Chapter 2.12 — Geopolitics, Policy & Risk

Policy rarely makes semiconductor demand disappear; it redirects the order, changes the eligible supplier, and alters the cost of capacity. The relevant investor question is therefore not “more controls or fewer controls?” but who receives the rerouted demand, how long the advantage lasts, and which tail risk cannot be hedged.

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.

For a company trying to sell an AI system across borders, two policy gates now matter. The first asks whether the seller is legally allowed to export the chip, equipment, software, or material. The second asks whether the buyer is legally or politically allowed to procure it. Clearing one gate does not clear the other. An American regulator may approve a restricted product for China while a Chinese procurement rule directs customers toward domestic hardware; a Chinese supplier may be able to export a system while a third-country project tied to US financing excludes it.

The distinction explains why policy analysis cannot stop at a list of bans. A rule changes the eligible product, buyer, route, and date of revenue. Demand may be destroyed, delayed, or redirected to another supplier. The affected company may redesign a lower-performance product, move production, seek a license, or abandon the market. The beneficiary may receive protected demand but also inherit higher cost, weaker technology, and an obligation to invest before the market is commercially attractive.

The investor’s unit of analysis should therefore be the rerouted order. Where did the customer originally intend to buy? Which policy blocked or conditioned that purchase? Which alternative supplier or geography receives it? What additional capital is required? And how long is the policy advantage likely to last? These questions turn geopolitical headlines into an auditable revenue path.

Restrictions have begun to redirect orders in both directions

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.

Different policy tools alter different cash flows

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

Each lever affects a different part of cash flow. An export ban can remove revenue immediately while leaving research and operating costs in place. A tariff can preserve the sale but raise the buyer’s price or compress the seller’s margin. A subsidy or government equity stake can lower financing risk while attaching domestic-production obligations. A procurement mandate can create demand for a local supplier before its product is globally competitive. A mineral restriction can raise input prices far upstream from the policy target.

Timing matters as much as direction. Orders may be accelerated before a rule begins, frozen while licenses are reviewed, or released suddenly after an exemption. Inventory can hide the impact for several quarters. A customer may qualify a domestic substitute during the delay and never return even after the rule is relaxed. Policy therefore changes both the level and the calendar of revenue, and a temporary rule can create a permanent change in supplier share.

Two stacks are competing for the same third markets

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.

Adoption happens in stages. A government may first purchase training capacity or build a national cluster. Universities and agencies then train engineers on its software. Local cloud providers expose the stack to companies. Applications, procurement rules, and technical standards grow around it. Once skills and workloads accumulate, changing the underlying system becomes much more expensive than changing the first hardware order.

That is why the third-market contest matters beyond near-term chip revenue. The winning stack receives operating data, developer attention, service relationships, and later replacement demand. The losing stack may remain technically capable but lack the installed base required to improve. A subsidized first deployment can therefore be rational if it establishes a decade of follow-on software and system dependence.

Sovereign AI creates demand before it proves utilization

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.

The difference between an allocation and a productive asset is the same as in the power chapter. The project needs a financed site, commissioned electricity, trained operators, software, workloads, and users. A government can rationally buy strategic capacity before commercial utilization exists, just as it funds defense readiness or a national laboratory. That may be valuable public policy while producing a lower financial return than a hyperscaler cluster filled by paying customers.

For suppliers, the first hardware order is real revenue. For investors, its durability depends on utilization and follow-on demand. Watch whether the country develops local models and applications, whether the cluster attracts private customers, whether operations depend indefinitely on foreign staff, and whether replacement generations are purchased after the initial diplomatic announcement.

Industrial policy buys resilience at a higher cost

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.

Resilience has an income-statement owner. Domestic capacity may require higher wages, duplicated facilities, lower initial yields, and operation below efficient scale. Governments can absorb some of the difference through grants, tax credits, guarantees, or procurement. Customers can absorb it through higher prices. Suppliers can absorb it through weaker margins. A policy announcement is not complete until the investor knows which balance sheet carries the premium.

The same point applies to China’s localization drive. Protected domestic demand can give suppliers the volume needed to learn, but customers may pay more power and capital for lower effective compute while the ecosystem matures. Strategic independence can improve rapidly even when near-term economic efficiency deteriorates. National success and minority-shareholder return must be evaluated separately.

Taiwan is the tail risk the stack cannot readily hedge

Leading-edge manufacturing remains concentrated in Taiwan, as Chapter 2.8 shows. Overseas capacity will not provide a full substitute before roughly 2030, and a Taiwan Strait disruption would affect Nvidia, Apple, AMD, Broadcom, and their customers simultaneously. Listed securities cannot provide a complete physical hedge. More limited risks include tighter export controls reducing China revenue for equipment suppliers, renewed mineral restrictions, tariffs that raise project cost, and case-by-case licensing that changes individual company access with little notice.

Licenses and procurement rules reset revenue timing

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.

Invest around rerouted demand, not permanent policy assumptions

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 every base case

A Taiwan Strait crisis would interrupt leading-edge manufacturing across both systems and is not fully hedgeable through listed securities. Less severe outcomes remain material: tighter export controls would reduce China revenue for equipment suppliers and Nvidia, while a durable détente would relieve some pressure. A lapse of the minerals truce would raise input costs across defense and semiconductors. Evidence of renewed technology exchange would weaken the bifurcation case; third-country adoption dividing cleanly between the two standards would strengthen it.

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

The practical conclusion is to treat policy as a changing state variable, not a permanent competitive moat. Build a base case from rules already in force and orders already eligible. Put licenses, elections, scheduled expiries, and negotiations into scenarios rather than silently assuming they continue. Give a policy beneficiary credit for the demand it can actually serve, then subtract the cost of localization and the possibility that protection eventually creates new competition.

For every affected security, follow four lines: lost revenue, rerouted revenue, required capital, and the date each becomes visible. Then ask whether the market price reflects a temporary headline, a multiyear installed-base shift, or the Taiwan tail that overwhelms both. Geopolitics matters because it changes who can deliver. Investment returns still depend on whether the newly eligible supplier can deliver profitably before the rule changes again.

That final timing test prevents two opposite mistakes. One is treating every restriction as permanent and capitalizing a temporary windfall forever. The other is dismissing policy-created demand because it began with a mandate. If a protected supplier uses the window to improve yield, software, service, and customer trust, a temporary rule can create a durable installed base. The evidence is not the announcement itself; it is whether the company becomes competitive enough to keep the customer when the policy window changes.

Sources

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