A 40x Compute Gap — So Why Hasn't China Fallen Far Behind in AI?

A Counterintuitive Question

There's a widely held belief in the AI world: compute is the moat. Whoever controls more GPU clusters can train more powerful models. Yet reality presents a fascinating paradox — despite the enormous gulf in compute resources between American and Chinese AI labs, China's leading labs haven't fallen as far behind as one might expect in terms of model performance.

In a recent interview, renowned semiconductor and AI industry analyst Dylan Patel offered a remarkably insightful explanation for this phenomenon. Patel is the founder and chief analyst of SemiAnalysis, a research firm known for its deep dives into chip architectures, data center economics, and AI training infrastructure. Their reports are widely cited in Wall Street investment firms and tech companies' strategic planning. Patel's ability to access such granular data on compute allocation stems from his long-standing tracking of NVIDIA's supply chain, data center construction plans, and deep engagement with major AI labs.

His core argument: The absolute gap in compute is real, but its role in determining a model's "current performance" may be overestimated.

A Staggering Compute Chasm

Let's start with the numbers. According to Dylan Patel, China's leading AI labs have at most 100 to 200 megawatts (MW) of total compute capacity. By comparison, Anthropic's compute capacity will approach 5 gigawatts (GW) by the end of this year.

It's worth explaining why the AI industry measures compute scale in electrical power (watts) rather than GPU counts. The reason is that modern AI clusters have grown so large that they're planned at the data center level, and the core constraint for data centers is power supply. A 100 MW AI data center can house roughly 20,000 to 30,000 NVIDIA H100 GPUs (accounting for the additional power draw of cooling, networking equipment, and storage). For reference, a typical coal-fired power plant has a capacity of about 600 MW, and Anthropic's planned 5 GW by year-end is equivalent to the total output of roughly 8 such plants — which is why American tech giants are investing heavily in nuclear power and renewable energy. Electricity is becoming the most fundamental physical bottleneck in the AI race.

Do the math: 1 GW equals 1,000 MW. This means Anthropic's compute reserves are roughly 25 to 50 times those of China's top labs. That's an order-of-magnitude, crushing advantage. Common sense would suggest the quality of models trained by the two sides should be worlds apart.

But that's not the case. In Patel's words: "Chinese companies today, on AI models — at least from a public perception standpoint — are not that far behind relative to the compute they have." This leads to the question that truly deserves investigation: Where is all that massive compute actually going?

Breaking Down an AI Lab's Compute Budget

Patel's answer upends many people's intuitive understanding of what "training a model" entails. He provides a detailed breakdown of an AI lab's compute budget.

The Traditional Training-Inference Split

Historically, a lab's compute is allocated roughly 60% to training and 40% to inference. Inference is the compute consumed when serving users after a model goes live — every time you ask Claude or ChatGPT a question, GPUs in the backend are running the model to generate a response. That part is straightforward. The real mystery lies within that 60% labeled "training."

A Deeper Look at the Training Budget

When Patel further dissects the "training" category, a critical insight emerges:

• ~50% of total compute goes to Research
• ~10% of total compute goes to Development
• ~40% of total compute goes to Inference

"Research" here refers to researchers generating new ideas, testing novel model architectures, and experimenting with different data mixes. Data mix tuning is one of the highest-leverage research directions in AI today — it refers to the proportional blend of different data types (such as code, math, scientific papers, web text, books, conversational data, etc.) in the pre-training dataset. What seems like a simple ratio adjustment can profoundly impact a model's capability distribution. For example, increasing the proportion of code data has been found to not only improve coding ability but also enhance logical reasoning. Google DeepMind invested heavily in data mix experiments when training Gemini, as did Meta when training Llama. Each mix configuration requires training a mid-sized model and running comprehensive evaluations to assess its effects — this is a textbook example of how "research compute" gets consumed in large quantities.

"Development," on the other hand, is what we typically think of as "training a production model."

In other words, the compute actually used to train the final released model represents only a small fraction of the overall budget.

The Truth Behind Anthropic Training Mythos

To back up this point, Patel cited a specific case from Anthropic's training of its model (codenamed Mythos) — and it's highly compelling.

Anthropic was founded in 2021 by former OpenAI Research VP Dario Amodei and Daniela Amodei, with "AI safety" as its core mission. It's one of the most closely watched AI labs in the world today. Its Claude model series stands out for its reasoning capability and safety properties. As of 2025, Anthropic has raised over $13 billion in cumulative funding, with major investors including Google and Amazon. A significant portion of these funds has been used to build and lease large-scale GPU clusters, supporting its leap from hundreds of megawatts to gigawatt-scale compute. The "Mythos" mentioned here is an internal codename at Anthropic, corresponding to a specific generation of its public Claude model lineup.

When Anthropic actually executes the model's pre-training, its peak compute consumption is under 200 MW, sustained for approximately two months. Pre-training is the first and most compute-intensive phase of building a large language model — the model performs unsupervised learning on massive text data (typically trillions of tokens), learning statistical patterns of language and world knowledge by predicting the next word. This process requires thousands or even tens of thousands of GPUs working in concert for weeks to months. The subsequent reinforcement learning (RL) phase consumes "even less." The RL phase follows pre-training and uses techniques like Reinforcement Learning from Human Feedback (RLHF) or more recent rule-based reward methods (such as GRPO) to align the model's behavior — teaching it to follow instructions, refuse harmful requests, and provide more useful responses. RL computation is typically far less than pre-training because it fine-tunes a model that already has foundational capabilities rather than learning from scratch.

What does this mean? Anthropic sits on multiple gigawatts of compute, yet at any single point in time when training a production model, it uses at most about 200 MW — a figure that roughly equals the total compute capacity of China's leading labs.

Patel specifically emphasized a key concept: sequentiality. Anthropic's total compute consumption is indeed enormous, but it's distributed across a timeline, occupied by different research experiments running in parallel or in sequence. This concept reveals a fundamental constraint of AI research: many experiments must wait for the results of a previous step before designing the next one. For example, researchers may need to first determine whether a new attention mechanism works before deciding whether to adjust training hyperparameters on that basis. This dependency means that even with unlimited compute, the speed of progress along any single research thread is limited by experimental cycles and the pace of human cognition. But the real advantage of massive compute lies in "breadth" — the ability to simultaneously pursue dozens or even hundreds of independent research tracks, each exploring different architectural innovations, training strategies, or data processing methods. This is essentially a strategy of using parallelism to hedge against sequential constraints, similar to the "cast a wide net" approach in venture capital, dramatically increasing the probability of discovering breakthrough methods.

The vast majority of compute is poured into the "research" phase — continuous trial and error, exploring new architectures and methods — rather than the final "development" of any single model.

What This Means for Chinese AI

Synthesizing Patel's analysis, we can draw several thought-provoking conclusions.

Short-Term: The Compute Gap Has Limited Impact

Since the "critical compute" needed to train a specific model isn't actually that high (~200 MW is sufficient), Chinese labs are fully capable of training models on par with their American counterparts. What they lack isn't the compute to train a single model, but the research compute for large-scale parallel exploration. This also explains why, from a public perception standpoint, Chinese AI models "haven't fallen that far behind."

Long-Term: The Compounding Effect of Research Compute Cannot Be Ignored

However, the other side of this story deserves equal attention. Devoting 50% of compute to research is essentially a "scientific accelerator." Whoever can run more experiments simultaneously, testing more architectures and data mixes, will discover superior methodologies at a faster rate. This advantage produces a compounding effect over time.

This compounding effect can be understood across multiple dimensions. First, methodological compounding: every successful architectural innovation (like the invention of the Transformer, the optimization of Flash Attention, the application of Mixture of Experts) improves the efficiency of all subsequent training runs, effectively enabling the same compute to produce stronger models. Second, talent compounding: labs that can run more experiments validate researchers' ideas faster. This rapid feedback loop attracts and retains top talent, who in turn produce better ideas, creating a positive flywheel. Third, data compounding: extensive experimentation generates meta-knowledge about "what works and what doesn't" — knowledge that is itself an extremely valuable R&D asset, guiding the direction of future experiments and avoiding repeated mistakes. These three layers of compounding stack on top of each other, meaning the long-term return on research compute likely far exceeds linear growth.

In other words, the vast compute gap between the U.S. and China hasn't yet fully materialized in model capabilities, but it's quietly translating into an advantage in "R&D velocity" and "exploration breadth." As American labs leverage their massive research compute to continuously iterate on new paradigms, this lead may crystallize and manifest all at once at some future inflection point.

Conclusion

Dylan Patel's analysis offers us a more nuanced lens through which to view the U.S.-China AI competition: The compute gap is real, but it currently manifests primarily in "research capability" rather than "model training capability."

Training a top-tier model may require only ~200 MW of compute — a threshold that isn't prohibitively high. But to maintain a sustained lead and continually open up new technical frontiers requires gigawatt-scale research compute. The fact that Chinese AI "hasn't fallen far behind" is today's reality, but the long-term trajectory of this race may ultimately be determined by the compounding returns from that invisible 50% of compute invested in research.