AINEWS
The Two-Horse AI Race Has a Third Lane. It Is Already in Use.
A Rest of World panel at New York Tech Week mapped the U.S.–China AI duopoly and surfaced the question middle powers cannot avoid: how to act before the dependency is hardwired.
A Rest of World panel at New York Tech Week mapped the U.S.–China AI duopoly and surfaced the question middle powers cannot avoid: how to act before the dependency is hardwired.
The framing of the U.S.–China AI contest as a "two-horse race" with Washington and Beijing controlling roughly 90% of the world's AI capacity is no longer just pundit shorthand. It is the operating picture three policy and academic specialists sketched out this month at a Rest of World panel convened during New York Tech Week, and it is the picture regulators, buyers, and ministries outside the two poles are now budgeting against.
Sam Winter-Levy, a fellow at the Carnegie Endowment for International Peace, told the panel that the consolidation is not abstract. It shows up in compute, in capital, and in product decisions that quietly lock in dependency by design. Winter-Levy pointed to the Anthropic Mythos managed-access rollout as a concrete example. When the frontier is gated by one company's terms, the rest of the world is buying into a relationship, not just a model.
The structural numbers are not subtle. The two leading AI economies account for an estimated 90% of AI capacity, and roughly 70 to 80% of investment flows through the same corridor, according to figures cited in the panel summary. A small roster of companies, anchored by Nvidia on chips and OpenAI and Anthropic on frontier models, sits at the choke points. The reference point most panelists used was the end-of-2022 ChatGPT launch, the start of a concentration now widely treated as structural rather than cyclical.
The question the panel kept returning to, articulated by Peter Micek, general counsel and UN policy manager at Access Now, was not whether the divide exists but whether it can be navigated without becoming a client of one of the two poles. The constructive answer, as Aditya Vashistha of Cornell's Global AI Initiative framed it, depends on whether middle powers use the leverage they still have: procurement rules, data localization regimes, public-interest compute investments, and coalition diplomacy that treats AI as critical infrastructure rather than just another software market.
That is where the third lane actually lives. It is not a third model lab that will catch OpenAI on raw capability. It is the slow, regulatory, and procurement work of defining what dependency a country will and will not accept. AI-specific compliance regimes modeled on the EU AI Act, public-sector procurement with explicit data-sovereignty terms, and national investments in domestic compute all sit in this lane. None of them breaks the duopoly. All of them shape which duopoly pole a country is structurally tied to, and on what terms.
The risk of getting it wrong is not just commercial. Micek warned that middle powers face the worst of both worlds: job disruption and social effects from AI systems they do not control, exported to them on terms they did not negotiate. The exposure-without-benefit structure is the political problem the panel kept returning to, and the one that does not have a technical fix.
The Rest of World conversation, reported by Rina Chandran and edited for brevity, is itself a symptom. A respected international tech outlet convened a panel to ask, in effect, what is left to do. The honest answer from the panelists was: less than there was, more than the doom framing suggests, and almost entirely a function of choices being made now in procurement offices, foreign ministries, and coalition talks that do not make headlines. The two-horse race is real. Whether it stays a two-horse race depends on whether the rest of the field treats its current positioning as a starting point or a verdict.