AI Macro
A few weeks ago I published a note called The Four Stages of AI, and the Bet Hiding in Your Index Fund. It argued that the AI buildout had quietly become everyone’s bet, whether they chose it or not, and it tried to map where the money actually goes. Since then, I’ve started seeing the same three questions, asked in a hundred different ways:
If AGI is really coming, does it matter whether America or China gets there first? Wouldn’t a true AGI just replace all our systems anyway?
Isn’t China about to do to AI what it did to steel and EVs, and won’t that be catastrophic for the Western AI companies whose valuations assume decades of subscription revenue?
The funding for data centers and power looks shakier every week. If it cracks, doesn’t it take down a stock market that has never been this expensive or this correlated to one theme?
And then, on the first of July, Palantir CEO Alex Karp went on CNBC and answered all three questions at once, in his way, by calling the business model of the frontier AI labs “effing insane,” accusing them of overcharging enterprises while absorbing their intellectual property, and declaring that the anger of American business was being channeled through him. Palantir’s stock jumped 8 percent that day. Zerohedge ran the clip under the headline “Something Has Gone Completely Wrong,” and a few million people came away with the vague sense that the whole AI edifice is a scam about to collapse.
Karp is not wrong about everything. Zerohedge is not wrong about everything. Neither is telling you the whole story, because neither is in the business of telling whole stories. Karp sells the alternative to the thing he was attacking. Zerohedge sells alarm. The only defense a reader has is to see the whole board: every variable, how they connect, and which piece a given talking head is paid to move.
So this note is a field guide. First the board, then the physics of how the pieces interact, then a decoder I am applying to any AI headline, then my scorecard by region, and finally how I currently see the future, year by year, to the end of the decade.
Start with the single most useful idea in this entire subject, because it dissolves about 80 percent of the confusion.
“AI” is not one story. It is three different stories wearing one name, and almost every argument you encounter is people shouting across the gap between them.
Story one is the frontier: who can build the smartest machine. This is a race between perhaps five American labs and a handful of Chinese ones, decided by talent, compute, and technique.
Story two is distribution: whose machines actually get used, by billions of people, millions of businesses, and nearly two hundred governments. This is a war over price, trust, and sovereignty.
Story three is the money: whether the roughly $700 billion a year now being poured into chips, data centers, and power plants ever earns a return, and what happens to the most concentrated stock market in modern history if it doesn’t.
Here is why the distinction matters. “China is months behind on the frontier” is a story one fact. “Chinese models dominate real world developer usage” is a story two fact. “Private equity is getting cold feet on data center deals” is a story three fact. All three are true right now, simultaneously, and none of them contradicts the others. A pundit who mixes them, deliberately or not, can prove anything. America is winning. America is losing. The bubble is bursting. The buildout is unstoppable. Pick your headline; there’s a true fact from the wrong story to support it.
Keep the three stories separate and the news starts making sense. Now, the variables.
One note before the board, because I intend to hold this note to the same standard I’m asking you to hold everyone else to. The evidence below is not all the same weight class. Some of it is documentary: government directives, legislative texts, companies’ own statements about themselves. Some is serious financial and trade reporting. Some is surveys and disclosures that point in a direction without proving a magnitude. I’ve tried to flag which is which as we go, and where a claim rests on the lighter stuff, I say so. What follows is a scenario built from that mixed evidence, argued as the most probable one, not a prophecy. The odds at the end are there so you can grade it later.
The variables
1. The American kill switch. The cleanest demonstration is the most recent one. In June 2026, the Commerce Department ordered Anthropic to cut off its newest models, Fable and Mythos, from every foreign national on earth, including the company’s own foreign-born employees, three days after launch and without a detailed public justification. Access was restored on July 1, three weeks later, again at government discretion. That one is documented in the company’s own statements and is not in dispute.
The earlier episodes are messier, and the mess is instructive. In May 2025, after US sanctions were imposed on the International Criminal Court’s chief prosecutor, his Microsoft-hosted email account in The Hague went dark. The New York Times reported that Microsoft helped disable the account to comply with sanctions; Microsoft’s president insists the company never cut services to the ICC as an institution and that the ICC itself made the operational choice. Both can be true, and it barely matters which version you believe, because the mechanism is identical in either telling: an American executive order reached into a Dutch inbox, and somebody had to comply. The ICC subsequently migrated to a German open source platform, and European ministries reacted to the episode as if the worst version were true. Add the reports (unconfirmed by the company) that Ukraine temporarily lost access to Maxar satellite imagery during a US pause in intelligence sharing, and the pattern that matters for this essay emerges: it is not that America flips the switch constantly. It is that America has shown the switch exists, that its reach follows US law rather than server location, and that foreign governments now plan around it. In geopolitics, a demonstrated capability does not need to be exercised cleanly, or often, to change everyone’s behavior. Perception moved policy, and the policy is variable 2.
2. Allied trust, now hardening into draft law. The response is no longer sentiment; it is legislation in motion. On June 3, the European Commission unveiled its proposed Tech Sovereignty Package, with its tech sovereignty chief stating the goal in one sentence: nobody should have a kill switch. The centerpiece, the Cloud and AI Development Act, would create mandatory sovereignty risk assessments and a tiered framework for sensitive government workloads, and at its top tier it sets conditions that American hyperscalers may be structurally unable to satisfy, not for anything they did, but because the US CLOUD Act legally compels them from home. To be precise about status: this is a proposal that still needs the approval of 27 member states, it will be lobbied hard, and it may emerge softer than drafted. But the direction is not in doubt, and the money is already moving ahead of the law. Gartner projects European sovereign cloud spending will more than triple between 2025 and 2027, from roughly $7 billion to over $23 billion. The Commission ran its first procurement with explicit sovereignty criteria in April, awarding €180 million to European provider groups. France, Germany, and the Netherlands are migrating government offices off American software. This is shaping up as the most durable economic consequence of the second Trump term in technology: allies increasingly treat American digital dependence the way they learned to treat Russian gas.
3. The sovereignty washing gap. The hyperscalers’ answer to variable 2 is “sovereign cloud” regions: your data, kept in your country, under local compliance. Critics have a name for this: sovereignty washing. A local data center provides sovereignty of location, not of control, because no contract overrides a US executive order. This gap caps how much of the trust market American landlords can actually win, and it is the specific crack through which European providers, open-source software, and Chinese models enter markets American firms thought they owned.
4. AGI as a strategic object, not a product. Governments no longer see frontier AI as software; they see it as a strategic capability adjacent to weapons. The Fable episode made this official: the US government treated a commercial chatbot as a munition subject to export control and asserted the power to recall it globally. The open question it left behind, asked across the industry, is whether Washington now effectively approves every frontier model release. Once intelligence is classified as a weapon, buying it from a foreign country becomes a defense procurement decision. Every government on earth is now making that calculation, and it explains most of what follows.
5. The hardware choke points, running in both directions. America controls the chips; China controls much of the rare earth processing the chips (and everything else) depend on. In January, Washington moved sales of Nvidia’s H200 class chips to China from near total denial to case by case licensing, explicitly usable as bargaining leverage; Beijing has retaliated with rare earth restrictions, blocked chip imports at customs, and warned its firms off American hardware. Both sides agreed to a one year truce on their harshest measures, which expires in 2026. The detail most commentary misses: both superpowers are trading the AI race off against other objectives. The race is real, but it is negotiable, which means neither side is behaving as if AGI arrives next year, whatever their press releases say.
6. The electron gap. Intelligence is now made of electricity, and here the positions reverse. In 2024, China added roughly 429 gigawatts of new generating capacity; the United States added about 51. The International Energy Agency projects data center electricity demand roughly doubling worldwide to nearly 945 terawatt hours by 2030, with American data centers alone approaching 9 percent of national electricity consumption. The constraint is already binding: Microsoft has disclosed an $80 billion backlog of cloud orders it cannot fulfill for lack of power. The shorthand analysts use is apt: America has the brains (chips) and a shortage of muscle; China has the muscle and restricted brains, and its labs are engineering around the brain shortage by making models radically more efficient. Then add the politics. Average American residential electricity rates have climbed sharply since early 2025, polls show most Americans would oppose a data center in their own community, and utility rate hikes tied to data centers are now live campaign issues in Pennsylvania, Georgia, Michigan, and Arizona heading into the midterms. A Utah official lost his seat over a single mega project; Virginia moved to strip a $1.6 billion data center tax break; New York is weighing a construction moratorium. Power is the one input the American AI project cannot import, cannot classify, and cannot build without the consent of ratepayers who are turning against it.
7. Taiwan. Nearly every advanced AI chip on earth passes through one island that China claims and America defends. This is the variable under all the other variables. It cannot be hedged, only acknowledged: any confident prediction about 2030 carries a silent asterisk that reads “assuming Taiwan.”
8. China’s counter offer to everyone else. While Washington restricts, Beijing recruits. China has proposed a World AI Cooperation Organization, pushed a Global AI Governance Action Plan, and passed a UN resolution on AI capacity building co-sponsored by more than 140 countries, precisely as the US retreats from international norms-setting. Its Digital Silk Road bundles free open-weight models, local-language services, financing, training, and “AI in a box” data center packages for countries that could never afford the American stack. The pitch writes itself: America rents you intelligence it can revoke; China hands you the weights and you own them forever. Every activation of variable 1 is a commercial for variable 8.
9. The dumping engine. China’s leading labs give their models away, weights and all, at a price of zero. You cannot tariff a download; there is no port, no customs desk. DeepSeek’s flagship prices output tokens around 87 cents per million against roughly $25 for a top American model, while trailing the frontier by an estimated eight months according to the US government’s own testing center (less than the two or three months Chinese benchmark claims suggest, but close enough for most work). The result: Chinese models now lead the rankings on OpenRouter, a marketplace that routes developer traffic to hundreds of models. Be precise about what that measures. It is a developer-routing signal, the earliest and least filtered indicator of where price-sensitive technical users go; it is not a ledger of enterprise contracts, government procurement, or revenue, where American models still dominate. Leading indicators can reverse. But they are called leading for a reason, and this one points the same direction as everything else on the board: the steel and EV playbook executed in its purest form, aimed not at capability but at revenue. The purpose is not profit. It is to ensure no one else profits from selling intelligence either.
10. The economics of agents. The industry’s own products are accelerating variable 9. Agentic AI, systems that work on tasks for hours rather than answering single questions, burns tokens at rates that break every flat-rate pricing plan. The evidence here is a mix of surveys, company disclosures, and trade reporting rather than audited financials, so weight it accordingly, but it all points one direction. A FinOps Foundation survey found 73 percent of enterprises exceeded their AI cost projections this year. Uber reportedly burned through its 2026 AI coding budget within months. Microsoft has reportedly canceled most direct Claude Code licenses in at least one division and redirected engineers to its own tools. Coinbase’s CEO said publicly that the company cut internal AI spending by nearly half by defaulting engineers to Chinese open-weight models. And in the clearest tell of all, confirmed by Microsoft itself to Axios: the company is evaluating an Azure hosted, fine tuned version of DeepSeek to power a cheaper tier of Copilot, its flagship product, built atop its $100+ billion partnership with OpenAI. When cost pressure makes a company shop against its own partner, the pressure is structural, not cyclical.
11. The capex overhang. By their own earnings guidance, the five biggest American tech companies plan to spend roughly $660 to $690 billion on infrastructure in 2026, a figure compiled across analyst coverage of the January-February earnings cycle, and industry analysis notes the spending is increasingly bridged with debt rather than cash flow. OpenAI’s data center commitments have been widely reported at around $1.4 trillion, against a small fraction of that in revenue; treat the precise number as directional, but no reporting disputes the order of magnitude. And there is a loop inside the ledger: an analysis of first-quarter filings found that nearly 60 percent of the combined income at Alphabet and Amazon came from “other income,” largely paper gains on equity stakes in the very AI startups that spend the money back on their clouds. That figure comes from filings, but the framing is the analyst’s; read it as a symptom of circularity, not proof of fraud. The harder physics is undisputed: unlike railroads or fiber, which lasted generations, AI chips are largely obsolete in three to five years, so the depreciation bill from the 2025-26 buildout arrives on earnings statements in 2027-28, right on schedule to collide with variable 9’s price collapse.
12. American political volatility as a priced risk. Foreign buyers can no longer model American AI policy, only American AI policy this quarter. The Pentagon labeled Anthropic a supply chain risk in March over the company’s refusal to drop restrictions on surveillance and autonomous weapons; three months later Commerce shut off its models; three weeks after that it turned them back on. Chip rules flip between denial and licensing with the news cycle. Unpredictability is itself a sovereignty argument, and it is one America’s competitors did not have to invent.
13. The wildcards. Model theft and distillation (a suspected China linked intrusion reportedly contributed to the Fable panic; Alibaba, from the other side, has accused Anthropic-related actors of a “distillation attack”), a first genuinely catastrophic AI-enabled cyberattack, a capability discontinuity nobody predicted, or a domestic political rupture in either superpower. Any one of these can reset the board in a week. You cannot predict them. You can only refuse to be surprised that surprises happen.
The loops
A list of variables is a parts diagram. The machine only makes sense when you see how the parts push on each other. Four feedback loops drive nearly everything, and at this moment all four run in the same direction.
The trust loop. The more Washington treats frontier AI as a weapon (variable 4), the more it uses its kill switch (1), which accelerates allied de-risking (2), which shrinks the trusted global market for American AI, which reduces the revenue that has to justify the capex (11), which raises the pressure on American firms and their government to monetize control even more aggressively. Around it goes. Every flip of the switch is a sales event for the alternative.
The controls loop. American chip restrictions (5) forced China to build a self-sufficient stack and to open-source aggressively as an asymmetric weapon (8, 9), which collapsed the world price of ordinary intelligence, which pulls even American companies onto Chinese models (10), which intensifies Washington’s urge to restrict further. Export controls, whatever their national-security merits, created the very ecosystem that is now dismantling American AI pricing power. Policy built the competitor it feared.
The sovereignty loop. AGI race logic makes every nation want its own AI capability (4), but only two countries can build frontier models, so for the other 190, “sovereignty” means choosing which dependence to accept. American dependence is more capable but revocable (1, 3). Chinese dependence is less capable but ownable, because open weights, once downloaded, cannot be recalled by anyone. Therefore, the hotter the superpower race burns, the more the middle of the world hedges toward the thing that cannot be switched off. This is the central irony of the whole era: America’s capability lead and America’s control instinct are the same policy, and the control instinct is what markets Chinese AI to the planet.
The power loop. The buildout (11) drives electricity demand up faster than a slow-permitting American grid can add supply (6), which pushes rates up, which feeds a ratepayer revolt that is already deciding local elections, which slows permitting further, kills tax breaks, and imposes special tariffs on large loads, which raises the cost and delay of every new American data center, which pushes the physical layer of the industry toward whoever offers cheap, fast, politically frictionless power: the Gulf states first, and eventually anywhere that is not a contested American grid. Compute follows electrons. Sovereignty questions follow compute. American grid politics are quietly exporting the physical layer of America’s own AI industry, one canceled substation at a time.
Notice what the loops share: none of them is primarily about which model is smartest. The frontier race gets the headlines. The loops decide the outcome.
3 Questions
Now put the board to work. When you encounter any confident AI claim, from a CEO, a senator, a newsletter, or a guy on a forum, run it through three questions.
Question one: what does the speaker sell? Karp’s July 1 appearance is the perfect specimen. He attacked token pricing, warned that enterprises were handing their intellectual property to the labs, and invoked the voice of American business. What was he there to announce? A Palantir-Nvidia partnership selling sovereign AI deployments where customers own their compute, models, and data: precisely the product his diagnosis prescribes. The day before, Palantir published a nine-point manifesto on AI sovereignty. None of this makes him wrong. A FinOps Foundation survey found 73 percent of enterprises blew past their AI cost projections this year; the pain he describes is real. The most effective salesmen tell true stories that happen to end at their product. And the test cuts every direction, including at the fix: France’s government has begun backing away from Palantir itself, warning against new strategic dependencies, and British health-service leaders want its contract dropped. The cure for dependency, it turns out, is sold by another dependency. Lab CEOs warning of AGI risk are selling the significance of their product. Hyperscalers projecting serenity are selling continuity. Zerohedge is selling alarm; alarm is the subscription. Knowing the product doesn’t tell you who’s lying. It tells you where to apply the discount.
Question two: which story is being told? Karp’s rant was a story two argument (distribution: who controls the stack) wearing story-one clothes (the labs oversold their models). The Zerohedge framing converted it into a story-three claim (the whole edifice is cracking). Three different stories, one seven-minute clip. Most bad AI takes are exactly this: a true fact from one story deployed as evidence in another. “China is behind” (frontier) does not refute “China is winning adoption” (distribution). “Data center financing is wobbling” (money) does not refute “the technology works” (frontier). Sort the claim into its story before you evaluate it, and half the apparent contradictions in the news dissolve.
Question three: what would have to be true for the speaker to be wrong? For Karp: if frontier models keep pulling far enough ahead of open ones, renting the best beats owning the mediocre, and the token model thrives at the high end. For the doomers: if enterprise revenue accelerates faster than depreciation lands, the financing holds and the buildout looks visionary. For the boosters: if “good enough” intelligence at 3 percent of the price captures most real work, the premium tier never gets big enough to pay for the cathedral. If you cannot state the falsifying condition, you don’t have a view; you have a mood.
Does it matter who reaches AGI first, if a true AGI would just replace our systems anyway?
This is the best question amateurs ask and experts dodge, and the honest answer is yes, but not for the flag planting reason the talking heads imply. The question assumes AGI arrives as an independent mind that surveys our governments and economies and selects better ones. But AGI will not be discovered like a comet. It will be built like a cathedral, and it will carry its builders’ fingerprints: what it values, what it refuses, whose interests it treats as the default. That gets decided during training, by people, under a government.
This is measurable today. Security researchers found that a leading Chinese model produced sharply more insecure computer code when prompts merely mentioned politically sensitive subjects like the Uyghurs or Falun Gong, and refused sensitive requests nearly half the time. The politics of the model’s birthplace live inside the machine, all the way down to the quality of its code. Scale that to a system running infrastructure, drafting law, and educating children.
So run the three scenarios. If AGI remains a tool, it matters enormously who holds it, because it amplifies its owner’s system rather than replacing it. If AGI is semi autonomous but shaped by training, it matters what values were instilled, because those become the defaults for civilization, and almost nobody changes defaults. And if a fully independent superintelligence escapes its makers entirely, then “who wins” collapses into a darker question: did the people who built it first care about caution at all? First place matters most in that scenario, because the leader chooses how much safety margin goes into the final steps. There is no branch of the tree where it doesn’t matter. What the question gets right is the word “war.” Nobody plants a flag on AGI. The winner doesn’t conquer the world; the winner sets its defaults.
Is China dumping AI the way it dumped steel and EVs, and is that catastrophic for Western AI companies?
Yes to the first, with a twist that makes it more potent than steel: the dumping price is zero and the product is a download, immune to tariffs by its nature (variable 9). The strategic logic is identical, though: not to profit, but to deny profit. Beijing cannot beat America to the smartest model, so it is making the second-smartest model free, which destroys the pricing power of everyone selling the smartest one. And the agent-economics problem (variable 10) means the dumping has willing customers inside the citadel: Microsoft, Coinbase, and thousands of engineering teams choosing price over patriotism every day.
“Catastrophic,” though, depends on a distinction the forum posts miss. Dumping crushes the price of average intelligence; it does not, by itself, touch the value of frontier intelligence. The question that decides the fate of the Western labs is how large the premium tier turns out to be: the work where being eight months smarter, more reliable, and legally accountable is worth real money. Drug discovery, defense, chip design, high stakes finance, and the agentic coding that is currently the labs’ best business. If that tier is large, the labs prosper as premium suppliers. If most of the world’s work turns out to be “good enough is good enough,” their revenue projections are fiction and the correction will be brutal. Nobody knows yet, and anyone who tells you they know has failed question three of the decoder.
Is the infrastructure financing cracking, and does it take the market with it?
The timing is arguable. Review variables 6, 9, 10, and 11 together: revenue per unit of intelligence is being compressed by dumping and by customers’ own cost revolts, at the same moment the depreciation from a $700-billion-a-year buildout, funded increasingly by debt and booked partly through circular equity gains, arrives on income statements, and the power to run the machines is getting more expensive and politically harder to secure. Chips don’t last like railroads. The bet has to pay fast, and both the price of the product and the availability of its fuel are moving the wrong way.
The likeliest resolution is not a 1929 pop but a sorting. The hyperscalers can absorb enormous pain: they have real profits from other businesses, and they will own useful infrastructure either way, just as the dark fiber of the dotcom bust powered the following twenty years of the internet. The danger concentrates in the leverage: debt-funded data center developers, the “neocloud” GPU landlords, energy projects underwritten by a single tenant’s projections, and any lab whose trillion-dollar commitments assume revenue the dumping is currently eating. When markets reprice the difference between owning infrastructure and owing for it, the move can be violent for an index this concentrated without the underlying technology failing at all. That is the honest shape of the risk: not “AI was fake,” but “the financing assumed a price that the product no longer commands.” If you own an index fund, you own this question. That was the argument of the last essay, and every month since has sharpened it.
Run the variables and loops region by region, and the geopolitical and economic outcomes come into focus. Not who “wins the AI war,” but what each player actually holds at the end of the decade.
The United States: wins the technology, strains the society, holds the risk. America keeps the three things hardest to replicate: the frontier labs, the chip design monopoly, and the deepest capital markets on earth. What it is losing is quieter: the presumption of trust that made American technology the world’s default, and the domestic consensus needed to build the physical layer at home. The trust loop shrinks its export market; the power loop raises its costs and fills its town halls; and the capex overhang means the financial risk of the entire global buildout is concentrated in American equities, American credit, and, through index funds, American households. The macro picture: enormous productivity gains if the technology delivers, captured disproportionately by capital, alongside rising electricity bills, land fights, and a political backlash that has already started deciding elections. America’s strategic problem is that it is trying to run a global technology empire and a domestic affordability politics off the same overloaded grid, and something has to give. Watch the dollar, too: a world de-risking from American clouds and American discretion has one more reason to diversify from American assets generally, and the AI trade and the dollar’s privilege are more entangled than either debate acknowledges.
China: loses the frontier, wins the map, and pays for it in deflation. China ends the decade months behind on capability and years ahead on presence: its models running on every continent, its standards embedded in the digital infrastructure of the Global South, its energy abundance letting it build compute the way it once built highways. The economic outcome is stranger than winning: China gives the product away, so AI adds little direct profit to an economy already fighting deflation and weak domestic demand. What Beijing buys instead is strategic position: it denies America the rents, anchors a hundred countries to its stack, and converts second place technology into first place influence at a discount no democracy could sustain. The risks: the chip ceiling if the truce collapses, the possibility that efficiency tricks stop compensating for restricted hardware, and the chance that dumping intelligence into its own labor market accelerates the employment problems it already has. China’s bet is that owning the world’s defaults beats owning the world’s margins. It is not an obviously bad bet.
Europe: no frontier, but the rulebook, and finally a reason to build. Europe will not produce a top three model this decade; that argument is over. What Europe is doing instead is converting its weakness into jurisdiction: the Tech Sovereignty Package, the four tier cloud rules, the €120 billion the Commission says the continent needs to invest in its own capacity by 2035. The near-term economics are unflattering: Europe pays more for less capable AI in exchange for autonomy, and its productivity gap with America likely widens before it narrows. But three assets compound in its favor: cheap clean power in the Nordics and Iberia that the power loop makes more valuable every year, a sovereign-cloud and open-source industry being legislated into existence with guaranteed government demand, and the Brussels effect, by which European compliance rules become the world’s rules because global companies build to the strictest standard. Europe’s endgame is to be to AI what it is to privacy and antitrust: not the inventor, but the referee everyone must satisfy. Referees are never rich, but they are never irrelevant either.
The Gulf: the swing vote, selling the one thing everyone is short. The UAE and Saudi Arabia hold the decade’s scarcest combination: unlimited cheap energy, sovereign capital that can write $100 billion checks, and no ratepayers to revolt. Stargate UAE, the Microsoft Saudi regions, and the sovereign funds’ stakes in every American lab make the Gulf the place where the power loop’s exported compute lands. The Gulf states are running the Switzerland play with electricity: hosting American frontier infrastructure while keeping Chinese options warm, converting oil wealth into compute wealth before the oil era closes. Their risk is that they are renting relevance: the chips are American, the models are American or Chinese, and a serious US-China rupture forces the choice they have so far avoided. But in the base case, the Gulf ends the decade as the world’s neutral compute landlord, which is a better business than oil ever was per barrel of political risk.
India and the middle powers: leverage without alignment. India is the swing market the way the Gulf is the swing supplier: the second-largest user base for American AI products, exempted from the chip restrictions that bind others, hosting gigawatt-scale builds from both OpenAI and Google, while its data localization laws force everyone to build inside its borders on its terms. India, Japan, South Korea, and Brazil share a strategy: extract infrastructure, jobs, and technology transfer from both superpowers while pledging allegiance to neither. Japan and Korea add a quieter position: they sit inside the chip supply chain itself (memory, materials, equipment), which makes them beneficiaries of the buildout regardless of whose models win. The middle powers’ collective outcome is the best risk-adjusted deal on the board: they get the consumer surplus of nearly free intelligence, the capex of two competing empires, and none of the bill.
The Global South: the biggest windfall and the deepest lockin. For most of Africa, Southeast Asia, and Latin America, the practical outcome of everything above is that world class intelligence becomes effectively free years before anyone expected: open weights on cheap hardware, financed and installed through China’s Digital Silk Road, in local languages, with training included. Measured as consumer surplus, this is one of the largest transfers of capability in economic history, and it would be condescending to pretend otherwise. The cost is the oldest one: the stack comes with standards, dependencies, and switching costs, layer stacking on layer until leaving is unthinkable. These countries spent the twentieth century as price-takers for manufactured goods; the question of the twenty-first is whether taking your intelligence from someone else’s stack is different in kind, and whose stack leaves more room to grow. Russia, for completeness, is not a player in this game at all: it consumes Chinese AI as a client state and contributes weaponized disinformation as its only export in the field.
Prediction in a system with thirteen variables and four live feedback loops is a fool’s errand if you promise precision. It is a duty if you promise only structure. Here is the base case, in stages, with the branch points marked.
Stage one, now through 2027: the great sorting. The world stops pretending there is one AI market and starts building three. The frontier layer stays American and becomes progressively less commercial: security reviews, release approvals in fact if not in statute, labs entangled with the state whether they resist (Anthropic) or embrace it (OpenAI, whose international pitch is explicitly built on coordination with the US government, which is now both its shield and its liability). The deployment layer, where most actual usage lives, tilts toward open weights, increasingly Chinese descended, pulled by price and pushed by distrust; the developer-routing data already shows the tilt, and the open question in this window is whether it climbs the stack into enterprise contracts and government procurement, where American models still hold the ground and where liability, security, and trust weigh more than price. The trust layer fragments nationally: Europe legislates its tiers, the Gulf builds its prestige projects with both superpowers, the Global South takes the Digital Silk Road package because it is the only offer on the table. In America, the midterms are the first national election in which electricity prices and data centers are a top tier issue, and the result, whichever party benefits, slows domestic permitting and pushes more of the buildout abroad. Watch four tells in this window: whether Washington forbids American clouds from hosting Chinese models (a decision that would formalize the split and hand the neutral world to Beijing), whether the chip and rare earth truce survives its expiration, whether any second country follows the Fable precedent with a model recall of its own, and how many gigawatts of announced American projects quietly migrate to the Gulf.
Stage two, 2027 through 2028: the bill arrives. Depreciation from the buildout’s peak years lands on earnings at the same time commodity intelligence approaches free and power costs keep climbing. This is the maximum stress window for story three. The likeliest single event is a failure or forced rescue among the leveraged middle: a neocloud, a mega data center project, or a lab funding round that reprices everything downstream. Expect the first serious IPOs from the labs before or during this window, because private markets will not carry these burn rates forever; expect them to be oversubscribed and volatile, and note (per the decoder) that some of the loudest current attacks on lab business models come from companies whose stock competes for the same institutional dollars. If enterprise revenue has accelerated enough by now, this stage passes as turbulence. If not, this is when the index fund bet gets marked to market. Either way, the infrastructure survives its financiers, as it always has.
Stage three, 2029 and beyond: the settlement. The stable configuration, absent wildcards, looks like this. The smartest machines are American, and access to them is a managed strategic good, sold to allies under conditions, the way advanced weapons are now. The most-used machines are open-weight descendants of Chinese releases, running everywhere, owned by their users, generating almost no revenue for their creators, exactly as intended. Europe never builds a frontier model that matters, but writes the compliance rules for the space in between and rebuilds a real domestic infrastructure industry doing it. The physical layer has partially migrated to wherever electrons are cheap and politics are quiet, with the Gulf as the largest single beneficiary. The durable profits sit with the owners of buildings, power, chips, and trust arrangements: the landlords, some old (Microsoft, Google, Amazon) and some new (national champions Europe and the Gulf are subsidizing into existence right now). The labs either grow into the premium tier or are absorbed, gently or otherwise, into the hyperscalers and the state.
If you remember nothing else, remember the shape. The AI race everyone imagines, two superpowers sprinting toward a finish line, is real, but it is generating side effects larger than the race itself. By treating intelligence as a weapon it can switch off, America is teaching the rest of the world that American intelligence is a dependency to be escaped. China, unable to win the frontier, is giving intelligence away in a form no one can revoke, converting second place capability into first place distribution. The machines run on electricity that China builds faster than anyone, that America increasingly cannot build without a town-hall fight, and that the Gulf sells to both. And the money is flowing to whoever owns the ground the machines stand on, while the debts taken to build that ground come due just as the product’s price collapses toward zero.
Every alarming clip, every triumphant keynote, every forum thread is one of three stories, told by someone selling something, and now you can name the story, name the product, name the region whose scorecard it affects, and name what would prove the speaker wrong. That will not tell you the future. It will do something more useful: it will make you very hard to play.
The story I’m telling is mostly story two, distribution, because I think it’s the underpriced one. And what would prove me wrong is written into the piece: if the premium tier for frontier intelligence turns out to be enormous, if the EU package dies in the Council, if enterprise revenue outruns depreciation, or if the frontier gap widens instead of narrowing, then the sorting I’ve described stalls and the American labs’ decade looks very different. Watch those four things, not the headlines.