US-China War Guide for 2026: Part Two

TLDR;

A few weeks ago, we discussed the situation between the US and China from the hardware and infrastructure perspective.

Today, we take a look at the software side, why compute budget matters a lot, the market concerns in China, the trick Chinese Labs use to appear close despite not being close, the product moat, and what China will try to do in a situation of comparatively low access to compute to keep up the battle.

We will end with a series of insights that summarise how you should think about the US vs China AI war in 2026.

Together, we’ll learn that comparing the US and China solely based on chip specs and AI model benchmark scores is something only rookies do.

Let’s dive in.

We have bayonettes, they have tanks

To explain China’s software decisions, we need to recall the sources of its issues. And while there are very few things in this world that China is starving for, one of them could be instrumental in determining who wins the AI race: compute.

Summary of what we discussed

Without going into too much detail because I already did, one of the key takeaways from the previous piece was the serious struggles China was facing to keep up with the US in accessing compute, having a production capacity in the order of 50 times less than what the US and Allies can produce (in fact, it’s mostly a comparison with the US/Taiwan/Korean combo).

But why is this so important?

Unless algorithmic improvements change the paradigm completely, currently, there are two ways you make progress in AI:

Increase compute training budget: Training larger models on larger datasets for longer, also known as the first scaling law.

Increase inference-time compute budget: Let models use more compute per task during execution.

As you can see, the common denominator is the need for more compute both for training (first scaling law) and for inference (second scaling law).

Compute. Compute. Compute.

Nonetheless, Chinese AI leaders have been surprisingly outspoken about this great asymmetry.

“The truth may be that the gap is actually widening,” Tang Jie, founder of the Chinese AI startup Zhipu, said at a conference last weekend in Beijing.

Alibaba’s head of development of the Qwen model family gave them a “20% or less” chance of beating OpenAI or Anthropic over the next three to five years.

But then, why are we seeing the media saying China is closing in, that in some respects it’s ahead, yadda yadda yadda? Besides the media being as clueless as always, they have reasons to believe this because they are looking at it from the wrong perspective.

Reframing what is measured

Isn’t China releasing models that appear as smart as US ones? When you look at popular benchmarks, they are very close on many scores (e.g., three in the top 10 of Artificial Analysis main intelligence benchmark), and handsomely leading the way when only comparing open-source models.

Furthermore, Jensen Huang has said that China is “nanoseconds behind”.

Enjoyed the introduction? Subscribe or sign in to continue with this edition.