The New Theater of War

Most war historians post-atomic bomb would have told you the next major war would be over resources for building more of these. And for most of the time between then and now, that appeared to remain true. Wars and insurgencies over oil, uranium, gold, silver. Treaties and agreements over who is and who isn’t allowed to develop nuclear weapons. But at some point in the last decade, a new resource entered the global theater of war and national interest.

Silicon. CoWoS. And ultimately: compute.

The stakes have remained the same: global dominance. The war of the new age is over compute power, where you get your electricity from, and how well your models score on HLE.

First Blood

One of the first bloody battles of this new war was the release of DeepSeek R1. Chinese. Open source. Beat every US-made closed-source model on key benchmarks. Allegedly cheaper to train. And it introduced core architectural principles (chain-of-thought distillation, reinforcement learning without supervised fine-tuning) that are now the de facto gold standard in the race toward AGI.

This was the first loss in this AI War, because it proved that US dominance in technology is not a given, it is a position that has to be actively defended.

The Supply Chain is the Front Line

There’s a whole timeline here you could spend hours going through. VLSI lithography and the strategic importance of every TSMC fab currently in the crosshairs of the CCP. The 1.4nm gate-all-around FET production slated to begin this year, pushing high-NA EUV to its theoretical limits. The battle over HBM3e pricing, where SK Hynix and Samsung are locked in a capacity war while NVIDIA consumes every wafer they can produce. The US CHIPS Act and entity list restrictions attempting to choke China’s access to sub-7nm nodes.

Each of these is its own geopolitical chess game. But in my opinion, the next big battle will be over something more fundamental.

The Energy Problem

All of this compute does nothing if you don’t have the power grid to sustain it. China remains comically above the rest of the world in average CO2 emissions while simultaneously commissioning new coal plants. The Paris Agreement was a naive but humble first step toward tackling climate change. The US now faces a much more complex space to navigate: how do you power the infrastructure needed to win the AI race without gutting your climate commitments?

I want to reframe the question though. The function of how much power data centers consume is not strictly tied to the power going in. What’s maybe more useful to look at is: how can we make these systems more power-efficient?

The Hardware Efficiency Race

The compute landscape is no longer just GPUs. Multiple architectures are competing to be the most efficient way to run inference and training at scale:

The thermodynamic computing angle is worth watching. Extropic is building hardware that uses thermal noise, the thing every other architecture tries to eliminate, as the fundamental compute primitive. Instead of fighting Boltzmann distributions in silicon, they sample from them natively. Probabilistic inference at the physics layer, not the software layer. If it scales, it changes the efficiency curve entirely.

This only addresses power consumption at the hardware level. Improvements are still being made in the software layer too. Quantization (FP8, INT4, GPTQ). Mixture-of-experts routing that only activates a fraction of model parameters per token. Speculative decoding. Knowledge distillation. DeepSeek proved you don’t need the biggest cluster. You need the smartest use of the cluster you have.

The US has already shown an ability to learn and adapt to this landscape, but this year, and probably pretty soon we will find out who lands the next major blow.

The Recursive Horizon

It will remain increasingly important to protect these US interests: hardware, software, and infrastructure. The landscape is growing hypercomplex by the week as new models, architectures, and hardware are released. Export controls that made sense six months ago are already being routed around. This is an environment changing far more rapidly than the manufacturing capability of nuclear weapons, which will truly test the efficiency and velocity of our government to maintain proper sightlines on the landscape.

We are also still approaching the recursive improvement singularity. The point where the winner of this race is no longer tied to human input, but to model output.

When an AI system can meaningfully improve its own architecture, its own training data curation, its own optimization passes, the gap between first and second place stops being a lead.

It becomes a moat.