In 2023 I started a series on AI and geopolitics. The core thesis was that the race toward artificial general intelligence resembled the atomic race: whoever got there first would gain outsized influence over the global order. I predicted the United States would push a "let's unite forces" narrative and that China would resist it, betting on its centralized model.
Three years later, reality turned out more interesting than my prediction. China did not resist the global-cooperation narrative. It claimed it, on its own terms: Xi Jinping is championing international AI cooperation built on open-source models. And the United States, which I imagined as the champion of openness, ended up defending an ecosystem of closed models and premium-priced APIs.
The roles flipped. It is worth understanding why, and above all, who wins on this new board.
A word of caution, the same one that has run through this series: these are personal views, low-resolution arguments, not definitive analyses. I write to think out loud, not to dictate truth.
The move nobody expected
There is an old maxim in technology strategy: commoditize your complement. If you sell infrastructure, you want software to be free. If you sell distribution, you want content to be abundant.
Seen through that lens, China's open-source push is neither generosity nor ideology. It is a gambit: you sacrifice the model's margin to capture everything else. That means the cloud it runs on, the applications that use it, the standards that govern it, and the dependence of the countries that adopt it. When a company in Jakarta, São Paulo, or Mexico City builds on Chinese open weights, it signed no treaty, but it is already inside a technical sphere of influence.
The move strikes directly at the American business model. Frontier labs need premium token prices to justify investments in the hundreds of billions. Chinese open models make that premium harder to sustain every quarter.
The signals
This has stopped being theory. Three data points from the last month:
- Microsoft is testing Kimi K3, the model from China's Moonshot AI, for Copilot tasks that currently run on OpenAI and Anthropic. Estimated savings: up to $600 million, at a per-token cost roughly 60% lower. That OpenAI's largest investor is running these tests says everything about where cost pressure is pointing.
- Chinese models now capture around 63% of usage on OpenRouter, the largest neutral model router. A year ago, American models held ~70% of that traffic; today they hover around 30%. And roughly 70% of the world's newly created open derivative models are built on Alibaba's Qwen.
- Vercel's data points to the key asymmetry: open models already process close to 30% of enterprise tokens while representing less than 4% of spend. Intelligence is becoming abundant and cheap far faster than it is becoming profitable for whoever sells it.
That last data point is the heart of the matter. For those who use AI, this is the best possible scenario. For those who hoped to charge a premium for it, it is a forced repricing.
Who wins
Financial coverage obsesses over the losers: Nvidia, Microsoft, Alphabet under pressure. I think that is the wrong angle. When the cost of an input collapses, the useful story is not who was selling the expensive input. It is who builds on top of the cheap one. When electricity got cheap, the winners were not the utilities: the winners were everything that could be electrified.
My map of the most likely winners:
China's full-stack platforms. Alibaba is the clearest case: Qwen dominates the open ecosystem, its cloud sells the infrastructure to run it, and it was chosen to power Apple Intelligence in China. Give the model away and charge for everything else. The Android playbook, executed with intelligence. Baidu and Tencent are running versions of the same book.
Inference infrastructure. If the model is free, the business is serving it: routers like OpenRouter, specialized inference silicon, regional clouds hosting open weights. Token volume grows even as price per token falls, and whoever collects a toll on volume wins on traffic, not on model margin.
The application layer. Every company whose product uses intelligence rather than selling it just received a structural subsidy: its main input cost drops 10x while its selling price does not have to. The quiet winners of this rotation are software with distribution already built, and vertical agents that solve complete problems.
The non-aligned world. India, the Gulf, Southeast Asia, Latin America. Frontier-class open weights mean a country can host its own sovereign AI without asking permission from Washington or Beijing. The geopolitical optionality the developing world never had with oil or with semiconductors, it is getting with intelligence.
And a note on hardware. The consensus that Nvidia loses feels premature to me. When intelligence gets cheaper, more of it gets consumed. That is the Jevons paradox. Demand can rotate from training to inference without collapsing. Nvidia's real risk is not that fewer chips get used: it is that inference admits far more competition than training ever did.
Predictions
So this essay can age honestly, concrete bets on a 12 to 24 month horizon:
- Another American hyperscaler will deploy a Chinese open model (fine-tuned and rebranded) in a mass consumer product, following the path Microsoft is exploring with Kimi K3.
- Frontier token prices will fall another 50% or more, forced by the open alternative. The "GPT-4-class token" trends toward zero marginal cost.
- Washington's response will be regulatory, not technical: restrictions on Chinese weights in government and critical sectors. Competition will shift from capability to trust: audits, data provenance, certifications.
- Spend will follow usage. The Vercel gap (30% of tokens, 4% of spend) will close from below: open-model spend will exceed 15% of enterprise AI budgets.
- The market rotation will reward distribution and application over labs. We will see at least one consolidation or down round among frontier labs that own neither their own cloud nor their own distribution.
The question that changed
In 2023, this series asked: who builds superintelligence first? That question remains open. But 2026 forced an earlier, more uncomfortable one: who captures the value when ordinary intelligence becomes free?
The first is an arms race. The second is a trade war, and trade wars are not won by whoever holds the most advanced weapon, but by whoever redefines the terrain so the rival's weapons lose their value. That is exactly what the open-source gambit is trying to do.
The chessboard is the same one I described three years ago. What changed is that one of the players stopped protecting their queen and started giving away pawns across the entire board. History suggests you should not underestimate that player.