Every Sovereign AI Road Leads Back to Two Companies

In late March, SK Hynix placed the largest single order for ASML lithography equipment ever publicly disclosed: $7.97 billion, roughly thirty extreme ultraviolet scanners. ^[1] The order was part of South Korea's $518 billion national mobilization to build a sovereign AI ecosystem — President Lee Jae Myung's "triple axis for a great leap forward" spanning semiconductors, physical AI, and data centers. ^[2] The largest equipment purchase in that sovereign buildout was a rent payment to the monopoly it seeks to escape. The paradox is not South Korea's alone. Over the past six months, a wave of nations has announced plans to achieve what they call AI sovereignty — the capacity to design, manufacture, and deploy artificial intelligence without depending on foreign suppliers. The strategies differ in scale and rhetoric, but they converge on a single physical reality: every path runs through two gates, and both are owned by someone else.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

The warning came from Sarvam AI, an Indian startup, as New Delhi launched its own sovereign AI push. That push was triggered directly by US export controls: when Anthropic cut off model access to foreign nationals, India's government concluded that software dependence was a national vulnerability and pivoted to hard assets. ^[3] The result was an allocation of 88,000 GPUs to Indian startups — chips designed by Nvidia and manufactured by TSMC — and a new Jabil plant to build data center components, not chips. ^[4] India is buying sovereignty with other people's silicon. The same pattern holds across the data-center tier of the escape strategy. France secured a €75 billion SoftBank commitment for AI infrastructure, framed as digital sovereignty. ^[5] But the 5 gigawatts of data center capacity SoftBank's Masayoshi Son called "probably 50x bigger than dot-com" will run on Nvidia and AMD processors manufactured by TSMC. ^[5] Palantir and NVIDIA's jointly marketed "Sovereign AI Operating System" for governments runs on NVIDIA's Blackwell chips — fabricated by TSMC, on ASML machines. ^[6] The product is sovereign; the supply chain is not. The second tier — foundry sovereignty — aims deeper. Nations in this camp are not merely deploying foreign chips in domestic data centers; they are building the factories to make the chips themselves. And here the dependence tightens rather than loosens. South Korea's $518 billion plan is the most ambitious. Samsung committed a record $82 billion in annual AI chip investment, expanding fabs in Pyeongtaek, Yongin, and Taylor, Texas. ^[7] SK Hynix and Samsung are racing on HBM4 and HBM4E memory for AI accelerators, targeting 80% of the global high-bandwidth memory market by 2028. ^[2][8] But HBM manufacturing requires EUV lithography, and EUV lithography means ASML. The HBM duopoly is a chokepoint nested inside the ASML chokepoint — and the $7.97 billion order is the receipt. Japan's Rapidus project — the state's flagship attempt to restore domestic advanced chip manufacturing — is building a 2-nanometer foundry with ¥250 billion in government backing. ^[9] The equipment is ASML's High-NA EUV; the process technology is licensed from IBM. Mass production is targeted for fiscal 2027 at the earliest. Japan's broader 370 trillion yen investment drive through 2040 names semiconductors a priority sector among seventeen, but the chip-specific piece — the part that would actually deliver manufacturing sovereignty — runs through the same two foreign gates as everyone else's. ^[10] The European Union has taken the most candid approach. Its €2.5 billion NanoIC pilot line — the EU's first EUV-capable facility targeting beyond 2-nanometer chips — is built around ASML's own technology. ^[11] EU Commissioner Henna Virkkunen made the logic explicit.

It’s true that we do have some of the key technologies, like ASML, that everyone is dependent [on] globally — Henna Virkkunen

The EU is not escaping the chokepoint. It is weaponizing its ownership of it — ASML is a Dutch company — and betting that controlling the gate is better than finding another door. It is a rational bet, but it is not sovereignty in any ordinary sense of the word. Then there is China, which is attempting something genuinely different. Denied access to both ASML's EUV machines and TSMC's leading-edge foundry capacity by US export controls, China cannot take either of the first two paths. It is trying to route around the chokepoints entirely. Huawei's Tau Scaling Law is the most explicit statement of this ambition. Rather than shrinking transistors — the path that requires EUV lithography — Huawei aims to optimize system-level architecture to achieve what it calls 1.4-nanometer equivalence by 2031.

Given all the various constraints, we have found some pretty good solutions... I can confidently say in the coming 10 years our solutions for mobile computing and AI computing will be competitive. — Huawei

The word "equivalence" is doing heavy lifting. Huawei's current demonstrated manufacturing capability is 7 nanometers, produced by SMIC. The gap between 7 nanometers and TSMC's 2-nanometer production — to say nothing of the 1.4-nanometer process TSMC targets for 2028 — is not a design problem; it is a physics problem that architectural cleverness can narrow but not close. ^[12] China's LineShine supercomputer offers a preview of what architectural bypass looks like in practice. In June, it reclaimed the world's fastest ranking using a CPU-only architecture with 45,000 domestic processors — a design choice made explicitly to circumvent export controls on GPU accelerators. ^[13] The system is 42% less energy-efficient than its GPU-powered competitors and ranked only fourth in AI-specific benchmarks. ^[13] The adaptation is real. So is the performance penalty. Alibaba's Zhenwu M890 AI chip, positioned as a domestic alternative to Nvidia's H100, has shipped 560,000 units. ^[14] But it is manufactured on SMIC's 7-nanometer process while TSMC's leading edge is at 2 nanometers — a design achievement that outpaces the manufacturing capability available to produce it at parity. DeepSeek is developing custom inference chips to bypass export controls, but chip design is not chip fabrication; the foundry bottleneck remains. ^[15] The strongest counter-evidence to the chokepoint thesis comes from Intel's foundry. Google ordered 3 million TPUs from Intel for 2028, Tesla is using Intel's 14A process, Nvidia is evaluating Intel's 18A, and Apple reached a preliminary agreement to shift some production. ^[16] Intel's foundry revenue reached $5.4 billion, and shares surged 13%. ^[16] This suggests the TSMC monopoly is loosening into an oligopoly — a genuine diversification of advanced manufacturing. But the Intel counter has limits. That $5.4 billion in foundry revenue is roughly one-twentieth of TSMC's annual revenue. ^[16] Customers are going to Intel primarily because TSMC's advanced packaging lines are sold out through 2027 — a capacity constraint, not a competitive loss. ^[16] Intel's 18A process remains behind TSMC's 2-nanometer node in deployment. And Intel itself depends on ASML for lithography equipment. The oligopoly, if it materializes, still runs through the same chokepoint. Meanwhile, the frontier is moving. TSMC announced A14 chip production for 2028, with 20% density improvement and 25 to 30 percent power reduction over the current generation, followed by A13 and A12 nodes in 2029. ^[17] Samsung's 2-nanometer process reached 70% yield — progress, but TSMC maintains 90% yields on its own 2-nanometer node and targets 1.4-nanometer mass production a year ahead of Samsung's delayed 2029 target. ^[18] The gap between the chaser and the chokepoint holder is a full generation and widening. And then there is the next frontier. In June, IBM unveiled a sub-1-nanometer nanostack technology that nearly doubles transistor density — a laboratory breakthrough pointing toward 0.7-nanometer chips. ^[19] The technology was developed in collaboration with ASML, Lam Research, and Tokyo Electron. IBM plans to license it to foundries rather than manufacture it.

We're not just making smaller transistors, we're reinventing how chips are built to deliver dramatically more power and energy efficiency. — IBM

The breakthrough that will define the next decade of computing was built with the chokepoint holders, not around them. It will flow to market through the same gates every nation is spending billions to escape. The harder they push, the more rent they pay to the gates they seek to leave.