The Geopolitics and Governance of the Global AI Divide
Internal synthesis covering the tripartite geopolitical architecture of the global AI ecosystem (US / China / EU), the three governance models, and the divergent optima each engineers in the AI value stack.
Type
This is a secondary-research synthesis assembled in
raw/, not a single primary paper. It braids Anu Bradford (bradford-digital-empires), Jeffrey Ding (ding-diffusion-marathon), Helen Toner, Kai-Fu Lee (lee-01ai-pivot), and the LSE Brussels Mirage legal scholarship into one frame. Primary sources are listed in Annotated source list.
Central claim
The AI value stack does not have a single global shape. The optimum is engineered exogenously by the host bloc’s capital structure, regulatory philosophy, and geopolitical constraints. See divergent-value-stack-optima.
- United States — top-heavy optimum. Value concentrates at hardware (Nvidia) + foundation-model/cloud layer (Microsoft/OpenAI, Google, AWS, Anthropic). Strategic moat: capital + compute scale. Strategy: closed-API frontier models, AGI sprint framing.
- China — bottom-heavy optimum. Value pushed downstream to application, B2B enterprise integration, Global-South infrastructure. Strategic moat: rapid diffusion + open-weight commoditization that destroys the model layer’s economic value. Forced into this shape by US chip embargoes.
- European Union — regulatory-centric optimum. Value attempted at the governance + sovereign-infrastructure layer. Strategic moat (theoretical): compliance as global standard via brussels-effect-and-mirage. Currently functions more as handicap than moat.
Four structural axes of divergence
Per Bradford (digital-empires-tripartite):
- Locus of value extraction — US: closed APIs + cloud rent. China: open-weight B2B deployment. EU: sovereign infra + compliance licensing.
- Innovation vs. diffusion (diffusion-vs-innovation) — US optimizes for frontier breakthroughs; China is pivoting to mass diffusion across industrial sectors.
- Model accessibility — US: proprietary weights as national-security IP. China: open weights as asymmetric weapon (open-weight-asymmetry) — “Android of the AI era”.
- Global export mechanism — US exports private corporate power; China exports infrastructure power (Digital Silk Road); EU exports regulatory power (Brussels Effect).
Notable empirical anchors
- DeepSeek “Sputnik moment” — V4/R1 trained on H800 chips at a fraction of Western compute cost; explicitly optimized for Huawei Ascend. Folded into open-weight-asymmetry.
- 01.AI pivot from B2C → B2B — Kai-Fu Lee’s strategic reorientation toward banking, insurance, mining, energy. See lee-01ai-pivot.
- EU AI Act backlash — open letter from European founders and VCs (Kry, 20VC, Voi, Synthesia) demanding implementation pause. Compliance overhead drains scarce VC capital. See brussels-effect-and-mirage.
- Global AI Index (Tortoise) — US + China dominate; Germany 8th, France 11th, Sweden 25th. Quantitative backbone for EU structural deficit.
- Sovereign AI footprint — Denmark (Ozempic-funded supercomputer for biotech), Italy (LLM for government workers), Sweden (rebuild). See sovereign-ai.
How this bears on the wiki’s hypotheses
H2_u-curve-of-value — major refinement. The single-global-U is wrong. There appear to be three different shapes: top-heavy (US), bottom-heavy (China), regulatory-centric (EU). The U-curve, if it holds at all, may be a Western-stack artifact. See divergent-value-stack-optima for full treatment.
H1_L0-L7-ladder — not directly challenged. The ladder is a substrate-ownership map; the geopolitical layer sits orthogonal. But L6–L7 (own model + own compute + own energy) maps neatly onto China’s domestic-substitution drive (Huawei Ascend, sovereign chips) and EU’s “AI factories” — both are L7 plays under different banners.
bear-case-synthesis — partial counterweight. zitron-circular-economics argues the US bottom of the U is propped up by recycled VC. This source independently argues that the same US bottom is propped up by an over-indexed national-security framing (Toner’s critique). Two different attacks, same target: the US infrastructure-layer rent extraction may be more fragile than it looks.
Contradictions and open questions surfaced
- Is the EU AI Act a moat or a handicap? Source argues handicap in short-to-medium term; potential moat in the long term if global enterprise demand matures around verifiable safety. Existing H2_u-curve-of-value is silent on this — needs incorporation.
- Does the “Brussels Effect” actually work for AI? The LSE Brussels Mirage paper exposes that compliance form travels but rights-protective substance does not (case study: predictive policing in India). Undermines the optimistic EU thesis.
- Is the open-weight commoditization durable, or a temporary embargo workaround? If US export controls relaxed tomorrow, would China revert to closed-weight competition? Source assumes not (state-directed strategy), but worth pressure-testing.
- What happens when blocs collide? The source treats the three optima as parallel; the open question is what happens at the seams — e.g., European startups building on Chinese open weights to escape both US API rents and EU compliance drag.
Annotated source list
The synthesis cites the following primary materials. Direct ingestion of each is the next priority for citation-readiness:
| Source | Stub | Why it matters |
|---|---|---|
| Anu Bradford — Digital Empires (Tech Policy Press, 2023) | bradford-digital-empires | Tripartite governance frame; “digital empires” terminology. |
| Jeffrey Ding — Running the Right Artificial Intelligence Race (RAND, Sept 2025) | ding-diffusion-marathon | ”Diffusion marathon” reframing of US strategy. |
| Jeffrey Ding — Explaining China’s Diffusion Deficit (FPRI, Sept 2024) | ding-diffusion-marathon | Empirical case for China’s diffusion gap. |
| Helen Toner — Lawfare Daily / 80000hours podcast | (folded into source notes) | “Steerability” and “adaptation buffers” critique of US non-proliferation. |
| Kai-Fu Lee — Capgemini interview + FT coverage (2024) | lee-01ai-pivot | 01.AI’s B2B pivot; open-source-for-Global-South thesis. |
| DeepSeek’s Sputnik Moment (Economic Times, Jan 2025) | (folded into open-weight-asymmetry) | Empirical evidence for chip-embargo blowback. |
| EU AI Act backlash (FintechWeekly, July 2025) | (folded into brussels-effect-and-mirage) | Direct market evidence of compliance drag. |
| The Brussels Effect in AI Regulation (LSE Law Review, April 2026) | (folded into brussels-effect-and-mirage) | “Brussels Mirage” — de jure vs de facto diffusion. |
| What is ‘Sovereign AI’? (WEF, Nov 2024) | (folded into sovereign-ai) | European sovereign-AI initiatives and “sovereignty traps”. |
| Global AI Index (Tortoise) | (folded into sovereign-ai and source notes) | Quantitative baseline for the tripartite gap. |
Related
- divergent-value-stack-optima — the key concept this source contributes.
- digital-empires-tripartite — the governance frame.
- diffusion-vs-innovation — Ding’s reframing.
- open-weight-asymmetry — China’s strategic move.
- brussels-effect-and-mirage — EU regulatory imperialism and its limits.
- sovereign-ai — European hardware play.
- H2_u-curve-of-value — the hypothesis this source refines.
- bear-case-synthesis — the other major synthesis page; both pressure the U-shape from different angles.