🔌 The AI Dollar: Part 2/6: Compute Is the New Oil

January 29, 2026Bloomberg: $35/month. Financial Times: $42/month. The Economist: $17/month. Original analysis by Tatsu with 40+ footnotes: $8/month.Share this preview with others.The petrodollar is dying. Something else is being born.In 1974, Henry Kissinger negotiated a deal that would define American power for half a century: Saudi Arabia would price oil in dollars exclusively and invest its surplus in Treasury bonds. In exchange, the United States would guarantee Saudi security. Every country that needed oil, which was every country, would need dollars first. The dollar's reserve status was no longer just about American economic strength; it was encoded into the physical infrastructure of global energy.That system is slowly unwinding. Saudi Arabia has quietly begun accepting yuan for some oil sales. The energy transition, however incomplete, is reducing oil's stranglehold on industrial economies. The petrodollar architecture remains standing but the foundation is shifting.What the dollar doomers miss is that a new architecture is being constructed, perhaps consciously, perhaps accidentally, that could prove even more durable. Not oil, but compute. Not barrels, but teraflops. Not OPEC, but NVIDIA.Full investigation below. $8/month for novel, footnoted deep analysis.The Logic of the AI DollarThe petrodollar worked through a simple chain:Everyone needs oil → Oil is priced in USD → Everyone needs USDThe AI Dollar works through an analogous chain:Everyone needs AI/compute → US controls the chokepoints → Access is denominated in USDThis isn't metaphor. It's already happening, just without the branding.Every AI startup in the world, whether based in Munich or Mumbai or Melbourne, pays its cloud computing bills in dollars. AWS charges in dollars. Azure charges in dollars. Google Cloud charges in dollars. The three hyperscalers that control the overwhelming majority of global cloud infrastructure are American companies billing in American currency.Every company accessing frontier AI models pays in dollars. OpenAI's API is billed in dollars. Anthropic's API is billed in dollars. Google's Gemini API is billed in dollars. The marginal cost of intelligence is denominated in USD.And every country trying to build sovereign AI capability faces a choice: buy American chips at American prices in American dollars, or try to build your own from scratch. China chose the latter. We'll examine how that's going in Part 3 (spoiler: not well).The Chokepoint TableThe AI supply chain is a series of bottlenecks, each controlled by the United States or its close allies. Let me map them:Layer | US/Allied Control | Alternative --------------------+----------------------+----------------------------- Chip Design | Nvidia (80%+ of AI | AMD (distant second), Intel | GPUs) | (struggling) Chip Fabrication | TSMC (90%+ of | SMIC (China, stuck at 7nm | advanced nodes), | with terrible yields) | Samsung | EUV Lithography | ASML (Netherlands, | None. Literally none. | 100% monopoly) | Cloud Compute | AWS, Azure, GCP | Alibaba Cloud, Tencent | (dollar-denominated) | (China domestic only) Frontier Models | OpenAI, Anthropic, | DeepSeek, Baidu (catching | Google, Meta | up, maybe) AI Talent | Brain drain TO | Brain drain FROM everywhere | United States | else Capital | Unlimited (US | Constrained (China property | venture + corporate) | crisis, Europe risk-averse) Energy for Training | Abundant natural | Coal dependent, grid | gas, nuclear | instability | buildout |Look at that table and tell me which row China wins. Talent? 87% of top Chinese AI researchers who publish at elite conferences choose to stay in the United States. Capital? China's property sector defaulted on $300 billion while American AI companies raised $108 billion in 2024 alone. Energy? China runs on coal and has grid stability issues; the US has abundant natural gas and is building nuclear specifically for AI datacenters.The only row where China competes is frontier models, and even there, DeepSeek's recent success was achieved using smuggled Nvidia H800 chips, not domestic Huawei hardware. The software is good. The infrastructure underneath is American.The ASML ChokepointLet me dwell on one row of that table because it illustrates the depth of American structural advantage.ASML is a Dutch company that makes the machines that make advanced chips. Specifically, they make Extreme Ultraviolet (EUV) lithography systems, the $200 million devices that etch circuits at the 7nm node and below. There is no alternative supplier. There is no Chinese equivalent. There is no Russian equivalent. ASML has a 100% monopoly on the technology required to manufacture cutting-edge semiconductors.Here's what that means in practice: if you want to build a fab that makes advanced AI chips, you need EUV machines. If ASML won't sell you EUV machines, you cannot build that fab. Period. The laws of physics do not care about your geopolitical ambitions.China cannot buy EUV machines. The US government pressured the Dutch to block sales in 2019, and that ban remains in effect. SMIC, China's national champion chipmaker, is stuck using older Deep Ultraviolet (DUV) technology, which requires a workaround called "multi-patterning," exposing each layer three or four times instead of once. This works, technically, but with devastating consequences for yield and cost.TSMC's yield rate for mature 7nm processes exceeds 90%. SMIC's yield rate for 7nm using DUV multi-patterning is reportedly around 20%. That means 80% of SMIC's wafers are waste, scrapped silicon that cost money to process but produced nothing usable. The cost per working chip is roughly 4-5x what TSMC charges.This is the physics of the chokepoint. China can announce "breakthroughs" and produce sample chips for propaganda purposes. They cannot achieve the economies of scale that make AI affordable. Every H100 equivalent they manufacture costs them vastly more than it costs Nvidia. That's not a gap that closes with determination; it's a gap that widens with each process node shrink.The Cloud DenominatorEven if you could somehow build competitive AI chips outside the US-aligned ecosystem, you'd face another bottleneck: where do you train your models?Training frontier AI models requires massive clusters of GPUs, typically 10,000 to 100,000 or more H100s running in parallel for weeks or months. Very few organizations can afford to own that hardware outright. Most rent it from cloud providers.AWS, Azure, and Google Cloud control the overwhelming majority of this market. They're American companies with American headquarters paying American taxes (sort of) and billing in American dollars. When a German AI startup trains a model, the compute cost shows up as a dollar-denominated invoice to an American corporation. When a Japanese pharmaceutical company uses AI for drug discovery, the cloud bill is in dollars. When an Indian fintech deploys fraud detection, dollars.The alternatives exist but are marginal. Alibaba Cloud and Tencent Cloud are competitive in China but have minimal presence elsewhere, partly due to security concerns but mainly because their technology trails the American hyperscalers. European cloud providers (OVH, Deutsche Telekom's Open Telekom Cloud) are subscale and not competitive for AI workloads.This means that the global AI economy, the entire emerging industry that will reshape every sector from healthcare to finance to manufacturing, has the US dollar as its unit of account. Not by treaty or negotiation, but by infrastructure. The pipes carry dollars.The Export Control WeaponIn October 2022, the Biden administration did something unprecedented: it banned the export of advanced AI chips to China. Not just finished chips but the equipment to make them. The goal, explicitly stated by national security advisor Jake Sullivan, was to maintain "as large a lead as possible."This wasn't trade policy. It was technology warfare. The US was using its chokepoint position to actively prevent a rival from developing AI capability. And it worked, at least partially. China's AI labs scrambled. Smuggling networks emerged. Prices for grey-market Nvidia cards in China reportedly reached 2-3x list price.The Trump administration has modified but not abandoned this approach. In January 2025, Commerce rescinded a pending Biden rule that would have required strict compliance reporting for AI chip exports to third parties (like the UAE). But the core restrictions remain. H100s can't be legally shipped to China. ASML can't sell EUV machines to Chinese fabs.What's emerged is a strategy officials call "Market Share Weaponization." The idea: rather than trying to totally deny China access to AI chips (which just accelerates their indigenous development), allow export of slightly degraded chips (the H20, possibly the H200) to keep Chinese labs dependent on American silicon. Maintain the leverage. Keep the kill switch.This is the compute equivalent of the oil weapon. The US isn't just passively benefiting from dollar-denominated cloud infrastructure; it's actively using chip controls to enforce technological hierarchy. Countries that cooperate with American policy get access. Countries that don't, don't.TSMC: The Crown JewelTaiwan Semiconductor Manufacturing Company is, by some measures, the most strategically important company on Earth. It fabricates roughly 90% of the world's most advanced chips. Apple's iPhone processors, Nvidia's AI GPUs, AMD's data center chips, Qualcomm's mobile chips, all made in TSMC fabs in Taiwan.This creates an interesting alignment: the company that controls the most critical chokepoint in the AI supply chain is located on an island that the United States has committed to defend. Taiwan's "silicon shield" means that American and Taiwanese interests are fused at the technological level, not by treaty but by shared dependence on TSMC remaining operational and un-conquered.TSMC is building fabs in Arizona, but those facilities won't reach Taiwan's capacity or capability for years. The US is trying to diversify the supply chain through CHIPS Act subsidies, but the expertise, the supplier ecosystem, and the institutional knowledge are concentrated in Hsinchu. For at least the next decade, Taiwan remains the single point of failure for advanced semiconductor manufacturing.Which raises a question we'll address in Part 4: if TSMC is so critical, what happens if China tries to take it?The ReframeHere's the key insight that ties this all together: the correct framing isn't "AI is the new oil." Oil was a commodity. Anyone could produce it with the right geology and equipment. The OPEC nations weren't technologically advanced; they just happened to sit on reserves.AI is different. The correct framing is: "Compute is the new oil, and AI is why everyone needs compute."Compute isn't evenly distributed. It's concentrated in a few dozen hyperscale datacenters owned by a handful of American corporations, powered by chips designed by American companies, fabricated by American allies using equipment that only one Dutch company can build. The supply chain is deep, complex, and almost entirely aligned with US interests.Oil gave the petrodollar its power because energy was the universal input to industrial economies. Compute is becoming the universal input to knowledge economies. And unlike oil, which could be drilled in Nigeria or Venezuela or Iran, advanced compute cannot be manufactured outside the US-aligned ecosystem. Not today. Not for years. Maybe not ever, if the lead keeps widening.What This Means for the DollarIf the AI Dollar thesis is correct, then the conventional dollar doom narrative is missing the forest for the trees.Yes, the US is running unsustainable deficits. Yes, the debt-to-GDP ratio is alarming. Yes, BRICS nations are diversifying reserves. All true. But none of it matters if the dollar remains the currency of compute.Think about it this way: even if China successfully de-dollarized its oil imports, it would still need to pay in dollars for every cloud computing instance it rents from AWS. Even if Saudi Arabia prices some oil in yuan, every AI startup in Riyadh is still paying Nvidia in dollars. The petrodollar can weaken while the AI Dollar strengthens.This is what I mean by American hegemony being rebuilt on silicon foundations. The old architecture (oil, treasury purchases, military bases) is aging. The new architecture (chips, cloud, AI chokepoints) is being constructed. The transition isn't from hegemony to decline; it's from one form of hegemony to another.If the US plays this correctly, maintaining chip leadership, winning the AI talent war, keeping allies like Taiwan, Japan, and the Netherlands aligned, then dollar dominance extends for another generation. Not because America deserves it or because the fiscal situation is sustainable, but because there's literally no alternative infrastructure.That's the confidence trick's new hardware. In Part 3, we'll examine the rival who's supposed to replace the US: China. And we'll see why the "rising power" narrative is built on foundations of silicon sand.NotesNotes[1] Petrodollar history from multiple sources including "The Hidden Hand of American Hegemony: Petrodollar Recycling and International Markets" by David Spiro.[2] Saudi yuan oil sales reported by Wall Street Journal and subsequent confirmations in 2024-2025.[3] Nvidia market share data from company filings and Mercury Research. AMD's datacenter GPU share remains in low double digits.[4] TSMC advanced node dominance from company investor presentations and industry analysis by TrendForce.[5] ASML monopoly: The company's 2024 annual report confirms 100% market share in EUV lithography systems. No other company has successfully commercialized EUV.[6] SMIC yield estimates from "Pushing the Limits: Huawei's AI Chip Tests U.S. Export Controls", Georgetown CSET.[7] Chinese AI researcher retention: Carnegie Endowment report "Have Top Chinese AI Researchers Stayed in the United States?" finding 87% retention rate.[8] October 2022 export controls: Bureau of Industry and Security final rule, "Implementation of Additional Export Controls on Certain Advanced Computing Items."[9] H20/H200 licensing oscillation from Bureau of Industry and Security announcements and industry reporting throughout 2025.[10] TSMC Arizona status from company announcements and CHIPS Act implementation reporting. First Arizona fab began limited production in 2025 but at older process nodes than Taiwan facilities. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit tatsuikeda.substack.com/subscribe