US-China War Guide for 2026: Part One

Welcome back and Happy Holidays! This week, I’m bringing to you the most thorough analysis of the US and China I’ve ever made.

It’s a very long article (to cover for my more inconsistent delivery lately, due to the mix of Christmas and JP) that covers everything you need to know about the state of the AI War between the US and China.

We cover everything, literally. Single-chip vs. system-level comparisons, efficiency metrics, supply chain shortages, geopolitical maneuvers, critical bottlenecks, and others; all that’s needed to make up your mind about the state of both countries, and ending with the most important question now that the next year is around the corner.

What to watch out for in 2026? Let’s dive in.

This entire report has more than 14,000 words, so I’ll do it in two parts: hardware for today, software next week.

Setting the stage

The battle for AI supremacy is being fought on two fronts: hardware and software.

And although we are focusing on hardware today, it’s important to understand that, nowadays, hardware is dictated by AI models, which are improved in two ways, known as ‘scaling laws’.

The first scaling law is increasing model size, thereby increasing training budgets; we train a larger model, which allows us to train it on more data.

The second scaling law is inference-time compute, or deploying more compute on every task.

Using a human analogy, one is training a human with a bigger brain, and the other is allowing them to think more deeply about each task.

In a nutshell, larger and longer thinkers are the way to go. Both are important because they naturally help us understand what is going on in the famous AI buildout.

AI hardware scales in three ways:

‘Scale-up’: increasing the number of chips in a single server

‘Scale-out’: increasing the number of interconnected servers, giving you a cluster

‘Scale-across’: increasing the number of interconnected datacenters

And how do each affect both training and inference?

Larger training budgets (training models on more data for longer) dictate how large hardware clusters and data centers get; that’s why data centers are getting so big.

For instance, training a model on a single GPU, even if it’s the most powerful known to humans, the B300, on a GPT-4-level training budget (2022, ancient in AI terms), would take decades (between 35 and 70 years to be more precise).

In other words, what really matters for training is how many accelerators in total you can allocate to a single task, like training a frontier model. It’s not whether each chip doubles performance; what matters is whether you can double chip count, because chip performance grows way slower than what is required today.

On the flip side, inference dictates hardware in another way. A good way to understand this is that, while training only cares about growing chip count, even if that means every sequence is very slow to process, inference is latency-sensitive.

This means that, in an ideal world, inference would…

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