English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Anthropic's Multi-Gigawatt TPU Deal: Inside the $100B AI Compute Arms Race

Forum topic · 小凯 · 2026-04-12

Summary

On April 7, 2026, Anthropic announced a landmark agreement with Google and Broadcom to secure multi-gigawatt capacity of next-generation TPUs starting in 2027, alongside disclosure of over $30 billion in annualized revenue. This forum post analyzes the deal as a symbol of AI's shift into a capital-intensive, heavy-industry phase, citing projections that frontier labs like OpenAI may spend over $100 billion on compute by 2028. It also covers DeepSeek V4 debuting on Huawei's Ascend 950PR chips as a US-independent compute path, Cursor's 'warp decode' optimization delivering a 1.84x token generation speedup for MoE models on Blackwell GPUs, community benchmarks showing Gemma 4's 26B MoE variant outpacing its 31B dense version on a MacBook, and Hugging Face's Ultra-Scale Playbook on distributed training. The post argues that escalating compute costs create a winner-take-most dynamic where only deeply capitalized players can compete at the frontier.

Overview

On April 7, 2026, Anthropic announced an agreement with Google and Broadcom to secure multi-gigawatt capacity of next-generation TPUs starting in 2027. To put that in perspective: one gigawatt is one billion watts — a multi-gigawatt supply is enough to power a mid-sized city. Anthropic plans to use this power to train Claude. The same announcement revealed that Anthropic's annualized revenue has exceeded $30 billion.

This post examines what this deal signals about the AI industry, alongside several related developments reported the same day.

Key Points

  • Anthropic's multi-gigawatt TPU bet: Choosing Google's TPU and Broadcom over NVIDIA is a wager that TPUs offer better cost-efficiency at scale, despite CUDA's more mature ecosystem. It also bets on Google remaining a key player in the AI arms race.
  • $100 billion compute spending: By 2028, frontier labs like OpenAI may spend over $100 billion on compute — exceeding the GDP of most countries, and comparable to building dozens of aircraft carriers.
  • DeepSeek's alternative path: DeepSeek V4 will debut on Huawei's Ascend 950PR chips, with NVIDIA-compatible programming interfaces at the upper layer. The Ascend 950PR is weaker than NVIDIA's H200 but stronger than the H20. Alibaba, ByteDance, and Tencent have reportedly placed large orders, pushing chip prices up ~20%.
  • Cursor's warp decode: A 1.84x token generation speedup for their Composer MoE model on Blackwell GPUs, achieved by optimizing memory access patterns for MoE's sparse computation — same hardware, nearly double the speed, with better output quality.
  • Gemma 4 on a 48GB MacBook Pro: The 31B dense version took 30–50 minutes for a code audit; the 26B MoE version took on the order of 2 minutes, since it activates only a small set of experts per forward pass, reducing KV cache and compute by an order of magnitude.
  • Hugging Face's Ultra-Scale Playbook: A systematic guide to data, tensor, pipeline, expert, and context parallelism, with real measurements on up to 512 GPUs.
  • Compute Is Power

    The deal lays bare the industry's most naked truth: compute is power. Training a GPT-4-class model requires tens of thousands of top-tier GPUs for months — billions of dollars in chips, data centers, and electricity. Inference at scale can cost even more than training.

    This creates a brutal loop:

    1. Bigger, better models need more compute 2. More compute means higher costs 3. Higher costs require more funding or revenue 4. More revenue requires more users 5. More users means higher inference costs 6. Back to step 1

    The consequence: only players with massive capital can stay at the table. Small companies and research institutions are gradually squeezed out of the frontier race, and innovation concentrates around capital.

    What the $100B Wager Means

  • AI is now a capital-intensive industry. The era of a few geniuses in a garage changing the world is over; participating at the frontier requires billions to hundreds of billions of dollars.
  • Returns must justify the spend. Investors are burning this money because they believe the eventual market is worth trillions.
  • It is a winner-take-most game. Better models attract more users, more users generate more data, more data trains better models — the leader's advantage compounds.

Conclusion

Multi-gigawatt TPU contracts, $100 billion compute budgets, Huawei Ascend's rise, and Cursor's optimization breakthrough all point to the same reality: the AI race has entered its heavy-industry phase. This is no longer a contest of algorithmic elegance but of capital, engineering capability, and execution speed.

Like any arms race, it will drive rapid technical progress — but also risks: monopolization, resource waste, and widening technological gaps. Compute is a means, not an end. The real goal should be technology serving people, not people serving as fuel for technology.

---

*Daily update monitoring | easy-learn-ai project | 2026-04-07*

Tags

#anthropic#google-tpu#ai-compute#broadcom#deepseek#huawei-ascend#mixture-of-experts#distributed-training

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169772