Knowledge Capital, covered in Letter 171, determines who can invent. AI Capital determines who can compute — and in 2026, that access runs through a US export-control regime that looks nothing like the clean tiered system most coverage still describes. The tiers were real. They were also scrapped before a single shipment was ever governed by them.
Every government building a national AI strategy in 2026 describes it using the language of sovereignty — sovereign compute, sovereign models, sovereign data. The infrastructure data underneath that language tells a more concentrated, and more improvised, story than the phrase suggests.
Global spending on sovereign AI systems is projected to cross $100 billion in 2026, spanning data centers, GPU clusters, and national compute-access programs, which together make up 59% of all tracked sovereign AI projects worldwide. But of the more than 130 national infrastructure and model projects tracked by CNAS's Sovereign AI Index, roughly 70% involve at least one foreign technology partner — and four-fifths of those involve a US company specifically. NVIDIA alone supplies the GPUs underpinning 52% of every tracked sovereign AI infrastructure project on the planet, holding an estimated 85% share of the global AI GPU market.
The capital scale involved makes the concentration starker still. Combined 2026 AI infrastructure capital expenditure across just four hyperscalers — Google, Amazon, Meta, and Microsoft — is running at approximately $725 billion, up 77% from $410 billion in 2025, a figure that on its own exceeds the entire GDP of most individual European countries.
Most 2026 commentary on AI compute access still describes the world in terms of a clean three-tier export-control system: trusted allies with unrestricted access, a middle tier facing caps and licensing, and a fully embargoed bottom tier. That system is real, but it is also dead. The Biden administration's Framework for Artificial Intelligence Diffusion, published January 15, 2025, would have sorted the world into exactly that structure — 18 close allies in Tier 1 with unrestricted access, most other countries (India included) in a capped, license-based Tier 2, and arms-embargoed states such as China, Russia, and North Korea locked out entirely in Tier 3. The rule was rescinded by the incoming Trump administration on May 12, 2025, three days before its compliance date, and never took effect. No shipment was ever governed by it.
What has actually governed chip exports since is narrower, more transactional, and negotiated country by country rather than assigned by tier. For China specifically, a BIS rule effective January 15, 2026 shifted the review posture for commercially available H200- and MI325X-class chips (and anything less advanced) from a presumption of denial to case-by-case review, contingent on strict security and end-user certifications — while frontier Blackwell-class chips remain blocked outright. That followed an August 2025 arrangement in which NVIDIA and AMD agreed to pay the US government 15% of their China AI chip revenue in exchange for export licenses on the H20 and MI308. A 25% tariff on covered advanced chips took effect the same week as the China licensing shift. For India, there is no equivalent tiered classification at all — its access runs through a US-India interim trade framework announced in February 2026 that includes specific language protecting chip access, alongside ordinary commercial licensing channels. Under that arrangement, India's IndiaAI Mission has secured roughly 38,000 GPUs — NVIDIA H100s, H200s, and newer Blackwell-generation cards — none of them domestically designed or manufactured.
The tier system most people picture when they think about AI chip export controls existed for exactly four months, on paper, and governed zero transactions before it was scrapped. What replaced it is a set of bilateral deals and case-by-case licensing decisions — harder to summarize in a chart, and considerably easier for any single administration to rewrite on short notice.
The complication cuts against the countries that appear best-positioned. Building "sovereign" AI infrastructure on a foreign hardware and software stack — even one with negotiated, protected access — recreates the exact dependency the sovereignty framing was meant to escape: hardware refresh cycles, licensing terms, and the underlying legal authority all remain externally set, and revocable. That authority is not abstract: the same Commerce Department mechanism that cut China off from advanced NVIDIA chips starting in October 2022 is the one India's compute buildout now depends on staying favorably disposed. A bilateral trade-framework protection, negotiated in February 2026, does not carry the durability of a published regulation — it can be renegotiated, narrowed, or lapse at the next round of trade talks in a way a codified rule cannot. Europe's own build-out illustrates a parallel trap: despite genuine infrastructure like Germany's 10,000-GPU Industrial AI Cloud, its combined capital capacity still cannot realistically match the $725 billion in 2026 hyperscaler spending it is implicitly competing against, meaning GPU-as-a-service arrangements reinforce structural dependency even as they expand nominal access.
For a country with no bilateral protection and no Tier 1-equivalent standing under any current framework, the risk is not marginal disadvantage but exclusion from frontier compute entirely — which is precisely the dynamic sovereign AI programs were funded to prevent, and which a rescinded rule and a patchwork of bilateral deals have not durably solved.
AI Capital is downstream of both Energy Capital (169) — a sovereign AI program without firm power is not viable regardless of chip access — and Knowledge Capital (171), since the research base a nation holds increasingly determines which parts of the AI stack it can build domestically rather than license. The next letter turns to the capital that ultimately builds, operates, and maintains all of the above: Human Capital.
The tiered global export-control framework widely assumed to govern AI chip access was rescinded before it ever took effect, and even on paper it never gave India the unrestricted status often attributed to it. What actually governs access in 2026 is a set of narrower, bilateral, and reversible arrangements — case-by-case licensing for China's older-generation chips, a trade-framework protection for India's newer-generation access — neither of which carries the durability of the codified rule most sovereign AI narratives still assume is in place. The framework: before underwriting any country's AI Capital position, check whether its compute access rests on a published regulation or a negotiated arrangement that could unwind at the next trade round — the two carry very different risk profiles, and 2026's coverage frequently conflates them.