Chapter 2.12 — Geopolitics, Policy & Risk

Every layer in this atlas operates inside rules written by governments, and this chapter is about those rules and the risks they create. A major recent development is that, for a strategic slice of demand, Beijing is using procurement and market access to force domestic adoption while the United States combines restrictions with negotiated licensing and fees. The policy regime is increasingly transactional, raising political volatility around every name in the industry. Underneath it sits the risk that can be sized but not readily hedged: Taiwan.

The AI industry is not a free market; it is a managed one, shaped at every layer by export controls, tariffs, industrial policy, and the slow split of one global supply chain into two rival blocs. Understanding this layer means holding several things at once: the whipsaw of US chip policy, the localization drive on the Chinese side, the sovereign-AI demand pool now being courted by both, and the concentration risks, Taiwan above all, that a single event could turn into a crisis. This chapter draws together threads from every earlier one, because policy is the force acting on all of them.

The controls inverted

For years the story was American denial: Washington restricting what China could buy. That story has flipped.

Nvidia's share of the Chinese AI-chip market fell from around 95% in 2022 to effectively zero by 2026, but the final step was taken by Beijing, not Washington. After the US banned the high-end H20 in early 2025, then reopened exports under an unprecedented arrangement that routed 15% of the revenue to the US Treasury, and finally approved sales of the more capable H200 to a handful of Chinese firms, it was China's own regulators who told domestic companies not to buy American chips and barred foreign accelerators from state-funded data centers. The result is that the binding constraint on Nvidia in China is now Chinese policy, and the genuinely unresolved question is who is restricting whom. The mandate regime behind this, the "autonomous and controllable" (自主可控) localization drive, is detailed in its own deep-dive; the point here is that it has made China's demand for domestic chips a matter of law rather than competitiveness.¹²

The policy tools have also changed character, from a rules-based system of published controls to a series of discretionary, case-by-case deals. The timeline of the past eighteen months shows the whipsaw plainly.

Policy, Trade, and Regulatory Instruments

The government hand now reaches into the industry through a recognizable set of levers, each with its own status and its own expiry.

Two stacks, and the fight for the third bloc

The cumulative effect of controls and localization is bifurcation: the emergence of two largely separate technology stacks, an American one built on Nvidia and CUDA and a Chinese one built on Huawei's Ascend and CANN, each with its own hardware, software, and increasingly its own standards, right down to rival AI-governance forums. This is now the base case rather than a risk, and the important contest has moved to the middle. The prize is the "third bloc", the Gulf states, Southeast Asia, and the Global South, that belong to neither camp yet and whose choice of stack is up for grabs. The United States is using chip diplomacy to court them, while China offers subsidized, string-free Ascend systems to buyers who cannot or will not access US chips. The decisive question of the whole geopolitical layer has become which stack the uncommitted world adopts.

Sovereign AI: the new demand pool

A distinct new source of demand has emerged from all this: sovereign AI, national compute funded by governments, projected above $100B in 2026. It is anchored by US chip diplomacy. Following a 2025 Gulf tour, Washington authorized the UAE's G42 and Saudi Arabia's HUMAIN to buy Nvidia Blackwell; the Stargate UAE campus (with G42, Nvidia, OpenAI, Oracle, and Cisco) targets 5 GW with a first 200 MW cluster live in 2026, and HUMAIN is targeting up to 600,000 Nvidia GPUs over three years. Europe is backing up to five "AI gigafactories" with around €20B, India and Japan are funding national models and compute, and essentially every major economy now has a program. Sovereign AI is the demand pool that most offsets Nvidia's lost China revenue, but it is lower-quality demand than the hyperscalers': it is policy-driven, oil-price-sensitive, and comes with security strings (no Huawei, US-vetted operators), so it deserves more skepticism than a hyperscaler backlog, and the Gulf is explicitly hedging by courting China as well.

Tariffs and industrial policy

The domestic-policy tool has shifted from grants to tariffs and equity. The Section 232 semiconductor tariff landed at 25% on certain advanced chips in January 2026, narrower than the "100%" once threatened, which functions mainly as a lever to extract US-investment pledges. The marquee industrial-policy move is the US government taking a roughly 10% equity stake in Intel, converting unpaid CHIPS Act grants into ownership, a template that could extend to other recipients. And allied coordination has deepened, with Washington pressing the Netherlands (ASML) and Japan (Tokyo Electron) to restrict not just sales but the servicing of installed tools in China, though the allies push back where their own revenue is at stake. The through-line is a move from carrots to sticks and ownership, and it taxes and reshapes the whole build.

The Taiwan tail, and the risk register

Above every specific policy sits the structural risk that Chapter 2.8 introduced: the concentration of leading-edge manufacturing in Taiwan. It is the one exposure that cannot be diversified away before roughly 2030, and a disruption in the Taiwan Strait is the single event that would invalidate the entire AI build at once, upstream of Nvidia, Apple, AMD, and Broadcom alike. It can be sized and respected, but not hedged. Around it sit the more manageable risks this chapter has cataloged: the possibility that export controls escalate and cut the toolmakers' China revenue, that the minerals truce lapses and re-weaponizes rare earths, that tariffs raise the cost of the whole build, and that the transactional policy regime produces sudden, headline-driven swings in individual names. These are the risks to price into any position in the industry.

Licences, procurement rules and the Taiwan clock

The near-term calendar is unusually concrete. Two dates in late 2026 matter most: the scheduled snapback of the US "50% affiliates" rule extending controls to majority-owned subsidiaries, and the expiry of the US-China minerals truce, either of which could re-tighten the screws. Beyond those, watch whether a formal replacement for the rescinded chip-diffusion framework ever lands or whether case-by-case licensing becomes permanent; whether any Blackwell-class chip is approved for China; whether the Gulf and Southeast Asian sovereign projects energize on their US-supplied hardware or defect toward Ascend; and, above all, the temperature in the Taiwan Strait, which no calendar can predict.

Positioning for a transactional regime

Policy is a lens on the other chapters more than a set of standalone trades. For Nvidia (NVDA) and AMD (AMD), the read is that China is now optionality rather than a base case, already written toward zero, while sovereign AI is the offsetting demand to watch. The ex-China rare-earth names, MP Materials (MP) and Lynas (LYC.AX), are direct expressions of the minerals leverage. Trade-compliance and reshoring beneficiaries, including the US-footprint expanders of Chapters 2.8 and 2.9, gain from the complexity itself. And TSMC (TSM) carries the Taiwan tail that sits behind the whole complex. The overarching discipline is that the transactional, deal-driven policy regime makes the industry more headline-sensitive than its fundamentals alone would imply, so position sizing and a tolerance for policy whipsaw matter as much as stock selection.

The event that overrides everything

One risk dominates all others: a Taiwan Strait crisis would break not just this chapter's thesis but the entire atlas, and it is the reason to hold any AI exposure with humility about tail risk. Short of that, a sharp escalation of export controls would hurt the toolmakers and Nvidia's residual China business, while a durable détente would relieve both. A lapse of the minerals truce would spike input costs across defense and semiconductors. And the bifurcation thesis itself would soften if the two stacks found reasons to re-converge, or harden further if the third bloc splits cleanly between them, which is the direction the evidence currently points.

For the physical U.S.–China system comparison, including what is confirmed, reported, inferred or still roadmap-only, see §2.13.

Sources

Linked evidence for this chapter's figures and load-bearing claims: ^{1 2}