The Chip Wars: NVIDIA, TSMC, and China's Parallel Stack

Every frontier model you've heard of was trained on chips from one industrial park in Taiwan

Roughly 90% of the world's most advanced semiconductors are manufactured by a single company, in a single country, on lithography machines built by a single firm in the Netherlands. That sentence is not a metaphor.

This guide is about the physical layer of AI — the chips, the fabs, the export controls — and why a few hundred hectares of industrial park in Hsinchu, Taiwan, have become the most strategically important real estate on Earth. By the end you'll know who actually controls the supply, what the United States is trying to do about it, and why China is quietly building a second, parallel version of the entire stack.

NVIDIA: the company that sells the picks and shovels

NVIDIA was founded in 1993 to make graphics cards for video games. The pivot that turned it into the most important company in AI happened in 2006, when it shipped CUDA — Compute Unified Device Architecture, a software layer that let programmers use the GPU (graphics processing unit, originally designed to render polygons in Quake) for general-purpose math. That math, it turned out, looked exactly like what training a large language model requires.

Fifteen years of compounding network effects later, NVIDIA holds somewhere north of 80% of the AI training market — closer to 95% if you count only frontier model training at OpenAI, Anthropic, Google DeepMind, Meta, and xAI. The H100 (the Hopper-generation accelerator that shipped in 2022) became the unit of currency for AI labs; clusters were measured in thousands of H100s, not in dollars. The B200 (Blackwell, launched in 2024) roughly doubled it. NVIDIA's market capitalisation crossed $3 trillion in mid-2024, briefly making it the most valuable public company in the world.

The interesting thing is that the moat isn't really the silicon. AMD's MI300X is, on paper, competitive on raw FLOPS — floating-point operations per second, the unit that measures how fast a chip multiplies numbers. The moat is CUDA. Every PyTorch model, every Hugging Face notebook, every research paper since 2014 assumes a CUDA-compatible GPU underneath. Rewriting that ecosystem for AMD's ROCm or Google's TPUs is technically possible and commercially miserable. NVIDIA didn't win because its chips were ten times faster. It won because its software was fifteen years older.

TSMC: the last mile of Moore's Law

NVIDIA designs chips. It does not make them. The actual fabrication — etching billions of transistors into a wafer of silicon with features measured in nanometres — is done by TSMC, the Taiwan Semiconductor Manufacturing Company, founded in 1987 by Morris Chang. TSMC's Fab 18 in the Hsinchu Science Park is where Apple's M-series, NVIDIA's H100 and B200, and AMD's MI300 are physically made.

TSMC's leading-edge process — N3, the "3-nanometre" node — went into volume production in late 2022. N2 is targeted for late 2025. To do this, TSMC depends on a single supplier, ASML in the Netherlands, for EUV lithography (extreme ultraviolet, the only technology that can pattern features small enough for sub-7nm chips). EUV machines cost around $200 million each; the next-generation High-NA EUV system, of which only a handful exist, costs closer to $380 million. ASML has zero competitors. None.

The result is that the leading-edge supply chain is a chain of monopolies. ASML builds the machines. TSMC operates them better than anyone else. NVIDIA designs the chips that run the models. Remove any one link and the global frontier of AI training stops moving. Samsung Foundry has a 3nm process but is years behind on yield (the percentage of chips per wafer that actually work). Intel is betting everything on its 18A node to catch up. Both might. Neither has yet.

What most people miss is that "3nm" no longer refers to anything physical; the smallest features on an N3 chip are around 12 to 15 nanometres. The number is a marketing convention left over from the 1990s. The real story is that TSMC is roughly 18 to 24 months ahead of everyone else on the things that actually matter — yield, density, power — and that lead has been stable for a decade.

October 2022: when Washington drew the line

On 7 October 2022, the US Bureau of Industry and Security issued export controls that, for the first time, treated advanced semiconductors as a strategic weapon. The rules banned the sale of NVIDIA's A100 and H100 to China, prohibited US persons from working on Chinese sub-14nm chip development, and choked off the equipment China would need to make its own — ASML EUV machines, deposition tools from Applied Materials, inspection systems from KLA.

The logic was explicit. Then-National Security Advisor Jake Sullivan called it a "small yard, high fence" strategy: contain the specific chips that train large AI models, leave consumer electronics alone. Frontier compute was now a dual-use technology in the same legal category as enriched uranium and night-vision optics.

NVIDIA, predictably, tried to thread the needle. It built the H800 — an H100 with the inter-chip bandwidth turned down just enough to fall under the export threshold. Chinese hyperscalers bought billions of dollars' worth before the Bureau closed that loophole in October 2023. By some estimates, tens of thousands of banned or downgraded chips also reached China through second-hand servers routed via Singapore, Malaysia, and the Gulf. Smuggling a server rack is harder than smuggling cocaine, but only marginally so when the customer is willing to pay $40,000 per chip.

Then, in September 2023, Huawei quietly released the Mate 60 Pro. Inside was the Kirin 9000S — a 7nm chip fabricated by China's SMIC, using DUV (deep ultraviolet, the previous-generation lithography that ASML is still permitted to sell to China) and a technique called multi-patterning. It was not supposed to be possible at scale. The Bureau spent the next six months publicly explaining that this changed nothing. Privately, it changed quite a lot.

China's parallel stack

The thesis worth understanding is this: China is not trying to catch up to TSMC at the cutting edge. It is building a complete, parallel version of the AI stack — chips, packaging, memory, software, models — that works without any US or Dutch components. Slower, less efficient, but independent.

The flagship is Huawei's Ascend series. The Ascend 910B, in volume production since 2024, is roughly comparable to NVIDIA's A100 (a 2020 chip) on training workloads — two generations behind, but real silicon, made in China, sold in China. The 910C is sampling. SMIC, China's TSMC equivalent, is producing them on its N+2 process, approximately equivalent to TSMC's 7nm but without EUV. Yields are reportedly poor and costs are high. Beijing does not care.

The software layer is following. Huawei's MindSpore and CANN are CUDA alternatives that nobody outside China would voluntarily use, and that everyone inside China is now strongly encouraged to use. DeepSeek — the Hangzhou lab whose V3 and R1 models embarrassed several US frontier labs in early 2025 — trained its early generations on a stockpile of H800s acquired before the ban, and is migrating workloads to Ascend. Alibaba's Qwen team is doing the same. The point is not that the parallel stack is good. The point is that it exists, and that it gets a little less terrible every year.

The deeper question is what happens at scale. China graduates roughly twice as many engineers a year as the United States. State capital is effectively unlimited. The bet Washington is making is that physics — specifically, the difficulty of doing 5nm lithography without EUV — will hold the line long enough for the gap to compound. The bet Beijing is making is that brute engineering effort and a captive domestic market will close it. Both can be right for ten years before anyone knows which was right forever.

What this means for you

If you build with AI: the chip layer is the binding constraint, not the model layer. Compute prices will not fall as fast as you expect, because demand is growing faster than TSMC can pour new fabs. Plan capacity eighteen months out, not six. If you are choosing between cloud providers, the question worth asking is who has signed long-term allocation agreements with NVIDIA for B200 and B300 — that is the actual scarce resource, not engineering talent.

If you invest: the picks-and-shovels triangle is NVIDIA, TSMC, and ASML, and it has been the best risk-adjusted trade in technology for three years running. The next layer down — high-bandwidth memory (SK Hynix, Samsung, Micron) and advanced packaging (TSMC's CoWoS process, the actual bottleneck on Blackwell shipments) — is where the marginal dollar of upside has shifted. Watch CoWoS capacity announcements the way oil traders watched Saudi rig counts in the 1970s.

If you are trying to think clearly about geopolitics: Taiwan is now to AI what the Strait of Hormuz is to oil, except oil has substitutes and EUV lithography does not. A Chinese blockade of Taiwan would not be a regional crisis — it would freeze the global frontier of AI for at least three years, probably longer. That is the calculation that turns a fab in Hsinchu into a US national security asset, and that turns the chip wars into something that is not really about chips.

Ask Gradient

If you want to push further on any of this — how H100 supply was actually allocated across labs in 2024, what the Stargate datacentre programme means for TSMC capacity through 2027, or whether SMIC can realistically reach 5nm without EUV — try Ask Gradient, our RAG (retrieval-augmented generation, a technique that grounds an LLM's answers in a curated source library) tool. A good question to start with: "How much of NVIDIA's 2025 B200 production was committed to hyperscalers before launch, and what was left for everyone else?" The answer is more concentrated than you'd guess.