The Multi-Gigawatt Gamble: When the AI Race Becomes a Compute Arms Race
2027. Multi-gigawatts.
That number sounds like an energy unit from science fiction, but it is real: Anthropic has just 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 GW = 1 billion watts. Multi-gigawatt means billions of watts of computing power—enough electricity to power a mid-sized city.
This is not buying computers. This is building the energy infrastructure of an industrial empire.
Why Compute Became 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.
Training a GPT-4-class model 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—before researcher salaries, data labeling, and the cost of failed experiments.
That's why when OpenAI released ChatGPT in late 2022, the world was stunned—not because others couldn't build similar models, but because few companies could afford to train them.
Compute became the hardest currency of the AI era.
Anthropic's Big Bet
Anthropic, maker of the Claude series (strong in coding, long-context handling, and reasoning quality), is OpenAI's most prominent rival. But compared to OpenAI—which enjoys priority access to high-end GPUs on Azure thanks to Microsoft—Anthropic has consistently lagged in compute access.
This deal with Google and Broadcom is Anthropic's gamble. Multi-gigawatt TPU capacity means training multiple large models simultaneously, running more aggressive experiments, and having more room in model scale and training duration. Most importantly, it gives Anthropic a card to play head-to-head against OpenAI.
The cost is enormous. Industry reports suggest top AI labs like OpenAI may spend over $100 billion on compute by 2028—not a one-time investment, but a growing annual operating cost.
TPU: Google's Secret Weapon
TPU (Tensor Processing Unit) is Google's custom chip designed for machine learning. Unlike general-purpose GPUs, TPUs are optimized at the architecture level for matrix operations and neural network training, and can significantly outperform same-wattage GPUs on specific AI workloads.
For Anthropic, deep integration with Google—rather than buying GPUs on the open market—is a pragmatic choice: TPU price-performance plus long-term contract price locks provide cost certainty.
But there are risks. Tying core infrastructure to a single supplier means Anthropic's fate is bound to Google's TPU roadmap. If Google's next-gen TPU development stumbles, or if strategic interests diverge, this dependency becomes a double-edged sword.
DeepSeek and Ascend: Another Path
While Anthropic signed its mega-deal in the US, another compute contest is playing out on the other side of the planet.
China's DeepSeek is planning its next-generation model, DeepSeek V4—and this time it chose not NVIDIA GPUs, but Huawei's Ascend 950PR chip.
This is an unusual choice. NVIDIA's CUDA ecosystem has long been the standard for AI training. But under US export controls, Chinese companies face increasing difficulty acquiring high-end NVIDIA chips. The Ascend 950PR reportedly outperforms the H20 but falls short of the H200—not the top chip, but within usable range. The 950PR offers upper-layer compatibility with NVIDIA programming interfaces to ease migration, though peak performance requires optimization for Huawei's stack.
Alibaba, ByteDance, and Tencent have reportedly placed large Ascend orders, driving prices up ~20%—evidence that China's domestic AI compute stack can now form a closed loop, at least on the supply side.
How Does the Compute Race End?
The AI race increasingly resembles an arms race: not about whose technology is more elegant, but who can spend more money on more compute. Possible endings:
- Winner takes all: A few compute-rich companies train far superior models and form de facto monopolies; everyone else does application-layer innovation on their APIs.
- Efficiency revolution: New architectures (e.g., MoE), new training methods, or new chip designs drastically cut the compute needed for frontier models.
- Open-source catch-up: Distributed training, model compression, and engineering optimization let the open-source community reach comparable capability at far lower cost—Meta's Llama series has partly proven this path.
- Regulatory intervention: If the race triggers energy crises, environmental damage, or geopolitical conflict, governments may impose training quotas or carbon limits, fundamentally changing the competitive logic.
- Cursor's team implemented "warp decode" for their Composer MoE model on Blackwell GPUs, claiming ~1.84x faster token generation—same hardware, more users or faster responses.
- The Muon optimizer was found to have a fast path on Blackwell GPUs by reusing the matrix-multiplication main loop.
- Hugging Face's Ultra-Scale Playbook systematically documents data, tensor, pipeline, expert, and context parallelism strategies, with measured results on up to 512 GPUs—a valuable reference for teams scaling from single machines to training clusters.
- Short term: Cost structures may become more opaque as companies recoup massive investments through higher API prices, pricier subscriptions, or tighter usage limits.
- Medium term: The race could accelerate AI capability growth—bigger models, more experiments, better features.
- Long term: Unsustainability may push the industry toward new paths: more efficient algorithms, new computing paradigms (quantum, neuromorphic), or redefined AI capability boundaries.
Efficiency vs. Scale
When compute becomes expensive, smart companies ask: can we achieve the same results with less?
The compute race isn't just "who buys more"—it's also "who uses it better." Efficiency can't fully offset scale, but it makes limited resources go further.
Geopolitics Meets Business
US export controls on AI chips, intended to slow Chinese AI, have instead stimulated investment in domestic Chinese chips—Huawei's Ascend 950PR is a product of that pressure. Anthropic's binding to Google isn't just a business decision; in a market shaped by geopolitics, choosing an ally means choosing a camp.
OpenAI's internal governance turmoil—board changes, disputes over alignment resources, disagreements between Sam Altman and the CFO over compute spending and IPO timing—reflects the immense pressure AI companies face. When operating costs hit the $100-billion scale, every strategic decision can be existential.
What It Means for Ordinary Users
Conclusion
Multi-gigawatt compute contracts sound like a distant numbers game, but they mark a profound transformation. When training a 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, signaling that the AI race has entered a new phase: no longer just an algorithm contest, but a full-scale contest of resources.
History tells us no arms race lasts forever. When costs become unbearable, change comes—through technical breakthroughs, policy intervention, or open-source disruption. Until then, we'll keep watching astronomical contracts, hundred-billion budgets, and compute as one of this era's scarcest resources.
The multi-gigawatt gamble has begun. The final winner remains to be seen.