Key Takeaways

  • AI chips account for less than 0.2% of all semiconductors sold but generate roughly 500 billion Dollar in revenue this year, about half the total chip market.
  • The US now treats advanced chip exports as a geopolitical lever. Nvidia can sell H200s to China, but only with a 25% revenue tax to the US government. Blackwell remains banned.
  • TSMC manufactures roughly 90% of the world's most advanced chips. That single-point dependency defines the entire AI supply chain.

The Setup

A chokepoint is any single node in a supply chain where one player, or a small group, controls enough capacity to dictate terms for everyone else. Oil had OPEC. Diamonds had De Beers. AI chips now have a version of both.

The AI chip market is projected at 500 billion Dollar for 2026, according to Deloitte. That is roughly half the total semiconductor market, which sits near 975 billion Dollar. The twist: AI chips represent under 0.2% of all semiconductor units sold. Almost all the value is concentrated in a tiny number of very expensive processors.

What AI Chips Actually Are

Standard chips run phones, cars, appliances. AI chips are different. They are designed specifically for training and running large language models, image generators, and other AI systems. The dominant product right now is Nvidia's GPU lineup. Their H200 sells for around 32,000 Dollar per unit. The newer Blackwell architecture pushes that ceiling even higher.

Nvidia held roughly 80% of the AI accelerator market in 2025. AMD is the closest competitor with its MI300X series, but volume remains a fraction of Nvidia's. Google and Amazon build custom AI silicon for their own cloud platforms, but these chips are not sold on the open market.

The TSMC Bottleneck

Nvidia does not manufacture its own chips. Neither does AMD, Qualcomm, or most of the industry. They design chips, then send the blueprints to a foundry. For anything at the leading edge, that foundry is TSMC in Taiwan. TSMC produces roughly 90% of the world's most advanced semiconductors. Samsung can do some leading-edge work, and Intel is trying to rebuild its foundry business, but neither comes close to TSMC's volume or yield rates.

This means one company, on one island, in one of the most geopolitically sensitive regions on earth, controls the physical production of nearly all frontier AI hardware. Every major AI lab, every hyperscaler, and every government AI program depends on TSMC. The stock trades at 326.81 Dollar as of March 27.

The Export Control Lever

The US government cleared Nvidia to sell H200 chips to China in January 2026. But with conditions. Washington takes a 25% cut of the revenue. ByteDance, Alibaba, and Tencent got approval for over 400,000 units at around 32,000 Dollar each. That is roughly 12.8 billion Dollar in hardware, with about 3.2 billion going to the US Treasury.

Blackwell, Nvidia's most advanced architecture, remains completely banned from export to China. The calculation: let China buy the previous generation at a premium, keep the frontier technology out of reach, and generate revenue for the US government in the process. Nvidia stock sits at 167.69 Dollar, reflecting the complexity of navigating these restrictions while maintaining growth.

Why It Matters Now

Can China build its own advanced AI chips to get around the controls? Huawei's Ascend 910C is the best domestic alternative, but independent benchmarks suggest it trails Nvidia's current generation by a meaningful margin. SMIC, China's leading foundry, still works primarily at older process nodes. Building a fully independent advanced chip supply chain takes years, massive capital, and institutional knowledge that cannot be easily replicated.

The structural picture is clear. A small number of companies control the most valuable part of the semiconductor supply chain. The US government is using that concentration as a policy tool. And every major AI project in the world runs through this chokepoint, whether they like it or not.

FAQ

Why did the US allow H200 sales at all if the goal is to restrict China?
The calculation was that keeping older-generation chips available at a heavy tax generates revenue and maintains some commercial engagement, while the true frontier technology, Blackwell, stays out of reach. A full ban risks accelerating China's domestic chip development without any financial upside for the US.
Could TSMC be replaced as the dominant foundry?
Not quickly. Intel and Samsung are investing billions in foundry capacity, but TSMC's lead in yield rates, process technology, and volume is measured in years. The US CHIPS Act is funding domestic production, but new fabs take 3 to 5 years to reach full output.
What happens to AI development if chip supply gets disrupted?
Training frontier AI models requires tens of thousands of advanced GPUs running for months. Any supply disruption, whether from export controls, natural disaster, or conflict near Taiwan, would immediately slow or halt new model development globally. Inference, running existing models, is less vulnerable because it can use older hardware, but scaling would stop.