The entire AI trade runs through one company's factories in Taiwan
Nvidia's H100 and H200 GPUs are built exclusively by TSMC. So are AMD's competing AI accelerators, Apple's M-series chips, and the custom AI silicon that Google, Amazon, and Microsoft have each developed as an alternative to Nvidia's architecture. Manufacturing is concentrated across a cluster of facilities in western Taiwan, from Hsinchu and Taichung down to Tainan, Kaohsiung, and the newer packaging hub in Chiayi.
That's not a supply chain. It's a chokepoint with a stock ticker attached.
The bottleneck already moved once
Wafer supply for AI accelerators improved through 2025, and the industry read that as the shortage easing. It wasn't. The constraint shifted to CoWoS advanced packaging, the process that joins multiple chiplets into a single accelerator, which TSMC's own chair confirmed at the company's June 2026 shareholder meeting remains sold out through the end of 2026. Nvidia holds over 70% of TSMC's CoWoS-L capacity specifically, per Morgan Stanley's estimates. The rest is split not just between AMD and Broadcom, but also Google's TPU programme, Amazon's custom silicon, and MediaTek.
Memory is tight too. Micron told investors on its June 2026 earnings call that it expects the high-bandwidth memory shortage to persist beyond 2027, not 2026 as earlier guidance suggested. TrendForce has separately reported HBM3E prices up double digits year on year, with contract prices expected to climb further in 2027.
A single point of political failure, not just manufacturing failure
TSMC's share of the global foundry market reached 72.3% in the first quarter of 2026, according to TrendForce, up from 70.4% the previous quarter. At leading-edge nodes specifically, that share runs considerably higher. The concentration risk isn't only industrial.
Nvidia's own numbers make the point directly. Washington approved licences for H200 shipments to China-based customers in early 2026. Beijing didn't approve the imports. On its most recent earnings call, Nvidia's finance chief confirmed the company has generated zero revenue from China H200 sales and isn't forecasting any. An order book that looked like a demand driver at the end of 2025 became, within months, a live illustration of how much of this trade sits downstream of two governments that don't agree with each other.
The hedge doesn't fix the concentration
Custom ASICs are projected to reach roughly 27.8% of AI server shipments in 2026, rising towards 40% by 2030, with ASIC volumes growing around 44.6% year on year against 16.1% for merchant GPUs, according to TrendForce. That's hyperscalers reducing their dependence on Nvidia's allocation decisions specifically. It isn't reducing dependence on TSMC. Every one of those custom chips still gets built in the same facilities.
What this means for how the trade is priced
TSMC raised its 2026 capital budget to $60-64 billion in July, up from a prior $52-56 billion range, and lifted its own full-year revenue growth guidance to slightly above 40%. Markets are pricing this expansion as though it resolves the concentration problem. It scales the same single point of failure instead. Arizona remains roughly a node behind Taiwan on process technology, and chips made there still return to Taiwan for CoWoS packaging today.
This is analysis and opinion, not investment advice.