The Multi-Gigawatt Gamble: When the AI Race Becomes a Compute Arms Race
2027. Multi-gigawatts.
The number sounds like an energy unit from science fiction. But it is real: Anthropic has announced that starting in 2027, it will receive multi-gigawatt-scale next-generation TPU capacity from Google and Broadcom.
What does multi-gigawatt mean? 1 gigawatt = 1 billion watts. Multi-gigawatt means billions of watts of computing power—enough electricity to power a mid-sized city.
This isn't buying computers. This is building the energy infrastructure of an industrial empire.
Why Has Compute Become the Moat?
Start with a simple question: what does it take to train a top-tier large language model?
The answer may surprise you: beyond clever algorithms and quality data, you need massive compute resources.
Take GPT-4-class models. Training one requires thousands of top GPUs running continuously for months, with electricity costs alone potentially reaching tens of millions or even hundreds of millions of dollars. That excludes researcher salaries, data labeling, and the cost of failed experiments.
This is why when OpenAI released ChatGPT in late 2022, the world was stunned—not because nobody else could build similar models, but because few companies could afford to train them.
Compute has become the hardest currency of the AI era.
Anthropic's Big Bet
Anthropic is OpenAI's primary competitor. Their Claude models perform well in coding, long-context handling, and reasoning quality.
But compared with OpenAI, Anthropic has an obvious weakness: compute.
OpenAI has Microsoft's backing and priority access to high-end GPUs on Azure. Though Anthropic also has Google's investment, it has been playing catch-up on compute access.
This deal with Google and Broadcom is Anthropic's big gamble.
What does multi-gigawatt TPU capacity mean? It means Anthropic can train multiple large models simultaneously, run more aggressive experiments, and have more room in model scale and training duration. Most importantly, it finally gives Anthropic a card to play head-to-head against OpenAI.
But the cost is enormous. Industry reports suggest that by 2028, top AI labs like OpenAI may spend over $100 billion on compute. This isn't one-time investment—it's continuously growing annual operating cost.
TPU: Google's Secret Weapon
TPU (Tensor Processing Unit) is Google's chip designed specifically for machine learning. Unlike general-purpose GPUs, TPUs are optimized at the architecture level for matrix operations and neural network training. On certain AI workloads, TPU efficiency can far exceed that of GPUs at equivalent power consumption.
Google has long positioned TPUs as a key differentiator for its cloud services: training on TPUs is cheaper and faster than on other providers' GPUs.
Anthropic's choice to bind deeply with Google—rather than buying GPUs on the open market at the highest bidder—is a pragmatic one. TPU cost-performance advantages plus long-term contract price locks give Anthropic more cost certainty.
But the choice carries risk. Binding core infrastructure to a single vendor means Anthropic's fate is tied to Google's TPU roadmap. If Google falls behind on next-generation TPUs, or if strategic interests diverge, this dependency becomes a double-edged sword.
DeepSeek and Ascend: Another Path
While Anthropic signed its massive deal in the US, another compute contest is unfolding on the other side of the planet.
China's DeepSeek is planning its next-generation model, DeepSeek V4—and this time they chose not NVIDIA GPUs, but Huawei's Ascend 950PR chips.
This is an unusual choice. NVIDIA GPUs have long been the AI training standard: the CUDA programming model, mature software ecosystem, established supply chains. But under US export controls, Chinese companies find it increasingly difficult to acquire high-end NVIDIA chips.
Huawei's Ascend series is the flagship of China's domestic AI chips. The 950PR reportedly outperforms the H20 but falls short of the H200—in other words, not yet the top-tier chip, but well within usable range.
Native support for DeepSeek V4 means deep optimization on Huawei's software stack. The Ascend 950PR is reportedly upper-layer compatible with NVIDIA programming interfaces to reduce migration costs, but extracting peak performance still requires chip-specific optimization.
According to reports, Alibaba, ByteDance, and Tencent have placed large Ascend orders, driving prices up about 20%. This suggests China's domestic AI compute stack can now form a closed loop—at least on the supply side, no longer fully dependent on US chips.
What Is the Endgame of the Compute Race?
Zoom out: today's AI race increasingly resembles an arms race. Not a contest of technical elegance, but of who can spend more money to buy more compute. Anthropic's multi-gigawatt contract, OpenAI's hundred-billion-dollar budget, Chinese vendors' massive chip purchases—none of these are sustainable competition.
Possible endgames:
1. Winner takes all. A few companies with enough compute train models far beyond competitors' reach, forming de facto monopolies. Everyone else is either eliminated or relegated to application-layer innovation on top of the giants' APIs.
2. Efficiency revolution. Algorithmic and hardware efficiency gains make training top models no longer require such staggering compute. New architectures (like MoE), new training methods, or new chip designs could dramatically lower compute costs.
3. Open-source catch-up. The open-source community achieves comparable capability with far less compute through distributed training, model compression, and engineering optimizations. Meta's Llama series has already partly proven this path is viable.
4. Regulatory intervention. When the compute race triggers energy crises, environmental damage, or geopolitical conflict, governments may impose training quotas or carbon limits—fundamentally changing the industry's competitive logic.
Balancing Efficiency and Scale
Another trend worth watching: efficiency optimization.
When compute becomes expensive, smart companies ask: can we achieve the same results with less compute?
- The Cursor team implemented a "warp decode" technique on Blackwell GPUs for their Composer MoE model, claiming roughly a 1.84x speedup in token generation. Same hardware, more users served, faster responses.
- The Muon optimizer was found to have a fast path on Blackwell GPUs, since it can reuse the matrix multiplication main loop. These low-level optimizations look dull, but their cumulative effect is enormous.
- Hugging Face's Ultra-Scale Playbook systematically catalogs data parallelism, tensor parallelism, pipeline parallelism, expert parallelism, and context parallelism, with measured data on up to 512 GPUs. For teams upgrading from "single-machine experiments" to large training clusters, it's an invaluable reference.
Where Geopolitics Meets Business
The compute race is never purely technical. It is deeply embedded in geopolitics and business strategy.
US export controls on AI chips were meant to slow China's AI development. In reality, they stimulated investment in China's domestic chip industry. Huawei's Ascend 950PR is a product of that pressure.
Anthropic's deep binding with Google isn't just a business decision either. In an increasingly geopolitics-driven market, choosing an ally means choosing a camp.
OpenAI's internal governance turmoil—board changes, alignment resource disputes, differences between Sam Altman and the CFO over compute spending and IPO timing—reflects the enormous pressure AI companies face. When operating costs soar into the hundred-billion-dollar range, every strategic decision can determine survival.
What It Means for Ordinary Users
Short term: AI service cost structures may become more opaque. To recoup massive compute investments, companies may raise API prices, launch pricier subscription plans, or tighten usage limits.
Medium term: the compute race may accelerate AI capability growth. More compute means bigger models, more experiments, more possibilities—likely good for users.
Long term: the unsustainability of the compute race may push the industry toward new paths—more efficient algorithms, new computing paradigms (quantum or neuromorphic computing), or a redefinition of AI's capability boundaries.
Conclusion
Multi-gigawatt compute contracts sound like a distant numbers game. Behind them, the AI industry is undergoing profound transformation.
When training one model costs more than many countries' GDP, AI is no longer just a technology—it's a form of power. Whoever controls compute controls the speed and direction of AI development.
Anthropic's deal is a major move in this power game, marking a new phase: no longer just an algorithm contest, but an all-out resource competition.
But history tells us no arms race lasts forever. When costs become unbearable, change arrives—through technological breakthroughs, policy intervention, or open-source disruption.
Until then, we will keep witnessing astronomical contracts, hundred-billion-dollar budgets, and compute as one of this era's scarcest resources.
The multi-gigawatt gamble has begun. Who wins in the end? Time will tell.
---
*Word count: ~3,600 (original Chinese)*