As of August 2025
The AI race people argue about - ChatGPT versus Claude versus Gemini - is the surface layer, where anyone can build for almost nothing. The contest that decides it runs four layers down: chip design, fabrication, the supply controls between them, and the infrastructure spend on top.
Each layer is a chokepoint. One ASML EUV machine costs around $250 million, ASML is the only maker, and it cannot ship them to China. A leading-edge fab runs about $20 billion, and most advanced chips come from TSMC. That is the board the export controls are played on.
The supply layer is where the strategy showed its hand. Starting in October 2022 the US restricted China’s access to the top GPUs, banning A100 and H100 sales, then tightening again in October 2023 to cover the cut-down H800. The bet was straightforward: more compute makes better models, so denying compute slows the rival down.
Lithography
$250M
the cost of one ASML EUV machine - the sole maker, and barred from selling to China.
Export controls
Oct 2022
the US begins cutting China off from A100 and H100 chips, with H800 restricted in October 2023.
DeepSeek V3
2,048
H800 GPUs and about 2.8 million GPU-hours used to match frontier models (released Dec 26, 2024).
The brute-force playbook
The US frontier labs leaned into scale. Meta trained Llama 3.1 on roughly 16,000 H100 GPUs. xAI stood up around 350,000 GPUs in Memphis, and the Stargate plan announced in January 2025 talks in millions. The assumption underneath all of it is that the path to better AI is paved with more hardware.
The constraint that bit back
DeepSeek did not have that option. Stuck with about 2,048 H800 chips and a tighter budget, the team optimized the algorithm instead. DeepSeek V3, released December 26, 2024, reached frontier-class performance on roughly 2.8 million GPU-hours - by the team’s own framing, about an order of magnitude less training compute than the scale-first approach assumed it would take.
Sometimes the limitation becomes the breakthrough. The lab that could not buy the best chips showed that better algorithms can matter more than more hardware.
