Every part of this series opens with a table of dated, sourced numbers and closes with an argument about what they mean. This page exists so the six parts don't have to keep stopping to explain themselves. Read it once, and you'll know what a Tally is, what game this series claims is being played, and why "one-way" is a harder claim to earn than it looks, and a more defensible one than its critics assume.
/ 01The Premise, and the Argument It Has to Survive
The claim: in 2026, the United States is a net donor to China's AI buildout. Not through any single deliberate policy: through the compounding effect of structural behaviors nobody designed as a package. A visa regime that prices out mid-career researchers with tacit knowledge no customs form can seize. Open-weight model releases that hand distillation a running head start. Inference demand quietly rerouted to cheap Chinese compute because American power costs too much. And, underneath all three, an energy-cost gap that lets one side absorb, retrain, and redeploy faster than the other.
The genuine rival hypothesis (and it is genuine, not a strawman erected to be knocked down) is that this is not one-way at all. It is the ordinary shape of a great-power rivalry: a prisoner's dilemma dressed up as generosity, a mutual hostage exchange where both sides hold real cards. Chinese-trained researchers still make up a majority of America's own AI research base. American developers distill Chinese open-weight models the same way Chinese labs distill American ones. Critical-mineral and component supply chains for AI hardware run through China, not around it. Anyone arguing "one-way" who doesn't reckon with this case honestly isn't arguing in good faith.
Worth saying plainly, because it bears on how hard this series has to argue: the brief that seeded this series was drafted with an AI research assistant, and that assistant's own conclusion (stated twice, independently) was explicitly bidirectional. Its language: the US and China are "each other's Trojan horses," locked in "a mutually assured arbitrage" where each side's strategy depends on the other's continued production of gifts. The strict one-way claim this series opens with is a deliberate sharpening made beyond what even that source material concluded. Which means the burden of proof sits squarely on this series, not on the reader's skepticism: the one-way thesis has to be earned against a mutual-arbitrage baseline that a fair-minded reader should start from, not a straw version invented to lose.1
Here is how it survives. Robert Keohane and Joseph Nye's 1977 distinction between sensitivity and vulnerability is the tool that does the work, and every part of this series leans on it, so it's worth getting right once, here.2 Sensitivity is how quickly and how sharply a change on one side produces costly effects on the other, holding existing policies constant. Vulnerability is the cost of adjusting to those effects after a state changes its own policy in response: the cost of the best available alternative. A state that can absorb a shock cheaply is not vulnerable even if it's highly sensitive. A state with no good alternative is vulnerable even if the initial shock looks modest.
Run the rival hypothesis's own best evidence through that filter and the one-way claim survives, narrowed and more precise than it started. The flows genuinely are bidirectional in sensitivity: both countries feel researcher departures, model distillation, and hardware dependency quickly and visibly. What they are not is symmetric in vulnerability: who pays more, and for how long, to adjust away from the dependency once it's exposed. That is an empirical, channel-by-channel question, not a rhetorical one, and it is the actual argument each of the next five parts is making: not "there is no exchange," but "the exchange is felt equally and paid for unequally."
/ 02Three Axes, One Direction
This series inherits its first axis from The Stack and adds two of its own. All three converge on the same vulnerability-asymmetry argument, from three different directions.
The cost structure The Stack already tore down floor by floor: who holds equity in the AI buildout, who holds the obsolescence risk, who holds the electric bill. Cross-linked here, not re-argued: the grid strain that series documented is the same strain that makes continued American "generosity" toward China's compute buildout look increasingly irrational on its own numbers.
This series' own argument, built fresh. The sequencing matters and this series corrects it explicitly: DeepSeek's January 2025 release started the open-weight commoditization race. American labs' subsequent releases are participants in that race, reactions to it, not its ideological origin. The asymmetry isn't about who invented openness. It's about who captures the distillation benefit fastest, on the cheaper power documented in Axis 3.
From Cembalest's data and Social Physics Article 7, which broke it down first. The US-China energy-cost differential is what makes the one-way transfer economically possible: the vulnerability asymmetry underneath the other two axes, expressed in dollars per megawatt-day rather than researchers or model weights.
/ 03The Fourth Thread: Alliance Erosion
Not a fourth axis in the geometric sense: a compounding throughline, and it runs through at least two of the six parts with named, dated tallies, not a gestured-at aside. While the United States structurally gifts China strategic depth across the three axes above, the same period saw Trump-led actions (tariffs, "freeloader" rhetoric, inconsistent chip-export enforcement, friction with NATO over defense costs) alienate the allies who would otherwise help counterbalance China: the European Union, Japan, South Korea, Taiwan, Australia.
This is not a symmetrical complaint about "both sides." It's a second, simultaneous policy failure, and the parts ahead document it in specifics: a Dutch government caught between US pressure to tighten ASML's chip-equipment exports and a proposed US law that would let Washington impose restrictions on an ally directly if it disagrees. A US Commerce Secretary publicly telling Samsung and SK Hynix they have "no choice" but to invest on American soil, while raising the prospect of 100% chip tariffs as leverage. A Taiwanese vice premier publicly rejecting a US-floated "50-50" chip-production split as something "our negotiation team has never... committed to." A former EU Trade Commissioner writing, in mid-2026, that Europe "can no longer rely on the erratically-led United States as an ally in dealing with Beijing."3
Squeezing the coalition that could help while enriching the competitor is the harder story this series tells, and the more honest one. It also isn't a clean case against the current administration specifically: burden-sharing pressure on allies has bipartisan American precedent going back to Kennedy, and Obama used strikingly similar "freeloader" language in 2016. What's different, and what the parts ahead argue carefully rather than assert, is the scale and simultaneity: tariff threats, defense-cost renegotiation, and named-company chip-investment coercion running against the EU, Japan, Korea, and Taiwan all at once, in an eighteen-month window, while the same administration separately eased the chip-export enforcement that the axes above say benefited China. Prior burden-sharing episodes didn't have to answer for that co-occurrence. This one does.
/ 04Which Game Is This?
Named up front, so no part has to stop and explain it: this series argues the US-China AI relationship is an iterated game with asymmetric payoffs, in which both sides are rational and the structure, not intent, not naivety, favors China. Getting there means passing honestly through two weaker readings first, not skipping past them.
Reading one: single-shot, one side naive, the other defects. Defensible only for specific, genuinely irreversible transactions: an open-weight checkpoint, once published, cannot be un-published; a trained researcher cannot be un-trained. Indefensible as the master frame for the whole relationship, which has run continuously since the 1980s and shows every sign of both governments planning for continued multi-decade engagement.
Reading two: genuinely symmetric mutual hostage-taking. The steelman this series is required to build, and the honest version is stronger than most one-way narratives admit: Chinese capital exposure to US equities, Chinese researchers still hired into US labs, and (the single strongest fact undercutting a pure one-way reading) US dependence on Chinese-dominated supply chains for the critical minerals, transformers, and battery components that go into AI infrastructure itself.4 Real bidirectional flows exist and this series states them, per part, before arguing past them.
Where reading two fails is exactly the sensitivity/vulnerability distinction from Section 1: pointing to a real flow in the other direction proves sensitivity is symmetric. It does not prove vulnerability is. On the channels this series tracks (talent replacement cost, grid-power adjustment cost, alternative-sourcing timelines) the cost of adjusting away from the dependency currently runs higher for the United States than for China. That is reading three, reached by passing through, not around, reading two.
The theorists doing the actual work, named here so the parts ahead can use them without re-introducing them: Thomas Schelling on strategic commitment and the "threat that leaves something to chance," which is why export-control tightening functions partly as manufactured irreversibility, a commitment device more than a precise economic lever.5 Robert Jervis on the security dilemma: AI is close to a paradigm case of offense-dominant, low-differentiability technology, where a "safety" advance or a benchmark release is nearly impossible for a rival to read as defensive-only, producing exactly the action-reaction spiral peer-reviewed research has now documented running from ChatGPT's 2022 debut through DeepSeek's January 2025 shock and into 2025's export-control tightening.6 Robert Axelrod on iterated cooperation, and a genuinely useful asymmetry his framework surfaces without any single paper making the claim outright: China's multi-decade planning horizon plausibly gives it a longer "shadow of the future" than a United States whose four-year electoral cycle lets each administration reverse the last one's China policy, degrading its own ability to sustain a cooperative-punishment strategy independent of anyone's good faith.7 Keohane and Nye, already introduced above, doing the most load-bearing work in the whole series. And structural realism (Waltz's relative-gains logic, sharpened by Joseph Grieco and Mearsheimer's account of why a rising power's rise reads as threatening even when framed benignly) explaining why the correction, when it finally came, came late: a period of absolute-gains, engagement-era policy design that only updated after the DeepSeek shock made the relative-gains alarm impossible to ignore.8
RAND's 2025–2026 formal model of a US-China "prisoner's dilemma in the race to AGI" is real, peer-reviewed-adjacent, and cited accurately in the parts ahead, but it's a model of development-pace and catastrophic-risk, the question of whether racing itself is dangerous. That is a different axis from this series' distributional question: who captures how much of the resulting value. The two get conflated constantly in casual commentary. This series keeps them separate.9
/ 05How to Read the Tally
Every part that follows uses the same three devices, in the same order, so this is the only place they get explained.
The Tally
Each part opens with a table of dated, named, sourced figures, not a headcount or a headline number alone, but the concrete cost or time differential underneath it: years of experience, dollars of reported compensation, tariff rates paired against visa policy changed twelve days apart. A tally is not a verdict: some entries are independently verified, some are corroborated-but-unconfirmed, and some are explicitly inferential or hypothetical. Each is flagged as what it is in the footnotes, not smoothed into a single implied certainty.
The Pipes
The connective argument between the tally and the claim, traced through evidence rather than asserted. When a part says a policy shift caused a downstream effect, it shows the dated action and the dated response, not just the correlation.
The Framework
The game theory and IR scholarship named in Section 4 above, woven into the argument at the point where it's doing real work, not boxed off as a separate "theory" section. If a scholar's name appears, it's because their specific claim is load-bearing for the sentence around it.
One more device carries through every part, and it isn't optional: the honest-reckoning aside. Before any part argues its piece of the one-way thesis, it states the strongest real evidence against it (the genuine steelman, not a weakened version built to be dismissed) and only then explains why the net direction still favors China on the vulnerability measure, not the sensitivity one. Where research turned up real, verified symmetry that this series' own numbers can't argue past, the parts ahead say so plainly rather than burying it. That standard is not a hedge. It's the same discipline this platform's other investigative series have applied to themselves, and it's what keeps a one-way argument from becoming a one-sided one.
/ 06The Six Parts
Read in order, or start anywhere: each part's Tally stands on its own. In order: Part 1, The Returnee, opens the talent axis with a sourced scene and the aggregate data that complicates it before the net assessment. Part 2, The Weights, corrects the DeepSeek sequencing and follows the distillation loop onto cheaper power. Part 3, The Route, follows inference demand and the energy arbitrage underneath it, including where Chinese underpricing is unwinding rather than expanding. Part 4, The Corpus, asks what travels inside a routed API call (preference data, alignment technique) and what it's worth on the other end. Part 5, The Allies, is where the fourth thread gets its full, tallied treatment: the Netherlands, Japan, South Korea, Taiwan, named and dated. Part 6, The Ledger, totals every flow from Parts 1 through 5 and closes on the argument that the marginal cost of a watt, not the count of chips, decides the next phase of the competition.
“The chip embargo had removed the hardware. But the hardware had already been mapped by a mind that no customs officer could detain.”
/ 07A Note on Register
Written for readers who don't work a trading desk or sit in an IR seminar. Where an argument turns on game theory, the theory is named and applied, not gestured at: you've already met the full cast in Section 4, and no part introduces a scholar it doesn't put to work. Where an argument turns on a number, the number is sourced, and where sourcing is thin, secondary, or unverified at the primary text, the parts ahead say so in the footnotes rather than presenting a search-summary as a citation. Where an argument turns on a rival hypothesis, the rival hypothesis is argued on its own strongest terms (the mutual-arbitrage reading this whole series exists to argue past) before it's set aside. This series follows The Stack and carries forward the numeracy spine of Social Physics, Article 7, built from Michael Cembalest's August 2026 J.P. Morgan note. Figures reused from that piece are cited by reference, not re-derived or silently duplicated.