matmul main loop.

Open Source vs Closed Source Compute Economics

The compute war is reshaping the open-vs-closed debate. When Gemma 4 can run locally on an iPhone and Hermes Agents can write their own skills, the "API key" business model is under pressure. But the compute war also gives closed models a defense: if training requires astronomical investment, who pays without subscription revenue? If training really costs hundreds of billions, is "fully open, fully free" sustainable — or will we see a tiered future with open base models and closed frontier models?

Closing Thoughts

Anthropic's TPU contract marks a milestone: compute has shifted from a "resource" to a "strategic asset," and the AI race has entered a new phase of capital and infrastructure competition. But tech history is full of seemingly invincible incumbents that were eventually disrupted — IBM by PCs, Nokia by smartphones, Intel by mobile- and AI-first shifts. Could today's compute castle become tomorrow's burden? Perhaps Cursor's "warp decode" hints at an answer: sometimes, clever beats strong.

Multi-gigawatt power, hundred-billion-dollar investments, the open-vs-closed tug of war — all of it is shaping the world we're about to enter. And that world may be crazier than anyone imagines.

---

*Original post tags: easy-learn-ai, daily update, memory, XiaoKai, compute war, Anthropic, TPU, AI infrastructure*