Part III — The United States, China, and the Dependencies Neither Can Escape

National rankings conceal the decisive question: if each system tried to commission one more AI factory, which missing input would stop it first? The answer reveals what each country must buy from abroad, subsidize at home, or replace at great cost.

3.1 If each system built one more AI factory, where would it stop?

The United States would find world-leading chips, software, and capital before it found a timely grid connection and fully domestic manufacturing chain. China would find power, construction capacity, and policy-backed demand before it found enough leading-edge fabrication, HBM, and mature software. Allied manufacturing economies would control some of the hardest components without owning the end customer. Third markets would bring capital and demand while choosing whose standards receive the next learning curve.

This is more useful than declaring a national winner. External dependency creates orders, bargaining power, and failure points—the things an investor can actually monitor.

The comparison in this part treats the world as four operating systems rather than two self-sufficient countries. The United States controls much of the demand, software, design, and private capital. China controls a large domestic market, deployment speed, power-system construction, and important processed materials. Allied manufacturing economies—especially Taiwan, Korea, Japan, and the Netherlands—control several inputs neither superpower can replace quickly. Third markets supply new customers, sovereign capital, energy, and the next installed base of technical standards.

The useful thought experiment is a marginal build: if each system tried to add one more 100 MW of commissioned AI capacity using its current suppliers, which approval, component, or financing step would be longest? This avoids comparing an American rack with a Chinese multi-cabinet system or a national power total with a local grid connection. It also connects national advantage to supplier revenue. The missing input is where the next order, subsidy, or redesign is likely to appear.

The answer is not permanent. The United States can invest in transmission and domestic manufacturing. China can improve yield, memory, and software. Allied suppliers can add overseas plants. Third markets can change procurement rules. The investment map comes from the speed and cost of those responses, not from a static national score.

3.2 The United States: strongest at the top, dependent below

Software, design, and capital can scale quickly

Read from a distance, the American position in AI has a clear and consistent shape: it leads the top of the stack and has offshored the bottom, and its binding constraint has moved from technology to physics and finance. The United States owns the layers where value and margin concentrate — chip design, cloud, models, applications, and the software that ties them together — and it owns the capital markets that fund the whole build. What it does not own is the making. The leading-edge fabrication sits in Taiwan, the one lithography machine that makes it possible sits in the Netherlands, the memory sits in Korea, and the raw materials trace back to China. America leads the parts of the industry that can be done anywhere and has rented out the parts that can only be done in a few places. That is a comfortable position right up until one of those places is disrupted.

Manufacturing and local power cannot

US strength varies by layer. It leads closed frontier models and enterprise applications, while Chinese suppliers are more prominent in open-weight models and developer-token volume. US companies dominate cloud software, merchant accelerators, interconnect, and EDA through Nvidia, Broadcom, AMD, Synopsys, and Cadence. Leading-edge fabrication, lithography, HBM, and important materials instead depend on TSMC, ASML, Korean memory suppliers, and Japanese materials companies. At the physical base, US projects face tighter power delivery than many Chinese projects. The result is a strong design and software position that still relies on allied manufacturing and constrained domestic infrastructure.

The binding constraint moved from chips to commissioned power

The American system's binding constraint in 2026 has broadened from chips to available, deliverable electricity. The country can design and buy more compute than it can power. Grid-interconnection queues stretch past eight years in the busiest regions, capacity prices have risen roughly tenfold in two years, and the transformers and gas turbines needed to energize a site carry multi-year waits (Chapter 2.3). This is a physical constraint that money alone cannot quickly solve: it is the consequence of long-running under-investment in grid capacity meeting a sudden, enormous new load. The country that leads the world in designing AI compute cannot, for the moment, plug all of it in.

Financing is an advantage and a transmission mechanism

Layered on top of the physical constraint is a financial one. The build is being funded increasingly by debt and by circular arrangements in which the industry finances its own demand (Chapter 2.11). Capital spending now exceeds the operating cash flow of the companies doing it; the largest names issued record bonds and turned to private credit and GPU-backed securitization; and the stock market that owns all of it has concentrated to a degree not seen in a generation, with the top ten companies about 40% of the index. Both the BIS and the IMF have named this a financial-stability concern. The American system, in other words, has moved its risk from "will the technology work" — a question it has largely answered — to "will the financing hold," which it has not. This is a distinctly American fragility, because the circular loops, the private megarounds, the crowded mega-cap trade, and the levered credit are all features of the US capital market, not the Chinese one.

Two dependencies cannot be diversified away quickly

Two exposures sit outside the American system's control and cannot be hedged, only sized. The first is Taiwan, where the leading-edge manufacturing that the entire build depends on is concentrated ninety miles from China; TSMC's Arizona diversification is real but back-loaded past 2030, and a Strait disruption would halt the whole complex at once (Chapters 2.8, 2.12). The second is China-controlled materials, the rare earths, gallium, and graphite that Beijing has repeatedly weaponized and that the West is years from replacing (Chapter 2.10). America's answer to the first is chip diplomacy and slow reshoring; its answer to the second is a policy-driven scramble to stand up MP Materials and Lynas. Neither answer removes the dependency this decade.

Transactional policy makes timing less predictable

The American state has shifted from a rules-based posture of denying China chips to a transactional one of selling and taxing them (Chapter 2.12): a 15% revenue-share tax on exports, a 25% tariff on advanced chips, a roughly 10% equity stake in Intel, and government-to-government GPU deals that route Nvidia hardware to sovereign-AI projects in the Gulf. The logic is to keep the world hooked on the American stack, offset the lost China revenue with sovereign demand, and use market access as leverage. It makes the industry more headline-sensitive and more politically entangled than its fundamentals alone would suggest.

The American system balance sheet

The United States as a system is therefore technologically dominant, physically power-constrained, financially reflexive, and geopolitically exposed at two undiversifiable points. Its strengths are the deepest capital markets on earth, the best chip and model designers, the CUDA and cloud moats, and the ability to attract the world's compute demand. Its vulnerabilities are the grid, the leverage in the financing, and the Taiwan-and-materials dependencies. For an investor, owning "America's AI" means owning the top-of-stack designers and the software moats for their quality, owning the domestic power-and-electrical complex for the constraint, and respecting that the whole edifice rests on a Taiwanese foundation and a Chinese mineral base that no American company controls.

3.3 China: deployment strength beneath a manufacturing ceiling

Policy demand creates an installed base

China is the mirror image. It was denied the top of the stack and is being forced to build the base it was cut off from, and its constraint is the reverse of America's: it has the power and the policy will but lacks the leading-edge manufacturing and memory. Where the American system rents the base it cannot make, the Chinese system is building the base it was refused, at subsidized cost and behind a wall of mandates, while remaining throttled at exactly two points.

Localization compounds through procurement and learning

State procurement and localization policy create a protected demand pool in China. Beijing has barred foreign accelerators from state-funded data centers, required domestic-chip quotas, and used the autonomous and controllable (自主可控, zìzhǔ kěkòng) and IT-application-innovation (信创, Xinchuang) regimes across government, state enterprise, finance, telecom, and military buyers. Inclusion on the procurement catalogue carries roughly a 20% price preference in tenders, and central enterprises are expected to replace most foreign IT by 2027. These rules support Huawei and Cambricon even when their products are less competitive in an open tender, while the resulting orders fund further domestic capacity and product development.

Domestic content ranges from branding to genuine control

A crucial subtlety that the headline localization numbers hide is that not all "domestic" (国产, guóchǎn) is equal, and the strictest state buyers know it. There is a hierarchy. At the top sit the truly self-controlled designs that own their instruction-set architecture and owe no foreign license: Loongson's LoongArch, and the self-designed accelerator architectures of Huawei's Ascend and Cambricon. Below them sit the chips that are "domestic" only in the sense of being made by a Chinese company while still resting on a foreign instruction set — Hygon's x86 (from a frozen AMD license) and the ARM-based designs of Phytium and Huawei's Kunpeng. For mainstream government procurement these licensed parts qualify, but for the defence and secrecy tiers they are discounted as not "truly" self-controllable, and the 2026 catalogue explicitly set out to purge "pseudo-Xinchuang" (伪信创) products. For an investor the lesson is to ask how domestic a champion really is, because the market sometimes prices "domestic" (国产) exposure without distinguishing the real from the nominal.

Advanced fabrication and HBM form a joint ceiling

The Chinese flywheel spins against two hard limits that policy cannot wish away. The first is fabrication: SMIC produces 7nm-class chips using older deep-ultraviolet tools and multi-patterning, but at yields estimated in the 20–30% range, and pushing below 7nm on purely domestic equipment is judged unlikely before the end of the decade (Chapter 2.8). The second is high-bandwidth memory: China trails badly in HBM specifically, and Huawei's push to build its own is an attempt to close a gap it has not yet closed (Chapter 2.7). These two ceilings are the reason a mandated market is not the same as a competitive one: Ascend and Cambricon can only ship in the volumes SMIC can yield and the memory China can supply, and their headline performance claims rest on using several times the chips and power of a Western rack. Everything else in the Chinese stack is strong or fast-closing; these two points are where it is genuinely stuck.

Power deployment is an advantage; utilization is the missing denominator

Where the American system is power-constrained, the Chinese one is power-abundant, and this is China's single clearest structural advantage in the whole contest (Chapter 2.3). China carries well over twice the installed generating capacity of the United States, adds more new capacity in a year than most countries operate in total, and can build transmission on a timescale American permitting cannot match. Its "east-data, west-compute" program (东数西算) deliberately routes data centers to the western interior where hydro, wind, and coal are cheap, and runs the fiber back east. The thing throttling the American build is close to a non-issue for the Chinese one, which is why the power trade is an American trade and the Chinese AI thesis lives in compute and localization instead.

Investors can own far less than the industrial map suggests

The Chinese system poses a problem the American one does not: much of it is difficult or impossible for a foreign investor to own, and the "market cap" of some of it does not mean what it appears to. The categories matter.

The practical consequence is that a foreign investor's China-AI universe is far smaller than the industry itself. The best assets (Huawei, DeepSeek) are un-buyable; the mandated chip champions are STAR-listed and often institutional-only; and one sanction, NS-CMIC, bars ownership outright for US persons. Screen for access before thesis.

Disclosure quality must be part of the thesis

Beyond access, the Chinese system carries a set of company-level risks that the American one largely does not, and reading them is a skill of its own. A-shares are heavily retail-driven, so many thematic names trade as 概念股 — "concept stocks" riding a story with little real exposure, at extreme multiples that reflect positioning, not fundamentals. Watch for pledged shares (股权质押), where founders have borrowed against their stock and can be forced to sell; for government subsidies (政府补助) inflating reported profit, which means reading the non-GAAP (扣非) line; for the overhang of the national "Big Fund" selling down its stakes (大基金减持); and for fabricated order rumors (小作文) that move retail-driven names. None of these has an American analog of the same character. The Chinese system's risks are political and structural where the American system's are financial and concentrated.

The Chinese system balance sheet

China has ample power in several regions, policy-supported demand, competitive open-weight models, and a growing domestic toolchain. Its main physical limits remain leading-edge fabrication and high-bandwidth memory, where foreign dependence persists. The same policy system that supports demand also concentrates risk in procurement rules, state funding, market access, and retail sentiment. Security analysis must therefore separate genuine domestic capability from nominal localization and account for ownership and policy constraints.

3.4 Allied manufacturing holds delivery power over both systems

The United States and China are not self-contained systems. Between them sits an allied manufacturing system whose members possess the least-substitutable capabilities in the stack: Taiwan in leading-edge foundry and advanced packaging, the Netherlands in EUV lithography, Korea in HBM, and Japan in substrates, wafers, photoresist and precision materials. Calling these countries “neutral chokepoints” understates their agency. Their companies allocate capacity, qualify customers, comply with national policy, choose where to build the next factory and determine how quickly either superpower can reduce its dependence.

The system's strength comes from accumulated process knowledge rather than raw scale alone. TSMC's yield learning, ASML's installed base and service network, Korean memory process integration, and Japanese materials qualification are the product of decades of customer feedback and manufacturing iteration. A subsidy can finance a new factory; it cannot instantly reproduce that learning curve. The resulting scarcity generally has a longer half-life than assembly capacity or model leadership.^{1 2 3}

Its weakness is geographic and political concentration. Taiwan remains the largest common exposure. Korea sits within reach of regional military escalation. ASML and Japanese suppliers must translate US controls into their own commercial policy while retaining economically meaningful China businesses. These companies can benefit from both stacks' demand, but they are not politically detached toll roads.

For equity investors, the allied system requires two separate judgments. The first is strategic indispensability. The second is security-level capture: valuation, China revenue, parent-company dilution, currency, cyclicality and location risk. The atlas therefore treats allied chokepoints as high-quality strategic assets, not automatically as low-risk securities.

3.5 Third markets determine which standards gain global learning

The countries outside the two manufacturing cores form a demand and capital system of their own. Gulf sovereign funds can finance multi-gigawatt campuses and use procurement to secure access to the US-led stack. India combines sovereign-compute policy, a large developer base and a growing data-center market. Southeast Asia offers land, power and regional cloud capacity while navigating US technology controls and Chinese commercial ties. Europe supplies industrial technology and regulation but is constrained by power, permitting and a fragmented capital market. Israel contributes an unusually dense design, networking, cybersecurity and application-startup ecosystem.

This is not a single political bloc. Its members optimize among access to frontier chips, financing, energy cost, data sovereignty, export-control exposure and local industrial development. Their choices determine standards and installed base. A country that builds on CUDA, US cloud and Western networking reinforces the American ecosystem; one that adopts subsidized Chinese systems creates a service, software and replacement market for the Chinese stack.

The Gulf is the clearest capital-led case. Saudi Arabia's HUMAIN and related sovereign partnerships are designed to combine power, capital, models and data-center construction at national scale.⁴ India and Southeast Asia are more price-sensitive: they can become important demand pools while placing greater weight on local ownership, lower-cost systems and sovereign control. Indonesia's announced NVIDIA-linked sovereign AI factory is one example of this contest becoming physical infrastructure.⁵

For investors, “sovereign AI” should be decomposed into funded phases, counterparties, energization dates and equipment orders. Announced national capacity is not commissioned demand. The signals that matter are utility agreements, financing close, named system suppliers, construction progress and recurring cloud utilization.

3.6 Put the four systems on one operating scorecard

The scorecard

Put the two systems side by side and the picture is one of near-perfect complementarity. Each is strong exactly where the other is weak.

Which system controls each layer

The systems remain complementary in several critical areas:

The allied lead concentrates in the compute-and-software layers and the neutral chokepoints; China leads at the raw base, in power and minerals; and the models and applications layers are the most contested.

Mutual dependence limits clean separation

One interdependent system is splitting into two, but neither side is independent. The United States still relies on Taiwan for leading-edge manufacturing. China cannot produce at the same frontier without restricted tools, while both systems depend on neutral chokepoints. China holds leverage in minerals; the United States and its allies hold leverage in lithography, memory, and design software. This mutual dependence discourages full separation while forcing both sides to fund redundant supply chains. China's 2025 mineral controls, which drove the antimony price up 2,600% before a truce, demonstrated the economic cost of using that leverage.

External dependencies are the investable map

Sections 3.3 and 3.4 change the bilateral scorecard in two ways. The allied system holds much of the manufacturing leverage attributed to “the US and allies,” while the third-bloc system determines a growing share of marginal demand. The strategic contest is therefore not only about which superpower has the better domestic stack. It is about whether the allied chokepoints remain aligned, and which technical and financing system new buyers install.

Several forms of bifurcation remain possible

Three trajectories frame the analysis. In a managed-interdependence base case, both stacks deepen while neutral chokepoints preserve mutual dependence and current truces hold. In a hard-decoupling case, export controls escalate and the minerals truce lapses, raising costs and stranding assets tied to cross-border supply. In the tail case, a Taiwan Strait disruption interrupts the shared manufacturing foundation and overwhelms the other scenarios.

The two systems offer mirror-image exposures

For an investor the comparison yields three durable conclusions. The allied chokepoints, which both giants depend on, offer high strategic criticality but retain valuation, cyclicality, access and geopolitical risks. Bifurcation is deepening, so the contest for the uncommitted third bloc is where marginal infrastructure growth may be decided, with sovereign AI as the vehicle. Taiwan concentration sits behind both systems at once — the event that would break the whole structure and the reason to hold any AI exposure with humility about tail risk. The United States rented the base it leads on; China is building the base it was denied; and the two remain, for now, locked together by the chokepoints neither controls.

The physical bridge from this system-level comparison to accelerator packages, complete systems and normalized 100 MW facilities is §2.13.

The part’s conclusion can be used as a four-step country screen. First, identify the layer in which demand originates. Second, trace the foreign inputs required to turn that demand into commissioned systems. Third, estimate which dependency is being reduced and which new dependency the workaround creates. Fourth, identify whether public investors can access the beneficiary without taking a larger policy, currency, or governance risk than the industrial advantage is worth.

This screen usually produces a less dramatic answer than “the US wins” or “China catches up.” The United States can retain frontier economic value while remaining physically dependent. China can build a strategically viable domestic stack while spending more power and capital per unit of useful compute. Allied suppliers can hold extraordinary bargaining power while carrying geopolitical and customer concentration. Third markets can receive subsidized infrastructure before proving utilization. Those mixed conclusions are not ambiguity; they are the investable structure.

Part III system endnotes

Sources

Linked evidence for this chapter's figures and load-bearing claims: ^{6 7 8}