The AI Pipeline, Part 3: The Buyers and the Export Control Era

The short answer: five hyperscalers hold roughly 71 percent of global AI compute, and since August 2022 every layer of that concentration has been used as leverage. US license requirements cost NVIDIA 4.5 billion dollars in a single quarter, ASML halted advanced equipment exports to China in January 2024, TSMC lost its China waiver in September 2025, China banned NVIDIA chips that same month, and rare-earth export curbs pushed terbium prices up about 350 percent.

This is part 3 of the AI Pipeline series. Part 1 covered the minerals and the design software. Part 2 covered the machine, the fab, and the memory. This part covers the two ends of the funnel at once: the tiny group of companies that buys everything, and what happens when the governments behind the supply chain start using those chokepoints against each other. Because every layer of this pipeline is a chokepoint, every layer is also leverage, and since 2022 it has all been in active use.

Who buys all of this#

The chips get bought by a tiny set of companies. Three hyperscalers (AWS, Microsoft Azure, Google Cloud) hold roughly two-thirds of global cloud infrastructure, with estimates running 30-32% AWS, 20-23% Azure, 12-13% Google. The hyperscale data center count crossed a thousand with the US hosting over half the capacity. Epoch AI's tracking puts five hyperscalers at roughly 71% of global AI compute. The largest known AI data center runs around 1.1 million H100-equivalent GPUs, and a gigawatt-scale facility costs on the order of $38 billion in capex.

NVIDIA sits between the fabs and the clouds, and its position is even tighter than the fab layer's. Estimates of its AI accelerator share run from 70% to 95%. The moat is not only the chip; it is CUDA, the software platform (compiler, libraries, tooling, the CUDA-X stack) that every major deep-learning framework, including PyTorch, directly targets. Training compute for frontier models grows about 4.5x per year, and NVIDIA's Blackwell generation is the current flagship, with the Vera Rubin platform now in full production: a Rubin GPU with 336 billion transistors, 288GB of HBM4 and 22TB/s of bandwidth, a Vera CPU, NVLink 6, and a rack-scale system where seven chips across five rack-scale systems behave as one supercomputer, with a claimed 10x jump in agentic-AI throughput per unit of energy over Blackwell.

The buyers are not passive about this dependency. The trillion-dollar project to fragment NVIDIA's monopoly is the largest industrial story in tech right now: Google's TPUs (a line of custom AI processors with a history going back over a decade, wired into some of the largest clusters on earth), AWS's Trainium and Inferentia, Microsoft's and Meta's custom silicon, AMD's MI series. Each hyperscaler builds or buys alternatives because a single supplier's pricing power over a trillion-dollar capex cycle is intolerable to them. None has dented the top of the market yet. All of them keep trying, and AI is fundamentally a capital-expenditure story, which is why NVIDIA's valuation sits above OpenAI's, the company it enables: the market is pricing the infrastructure, not the chatbot.

On top of that infrastructure sit the frontier labs: OpenAI, Anthropic, Google DeepMind, xAI, Meta. Their products are the visible face of the pipeline. ChatGPT alone claims around 700 million weekly users with a roughly 30% work / 70% everything-else usage split, over a million direct business customers, and enterprise message volume that grew about 8x in a year. Google says about 75% of its new code is AI-generated. There are over 30,000 AI companies worldwide, roughly one in four American, and 44 of Forbes' AI 50 are US-based. The 2026 AI Index counted 59 notable models from US companies against 35 from Chinese ones, with only a scatter from anyone else. Adoption runs ahead of every earlier technology wave: 85% of surveyed healthcare leaders were exploring or had adopted generative AI, and on consumer apps, ChatGPT, Gemini, and DeepSeek account for nearly 90% of time spent on AI assistants globally.

The export control era#

The first formal move came in August 2022, when NVIDIA disclosed that the US government had imposed a license requirement on A100 and H100 exports to Russia and China. That single disclosure started the era we are still in.

It is worth remembering that gating who can access strategic technology has history. Between 1905 and 1936, 14 of the 38 Nobel Prizes in science went to Jewish scientists, and when Nazi Germany began persecuting them, the knowledge did not disappear. It relocated, and the countries that received it inherited a generation of scientific leadership. Gating access to knowledge and capability always produces movement, not obedience. Everyone designing export regimes in 2026 is, knowingly or not, running that experiment again.

The timeline since 2022, laid out flat:

• January 2024: ASML halts advanced chipmaking equipment exports to China after US request, and partially revokes licenses for its most advanced DUV systems; the Netherlands expands its control measures later that year
• 2024: TSMC is barred from exporting 7nm-and-below chips to Chinese customers, after TSMC silicon was found inside a Huawei Ascend AI processor (a Chinese designer had ordered chips through a third party, Sophgo, and TSMC suspended that customer immediately upon discovery); TSMC then told Chinese customers it would suspend fabrication of advanced AI chips at 7nm and below without regulatory approval
• May 2025: a new US license requirement lands on the H20, NVIDIA's China-market chip, triggering a $4.5 billion write-off and roughly $8 billion in lost H20 revenue in a single quarter; licences were later granted
• September 2025: the US pulls TSMC's waiver for shipping chip supplies to China. The same month, China bans NVIDIA AI chips outright, and Chinese buyers shift toward Huawei and Cambricon
• April 2026: the US proposes export restrictions targeting chipmaking tools and servicing for Chinese firms, notable because China was 33% of ASML's 2025 sales
• 2026: the US makes it harder for SK hynix and Samsung to manufacture chips in China, extending the squeeze to the Korean memory makers' Chinese operations

Russia is the cautionary tale here. Its chips were designed at home and fabricated abroad, and after the Ukraine invasion TSMC confirmed compliance with export controls, cutting off Russia's domestically designed Baikal and Elbrus processors, which had nowhere else to go. The machinery was ready for this: the foreign direct product rule first used against Huawei in 2020 (blocking any foreign-made chip built with US technology) was explicitly set up to be reusable against Russia. Huawei's smartphone business, once globally dominant, was crippled; the company itself estimated a $30-40 billion smartphone revenue loss in 2021 alone. A country can design processors and still lose its computing future if it does not control fabrication.

The US also tried to formalize a global AI hierarchy. The January 2025 "AI Diffusion" framework would have tiered every country on earth by how much compute and model access it could receive, covering both chips and certain model weights, effectively licensing AI like nuclear material. It was rescinded before its compliance provisions took effect, but the instinct did not go away; it moved to deal-by-deal compute diplomacy, most visibly with the UAE and Saudi Arabia, where export approvals are exchanged for security and investment commitments, and the Stargate campus in the Gulf. Model-layer controls are part of the same picture: Anthropic's frontier models faced US access restrictions that were later lifted, OpenAI limited a GPT-5.6 rollout after a government request while publicly noting restrictions should not become the norm, and NATO doctrine now talks about AI as a source of "cognitive edge" over adversaries, which tells you what category governments place this technology in.

China's counter-moves#

China's answers operate on the chokepoints it holds. Rare-earth export controls rolled out from late 2023 through 2025: processing technology exports banned outright in December 2023, custom license regimes for specific elements announced by the Ministry of Commerce in 2025, gallium and germanium exports falling to zero in the months after the first curbs, and prices moving accordingly (terbium up about 350%, dysprosium up about 450% at one point, germanium more than doubling). Japan has spent years trying to de-Chinafy its rare-earth supply chain and has made only partial progress, which tells you how hard refining substitution actually is. One US government estimate puts the cost of a 30% gallium supply disruption at about $600 billion in American output. If chips are the US leverage point, minerals and refining are China's, and the 2025 rare-earth curbs arrived in direct response to the chip curbs.

Meanwhile, China is building its way out of the lithography trap. Reports describe a covert EUV reverse-engineering program, with 13.5nm light generated in a domestic lab and prototypes targeted around 2028, plus limited production of homegrown immersion DUV tools as of mid-2026. The strategy underneath is what one think tank called the lithography loophole: stay on older DUV nodes but multiply patterning steps to approximate advanced density, at the cost of yield and economics. Alongside it, Huawei's CloudMatrix 384 clusters brute-force around the GPU shortage by wiring far more, older chips together (including HBM stacks reclaimed through grey markets), and domestic HBM production is scaling, with Chinese memory firms racing to close a gap their own industry measures in the hundreds of billions of dollars. Independent analysis (CIGI, SemiAnalysis, and others) converges on the same read: none of this matches TSMC-plus-ASML quality today. But it does not need to. It needs to be good enough to keep the lights on while the gap closes, and the scale of investment behind it ($1 trillion-class national chip targets by some measures) means the question is when, not whether, the quality gap narrows.

Two economic footnotes worth keeping, because they frame the whole era. First, the buildout is already visible in macro data: the St. Louis Fed tracks AI data centers, software and research contributing about 0.9 percentage points to quarterly US GDP growth. Second, every one of these measures is a calculated cost. NVIDIA eats billions writing off China revenue; ASML gives up a third of its sales; China pays inefficiency premiums on DUV multipatterning; the US pays subsidy bills for onshoring. Nobody is winning cleanly. Everyone is paying for optionality.

What the export-control era reveals about the pipeline is the subject of part 4: it turns out the power to grant or deny access to compute has a geography, and most of the world, including the countries doing the most AI adoption, does not sit inside any of the walled gardens. India is the most instructive case, and it is where the series goes next.

The full series#

• Part 1: Dirt, Sand and the Software That Draws Chips
• Part 2: One Machine and One Island
• Part 3: The Buyers and the Export Control Era (you are here)
• Part 4: The Geography of Compute and the India Lens
• Part 5: Nobody Owns the Stack, Everyone Owns a Switch