Between the opposition to new data centers and the recent progress of Chinese models, some have started to wonder whether the U.S. can hold its lead in the race to AGI. A recent conversation between Dylan Patel and Dwarkesh Patel puts those worries in perspective. Whoever can extract the most value from a scarce unit of compute will lead that race, and right now that heavily favors the U.S. Two of their arguments explain why. China is years behind on compute and the labs that generate the most revenue per megawatt can afford to outbid everyone else, including the companies hosting open models.

China’s Compute

According to Dylan, China is far behind the U.S. in compute despite remaining surprisingly competitive in model capability. He estimates that leading Chinese labs generally have only 100 to 200 MW of compute, while Anthropic alone is expected to exceed 5 GW by the end of 2026. By 2028, all of China might hold roughly 15% of global AI compute, around 30 GW. Dylan and Dwarkesh credit U.S. chip restrictions with a meaningful share of that deficit, though Dylan is careful to add that export controls don’t tell the whole story. The U.S. financial system has simply been far more willing to pour enormous amounts of capital into AI infrastructure.
The lead is not permanent. Dylan expects Chinese semiconductor production to accelerate sharply from 2028 onward, and denying Chinese labs the best chips has pushed them toward more compute-efficient models and an independent supply chain. Whether any of that matters depends on how quickly AI progress compounds. If increasingly capable AI researchers and automated R&D speed things up while the U.S. still holds a large compute lead, China's ability to produce its own advanced chips may arrive too late to change the outcome.

Why Expensive Compute Hurts Open Models

The more interesting argument is what the compute gap does to the economics of open models. Dylan and Dwarkesh expect OpenAI and Anthropic to keep outbidding other buyers for compute, pushing prices higher for everyone. That matters because an open model is only free on paper. A Western company using Kimi, DeepSeek, or Qwen still needs someone to own and operate the GPUs the model runs on, and that someone pays market rates.

A useful way to think about the economics is this: cost per useful task is roughly the price of compute divided by the number of useful tasks produced per unit of compute. Say you pay OpenAI $1 to complete a task (e.g., translating a long document and summarizing it). Then you discover you can do the same task with Kimi K3 through OpenRouter (an API marketplace where multiple inference providers host models 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 sell that same capacity to OpenAI or Anthropic instead. Your $0.50 Kimi task drifts toward $1.

How can OpenAI and Anthropic afford to outbid everyone? Dylan's argument is that they generate far more economic value per megawatt than anyone else. That advantage compounds through two loops.

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

Or, just as plausibly,

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 closed models versus open models. It is who can generate and capture the most value from a scarce unit of compute. If OpenAI and Anthropic keep a large advantage on that measure, they can pay more for compute, raise costs for everyone else, and erode one of the biggest economic advantages open models have.

Where the Argument Is Weaker

The pricing mechanism is cleaner in theory than in practice, and three things could blunt it.

First, compute is not one market. Frontier labs bid mostly for training-grade capacity on long-term contracts, while much open-model inference runs on older or cheaper hardware. The two markets are linked, but a bidding war at the top does not necessarily pass through to inference prices one for one.

Second, the opportunity-cost argument applies to cloud hosts, not to self-hosting. An enterprise running Qwen on GPUs it already owns has no one to sell that capacity to and no reason to care what OpenAI is paying. Owning the hardware is a big part of why companies choose open models in the first place.

Third, Chinese hosts running Chinese models on Huawei silicon are not in the market OpenAI is bidding in at all. The squeeze lands on Western hosts of open models.

There is also a risk on the other side. If OpenAI and Anthropic cannot release their best models because of safety or government restrictions while open models keep improving, their revenue per megawatt could stall or decline. That would weaken the very advantage that lets them outbid everyone else.

None of these caveats break the argument. They narrow it. Expensive compute is a real headwind for open models, but it hits hardest for companies renting Western inference capacity, and it depends on the frontier labs continuing to convert compute into revenue faster than anyone else.

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