Nvidia CEO Jensen Huang told the Computex audience in Taipei this week that his company will spend $150 billion a year in Taiwan — not on R&D, not on US data centers, but on chips, packaging, and the network of suppliers that turns sand into the silicon running the AI revolution.

According to Reuters, Huang called Taiwan the “epicentre” of the AI revolution. The $150B figure isn’t a five-year capex plan. It’s annual. Recurring. A floor, not a ceiling.

🔍 THE BOTTOM LINE: $150B a year in a single country, and a country China has not ruled out taking by force, makes Nvidia’s Taiwan bet the most concentrated geopolitical exposure in the history of the semiconductor industry. The AI race runs through a 100-mile strait. That’s not a supply chain. That’s a hostage situation with a profit margin.

What $150 Billion a Year Actually Buys

The number is so large it loses meaning. To anchor it:

• It’s roughly the entire 2025 GDP of New Zealand — spent every year, in one country, by one company
• It’s ~30% more than TSMC’s 2025 total revenue ($115B)
• It’s over 4x the entire US CHIPS Act allocation ($53B for the whole industry over five years)
• It’s enough to build 25 gigawatt-scale data centers at current construction costs

TSMC manufactures roughly 90% of the world’s leading-edge chips. Nvidia’s flagship H100, B200, and B300 GPUs are TSMC products. Every Blackwell chip that trains GPT, Claude, or Gemini comes out of a fab in Hsinchu or Tainan that sits 110 miles from mainland China artillery.

Huang isn’t diversifying. He’s accelerating into the most concentrated supply chain on Earth and calling it strategy.

The Concentration Problem

For years, the AI industry told itself a comforting story: that Taiwan exposure was a temporary bottleneck that US fabs (Intel, TSMC Arizona, Samsung Texas) would eventually ease. The 2024-2025 CHIPS Act money was supposed to onshore 20-30% of leading-edge production by 2027.

That hasn’t happened at the scale anyone hoped. TSMC Arizona is producing 4nm chips at modest volume. Samsung’s Texas fab is years behind. Intel’s foundry business is, charitably, rebuilding.

Meanwhile, Nvidia’s orders keep growing. The $150B annual figure implies Nvidia is buying more leading-edge wafers from TSMC than the entire global smartphone industry did at its 2018 peak. There’s no other customer. There’s no other supplier. The dependency is mutual — and unprecedented.

China’s live-fire exercises around Taiwan have not stopped. The PLA’s most recent simulation reportedly modelled a blockade, not an invasion — slow strangulation, not amphibious assault. The kind of operation that would halt TSMC shipments without a single shot.

The Rivals Who Can’t Replicate This

Google just scaled to 960,000 GPUs and is designing its own TPUs with Marvell. Anthropic is buying 300 megawatts from SpaceX and growing 8x year-over-year on Claude Code. Meta is firing 8,000 workers to fund its AI infra bill.

None of them can build a TSMC. The moat isn’t Nvidia’s GPUs — it’s Nvidia’s relationship with the only fab in the world that can make them at volume. AMD’s MI400 series is bottlenecked at TSMC. Every hyperscaler designing custom silicon (Google TPU, AWS Trainium, Microsoft Maia) goes through TSMC packaging.

The $150B figure is, functionally, a tax every AI lab pays to the same supplier — and Nvidia is the toll collector.

The NZ Angle

New Zealand is small in absolute AI compute terms. But the structural lesson applies: when a single supplier sits in a single geopolitically-exposed jurisdiction, every customer downstream carries the same risk.

Our local compute story — government-backed NVIDIA H200 clusters at NIWA, private-sector experiments at universities — runs on the same TSMC output. If the strait closes for 90 days, the global AI industry loses three quarters of leading-edge production. Kiwi researchers don’t get a carve-out.

For policymakers, the question is whether to mirror Australia’s approach (significant sovereign compute investment, AUKUS-style AI partnerships) or continue treating compute as a market externality.

For the rest of us: every AI product you’ve integrated into your workflow in 2026 runs through chips that were made in a place that, by Beijing’s own statements, may not be making chips in three years.

The Other Side

Huang’s argument: Taiwan is the right bet because the talent, the supplier network, and the manufacturing know-how aren’t replicable anywhere else — not in five years, not in ten. Building alternatives in parallel is good. Betting the company on them would be foolish. “Epicentre” is just an honest description of where the work happens.

The bull case: $150B a year is what it costs to stay on the leading edge. If Nvidia diversified prematurely, it’d be building 2nm fabs in Arizona that trail TSMC by a node — and TSMC node leadership is the entire reason Nvidia’s GPUs are the ones everyone wants. The concentration is rational, not pathological.

The bear case: The most concentrated supply chain in tech history sits inside the most contested piece of geography on Earth. No discount rate justifies the tail risk. And if the worst happens, $150B a year doesn’t insulate Nvidia — it just means there’s $150B a year of stranded capacity.

The Bigger Picture

Huang’s $150B figure is the most explicit admission yet that the AI industry is infrastructure-bounded, not capital-bounded. There’s more money than there are chips. The bottleneck is fabs, packaging, HBM memory, and the small handful of people who can run advanced process lines.

That bottleneck now has a name and a zip code: Hsinchu Science Park, Taiwan. The AI revolution is being built on an island the size of Maryland, by a workforce smaller than Apple’s, for a global market that will spend half a trillion dollars on AI infrastructure in 2026 alone.

We’ve covered the AI capex arms race before and the strange economics of GPUs in 2026. What Nvidia said at Computex this week is the end of the illusion: this is a Taiwan story, told in dollars, that the entire AI industry is now underwriting.

❓ FAQ

Is $150B a year actually new, or just a restatement of existing plans? Both. Nvidia had guided to roughly $80-100B in annual Taiwan spend through 2027. The $150B figure is a step-change, implying both higher volumes (more chips) and higher per-unit prices (more advanced packaging, more HBM).

Could US fabs absorb this if TSMC went down? No. Not within 24 months. Even at full ramp, US-based leading-edge capacity would cover ~20% of current Nvidia demand. The other 80% doesn’t exist anywhere else.

What about Samsung Korea? Trailing TSMC by roughly 12-18 months on leading-edge nodes. Useful for legacy production, not for Blackwell-class AI training chips.

Is Nvidia hedging at all? Minimally. Intel Foundry has been awarded some packaging and older-node work. TSMC Arizona is producing 4nm at modest scale. The vast majority of $150B flows to Taiwan proper.

What’s the worst-case scenario? A Chinese blockade of Taiwan that halts all TSMC exports. The AI industry would lose ~70% of leading-edge chip supply for the duration. Training runs would stop. Inference fleets would age out without replacements. Estimated global GDP impact: $1-3 trillion in the first quarter.