Multi-Gigawatt Wager: Anthropic's TPU Deal and the Trillion-Dollar AI Compute Arms Race

On April 7, 2026, Anthropic announced that it has signed agreements with Google and Broadcom to secure multi-gigawatt capacity of next-generation TPU chips starting in 2027. A gigawatt is one billion watts — multiple gigawatts could power a mid-sized city. Anthropic plans to use this power to train Claude. The company also disclosed that its annualized revenue has exceeded $30 billion.

Compute Is Power

The underlying truth of this news is that compute equals power in the AI industry. By 2028, frontier labs such as OpenAI may spend more than $100 billion on compute — more than the GDP of most countries, enough to build dozens of aircraft carriers. Why? Because whoever reaches superintelligence first gets to define the next era.

TPU: Google's Secret Weapon

TPUs (Tensor Processing Units) are chips Google designed specifically for machine learning, in development since 2016 and now in their fourth and fifth generations. Compared to general-purpose GPUs, TPUs offer higher efficiency and lower energy consumption for specific workloads.

Anthropic choosing Google and Broadcom over NVIDIA is an interesting bet. NVIDIA GPUs dominate AI training but are supply-constrained and expensive. Google's TPU ecosystem is less mature than CUDA, but with sufficient engineering capability, TPUs offer excellent cost-performance. Anthropic is betting that TPUs are the future — and that Google remains a key player in the AI arms race.

DeepSeek's Alternative Path

The same day's updates included another compute story from China: DeepSeek V4 will run first on Huawei's Ascend 950PR chips, with an upper-layer compatibility with NVIDIA's programming interface.

The Ascend chips were developed under US export controls. They are weaker than NVIDIA's H200 but better than the H20, and represent a path not dependent on US technology. Alibaba, ByteDance, and Tencent have reportedly placed large orders for Ascend 950PR, pushing chip prices up about 20%. The lesson: when supply chains can be cut off, owning your own chip capability equals strategic autonomy.

The Brutal Logic of the Compute Arms Race

Training a GPT-4-class model requires tens of thousands of top GPUs running for months — billions of dollars in power, chips, and data centers. Inference costs can exceed training costs at high call volumes.

This creates a brutal cycle:

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

The result: only players with massive capital stay at the table. Small companies and research institutions are being squeezed out of the frontier.

Cursor's Optimization Breakthrough

The same day, Cursor announced a technical breakthrough: "warp decode" for their Composer MoE model on Blackwell GPUs, improving token generation speed by roughly 1.84x.

MoE (Mixture of Experts) models activate only a subset of parameters per forward pass (e.g., 10B of a 100B-parameter model). Traditional decoding doesn't fully exploit this sparsity. Cursor's warp decode optimizes memory access patterns so GPUs handle MoE's sparse computation more efficiently — nearly double the speed on the same hardware, with better output quality. When others spend $1 billion to do something, doing it for $500 million is a competitive advantage.

Gemma 4 on a Mac: The Art of Efficiency

A notable comparison on a 48GB MacBook Pro:

• 31B dense version: a code audit takes 30–50 minutes
• 26B MoE version: takes only on the order of 2 minutes

Why? The dense version processes all 31 billion parameters per pass with a huge KV cache, while the 26B MoE activates only a few experts — an order of magnitude less actual computation. Sometimes smarter architecture matters more than raw scale, especially when compute is limited.

Scaling Challenges: Hugging Face's Ultra-Scale Playbook

Hugging Face released its Ultra-Scale Playbook, systematically covering data parallelism, tensor parallelism, pipeline parallelism, expert parallelism, and context parallelism. Simply adding more GPUs doesn't work — communication overhead, load balancing, and fault recovery are all complex problems. The guide includes measured results on up to 512 GPUs, a valuable reference for teams moving from single-machine experiments to large-scale training.

What a $100 Billion Bet Means

First, AI is now a capital-intensive industry — no longer an era where a few geniuses in a garage can change the world. Second, returns must be enormous to justify such investment — investors believe the eventual market is worth trillions. Finally, this is a winner-takes-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 forecasts, the rise of Huawei Ascend, Cursor's optimization breakthroughs — all point to the same reality: the AI competition has entered its heavy-industry phase. It's no longer a contest of algorithmic elegance, but of capital, engineering capability, and execution speed. Like any arms race, it will drive rapid technological progress but bring risks: monopolies, resource waste, and technology gaps.

As observers, perhaps the clearest takeaway is: compute is a means, not an end. The real goal should be technology serving people — not people becoming fuel for technology.

---

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