Sovereign AI and the Compute Arms Race

The AI revolution is not unfolding on a single, global playing field. It is fracturing along geopolitical fault lines that will determine which enterprises can access which compute, which models, and which markets for the next decade.

Throughout this report, we have treated the AI stack as a unified global system: NVIDIA designs chips, TSMC fabricates them, hyperscalers deploy them, and enterprises everywhere consume the resulting intelligence. That framing is increasingly incomplete. Three rounds of US export controls (October 2022, October 2023, January 2025) have systematically restricted China’s access to advanced AI chips. China has responded by investing billions in domestic alternatives, accelerating open-source model development (Chapter 15), and building parallel compute infrastructure. The EU, Middle East, and India are launching sovereign AI initiatives to reduce dependence on both US and Chinese technology stacks.

The result is an emerging bifurcation: a US-allied AI ecosystem (NVIDIA, AMD, TSMC, US hyperscalers) and a China-aligned ecosystem (Huawei Ascend, SMIC, domestic clouds). For multinational enterprises, this bifurcation creates strategic choices that cannot be deferred: which chips to deploy, which models to use, which data sovereignty rules to comply with, and which markets to serve from which infrastructure. The wrong choice locks you into a stack that may lose access to key markets. The right choice — or more precisely, the right dual-stack strategy — positions you for both ecosystems.

Impact on NVIDIA’s China Business

Before export controls, China represented approximately 25% of NVIDIA’s data centre revenue. The A100 and H100 were the standard training accelerators for Baidu, Alibaba, Tencent, and ByteDance. After Round 1, NVIDIA created the A800 and H800 with reduced interconnect bandwidth — a temporary workaround that generated billions in revenue before Round 2 closed the loophole. After Round 2, NVIDIA introduced the H20, a further downgraded chip with roughly 20% of the H100’s AI training performance but priced competitively for inference workloads.

The financial impact has been significant but not catastrophic. NVIDIA’s China data centre revenue declined from an estimated $10–12 billion annually to approximately $5–7 billion, largely through H20 sales. The company has compensated with accelerating demand from US hyperscalers and sovereign AI projects globally. But the strategic impact is deeper than revenue: every chip denied to China accelerates China’s investment in domestic alternatives.

Compute Sovereignty vs Model Sovereignty

The sovereign AI movement reveals a strategic distinction that many governments have not yet resolved: the difference between compute sovereignty (controlling the physical infrastructure — chips, data centres, power) and model sovereignty (controlling the AI models — training data, weights, deployment).

Compute sovereignty requires semiconductor fabrication capability, which only Taiwan (TSMC), South Korea (Samsung), the US (Intel), and to a limited extent China (SMIC) and the EU possess. For most countries, compute sovereignty means securing long-term supply agreements and building domestic data centres, not fabricating chips. Model sovereignty is more accessible: any country with sufficient compute (purchased or leased) and AI talent can train models on local data in local languages. The open-source ecosystem (Chapter 15) has dramatically lowered the barrier to model sovereignty — fine-tuning a Llama or Qwen base model for a specific language or domain costs orders of magnitude less than training from scratch.

The strategic implication for enterprises: compute sovereignty is a government problem; model sovereignty is an enterprise opportunity. Companies that build domain-specific, locally-compliant AI systems on top of open-source base models can serve sovereign AI markets without owning chip fabrication facilities. This is exactly the application-layer value capture thesis of this entire report, applied to geopolitics.

The arithmetic is stark. NVIDIA designs AI chips. TSMC fabricates them. Approximately 90% of chips at the most advanced process nodes (<7nm) come from TSMC’s fabs in Taiwan. The entire AI infrastructure buildout described in Chapters 16–18 — the $1.7 trillion compute stack, the 72 million H100-equivalents by 2030 — flows through TSMC’s fabrication facilities.

Taiwan sits 130 kilometres from mainland China in one of the world’s most geopolitically sensitive zones. Any disruption to TSMC’s operations — whether from military conflict, natural disaster, or political coercion — would create a global AI chip shortage that no amount of demand-side adjustment could mitigate. The AI industry has a single point of failure, and it is located in a potential conflict zone.

Diversification efforts are underway but slow. TSMC is building fabs in Arizona (first chips expected 2025–2026, full production 2028), Japan (Kumamoto, operational 2024 for mature nodes), and Germany (Dresden, planned 2027). Intel is expanding domestic US fabrication. Samsung is investing in Taylor, Texas. But advanced-node fabrication cannot be relocated quickly — a new fab takes 3–5 years from groundbreaking to production, and the most advanced processes (3nm, 2nm) remain in Taiwan for now.

Enterprises serving both US/EU and Chinese markets need a dual-stack architecture. This does not mean maintaining two completely separate AI systems. It means designing a modular architecture where:

• Application logic is stack-agnostic (the routing, workflow, and business logic runs anywhere)
• Model integration uses abstraction layers (so swapping GPT-4o for Qwen or Ernie requires configuration, not re-engineering)
• Data handling complies with local sovereignty requirements by design (Chinese data stays in China, EU data stays in EU)
• Compute is provisioned from local cloud providers (AWS in the US, Alibaba Cloud in China, local providers in regulated markets)

This architecture is precisely the intelligence routing framework described in Chapter 22, extended to include geopolitical routing. The router does not just select the optimal model for the task — it selects the optimal model that is legally available and data-compliant for the user’s jurisdiction.

What Comes Next

The geopolitical fragmentation of the AI stack creates winners and losers at the enterprise level. But it also creates winners and losers at the workforce level. Chapter 26 examines the other dimension of “what comes next”: the impact of AI on employment, wages, skills, and the very definition of work — building on the task-level analysis from Chapter 21 to map the human consequences of the disruption this report has documented.