When I see the recent progress of Chinese models and the opposition to data centers in the U.S., it starts to sow doubt in my mind that we can win the race to superintelligence. Two points from Dylan Patel and Dwarkesh Patel’s recent discussion make me more optimistic. America has, and should continue to have, a massive compute advantage for years to come. If OpenAI and Anthropic can consistently afford to outbid everyone else for compute, the economic advantage of open models starts to disappear.

China’s Compute

According to Dylan, China is currently far behind the U.S. in compute despite remaining surprisingly competitive in model capability. He estimates that leading Chinese labs generally have only 100–200 MW of compute, while Anthropic alone is expected to have more than 5 GW by the end of 2026. By 2028, all of China might have only 15% of global AI compute (roughly 30 GW). 

However, he expects Chinese semiconductor production to accelerate sharply from 2028 onward. How important this becomes depends heavily on the speed of AI takeoff, i.e.,whether increasingly capable AI researchers and automated R&D cause progress to accelerate before China has time to close the compute gap. If that happens while the U.S. still has a huge compute lead, China’s ability to produce more of its own advanced chips may come too late to matter.

This raises the question of whether U.S. chip restrictions have worked. Dylan and Dwarkesh seem to think they have meaningfully contributed to China’s current compute deficit. Dylan adds an important caveat. Export controls are not the entire explanation. The U.S. financial system has also been much more willing to pour enormous amounts of capital into AI infrastructure.

One counterargument is that denying Chinese labs easy access to the best chips has increased the incentive to develop more compute-efficient models while accelerating China’s push to build an independent semiconductor supply chain.

Implications for Open Models in a World With More Expensive Compute

In the same discussion, they argue that OpenAI and Anthropic will increasingly be able to outbid other buyers for compute, pushing prices higher. If compute becomes substantially more expensive, open models start to lose their cost advantage (some are actually quite expensive already). A Western company using Kimi, DeepSeek, Qwen, etc. still needs someone to own and operate the GPUs on which the model runs.

A useful way to think about the economics is like this: cost per useful task ≈ price of compute ÷ useful tasks produced per unit of compute. Let’s say that you’re paying $1 to OpenAI to do a useful task (e.g.,translate long texts and create summaries). Then you find you can complete the same task using Kimi K3 via OpenRouter (an API marketplace where multiple inference providers host models such as Kimi K3 and route your request) for $0.50. Now imagine demand from OpenAI and Anthropic pushes the price of GPU capacity higher. The companies hosting Kimi face higher operating costs and a higher opportunity cost because they could potentially sell that same compute capacity to OpenAI or Anthropic instead. Your $0.50 Kimi task might eventually become a $1 task.

How can OpenAI and Anthropic afford to outbid everyone? Dylan's argument is that they appear able to generate much more economic value per MW of compute than everyone else. 

better models → more valuable tasks → greater demand/willingness to pay → more revenue per MW → ability to bid more for compute.

And or

better models → more productive AI R&D → greater expected future value from internal compute → greater willingness to pay for compute.

So the real competition is not really closed models vs. open models. It’s who can generate and capture the most value from a scarce unit of compute? If OpenAI and Anthropic maintain a large advantage on that measure, they can afford to pay more for compute, raising costs for everyone else and potentially eroding one of the biggest economic advantages of open models. If, however, they can not release their best models because of safety or government restrictions while open models continue to improve, their revenue per MW could stall or decline, weakening their ability to outbid everyone else for compute.

Watch the full conversation:
https://www.youtube.com/watch?v=aV26V1UvkJw