Why the Country With the Best Model May Still Lose
There is a comfortable story Americans tell about the race for artificial intelligence. In that story, the contest is a sprint, the finish line is the most powerful model, and the United States is winning. The story is not wrong, exactly. By most public measures, the strongest models in the world are still built in American labs.1 But the story leaves out the part that may matter most. It assumes that whoever builds the best engine wins the race. It does not ask whether anyone is driving the car.
The better way to understand what is happening is to picture two runners on the same track, chasing two different finish lines. The United States is running toward raw capability: the smartest model, the hardest benchmark, the next breakthrough. China is running toward something less glamorous and possibly more decisive: getting usable AI into the hands of as many people, companies, and countries as possible, as cheaply as possible, with as few strings attached as possible. These are not the same race. And the quiet argument of this essay is that China may be winning the one that counts, not by building the most impressive machine, but by filling the stadium while America perfects a trophy that fewer and fewer people will ever touch.
This is not a prediction of American decline. It is an observation about what kind of advantage actually compounds over time, and a warning that the most-watched scoreboard may be measuring the wrong game.
The Two Runners
Start with the runners themselves, because they are built differently and they are trying to do different things.
America’s Bet: Build the Frontier
The American approach is a bet on capability. The logic is straightforward and, on its own terms, sound: if you build the most powerful AI in the world, the economic and military advantages will follow. Get to the frontier first, and the rest takes care of itself.
The companies making this bet are the largest in the world. Microsoft, Google, Amazon, and Meta — the four giants that build and rent out most of America’s AI infrastructure — are spending on a scale that has no real precedent. On their spring 2026 earnings calls, the four told investors they now expect to spend as much as $725 billion on capital expenditure this year, the bulk of it on AI data centers, up roughly 77 percent from about $410 billion the year before.2 Microsoft alone guided to $190 billion for the year; Amazon to about $200 billion; Alphabet to as much as $190 billion; and Meta to as much as $145 billion.3 These are not the budgets of companies hedging their bets. They are all-in.
The spending buys real leadership. The frontier models that top the hardest reasoning and coding evaluations — OpenAI’s GPT-5 family, Anthropic’s Claude Opus line, Google’s Gemini — are American.4 When a new benchmark is designed specifically to separate the best models from the merely good, it is usually an American model at the top.5 On the question the American strategy cares most about — who has the smartest model — the answer is still the United States.
The military has made the same bet. After years of treating AI as a research project, the Pentagon has begun fielding it. Project Maven, the program that uses AI to sift through surveillance data and identify targets, has moved from experiment to a core command-and-control system, with more than $1.3 billion in contracts behind it.6 The Replicator initiative is built to field large numbers of cheap, autonomous systems quickly.7 The bet is the same as the commercial one: superior capability, deployed for advantage.
But a bet this large carries risks that grow with it, and the American strategy has four of them.
The first is energy. AI does not run on ideas; it runs on electricity, and the United States is running short. A single large training run can consume a meaningful slice of a regional power grid at peak.8 New data centers wait years for the permits and grid connections they need, and the wait is getting longer, not shorter.9 The compute is only useful if you can power it, and power is becoming the binding constraint.
The second is money. Spending three-quarters of a trillion dollars in a single year only makes sense if the returns arrive. So far, much of that return is still a promise. Analysts and investors have begun to ask the uncomfortable question aloud: what happens if the revenue does not show up on schedule?10 Shareholder patience is not infinite, and a strategy that depends on ever-larger spending is also a strategy that depends on ever-larger faith.
The third is talent. For years the assumption was that the best AI researchers would always want to work in America. That assumption is weakening. Chinese labs have become genuinely attractive places to do frontier work, and the flow of talent is no longer one-directional.11
The fourth is the deployment gap, and it is the one that matters most for the argument here. Building the best model is not the same as getting people to use it. Across much of the economy, AI adoption is still stuck in what practitioners call “pilot purgatory” — endless trials that never quite become standard practice. Even in the United States, surveys through early 2026 found that only around half of businesses were paying for any AI service at all, and overall business AI usage hovered between 17 and 20 percent.12 The frontier is real. The question is how much of it is actually being used.
China’s Bet: Build the Plumbing
China is running a different race. Its bet is not on having the smartest model. It is on integration — on making AI cheap, available, and easy to bolt onto real problems, and then spreading it as widely as possible.
The cast of players is large and increasingly capable: DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu, MiniMax, Baidu’s ERNIE, and the hardware backbone from Huawei, much of it backed and coordinated by the state. Their strategy rests on a few simple advantages.
The first is openness. Many of the strongest Chinese models are released as open weights, often under permissive licenses that let any developer download, modify, and deploy them freely.13 A company does not have to ask permission or sign a contract. It just takes the model and builds.
The second is price. Chinese models are dramatically cheaper to run. The gap is not small. Where a frontier American model can cost on the order of $10 to $30 per million tokens of output, a strong open Chinese model can cost a fraction of a dollar — in some comparisons more than a hundred times cheaper for tasks where “good enough” is genuinely good enough.14 For a business deciding whether to put AI into a product, that difference is the difference between a line item and a rounding error.
The third is that the quality gap, the thing the American strategy is built around, has been closing fast. Chinese open models now trade benchmark wins with the American frontier in specific domains. Moonshot’s Kimi K2.6 leads several frontier models on a respected software-engineering benchmark; DeepSeek’s V4 is at or near the top of the open-source field for code and math.15 The honest summary is that American labs still hold a lead of roughly half a year on the hardest, broadest tasks — world knowledge, complex agentic work, multimodality — and that lead is largely holding.16 But for the bread-and-butter work that most real deployments actually involve, the gap is now measured in single-digit percentage points, while the cost gap is measured in multiples.17 When the quality difference is small and the price difference is enormous, most of the world does the math the obvious way.
China’s strategy has force multipliers that America’s lacks. The state coordinates: industrial policy, subsidies, and a domestic chip push align the public and private sectors in a way that is hard to replicate in a market economy.18 And the doctrine of civil-military fusion means commercial AI gains flow into military capability by design — the logistics model that runs a delivery company can run a military supply chain.19
China’s bet is not free of constraints either, and they are serious. Its AI capital spending is still far below America’s; even ByteDance, one of its most aggressive spenders, is weighing capex of up to $70 billion for 2026, a fraction of what the American four are deploying.20 And China’s labs still need access to the most advanced chips for their largest training runs — chips that American export controls are designed to deny them.21 But here the story takes a turn that the architects of those controls did not intend.
The Track Conditions
Neither runner races in a vacuum. A set of external forces — rules, controls, energy limits, and buyers — reshapes the payoffs for both, and they deserve their own attention because they are constraints on the contestants, not contestants themselves.
The first is regulation, and specifically the European Union’s AI Act. Europe is not building frontier labs, but it is writing the rulebook that frontier labs must follow if they want access to the European market, an effect often called the “Brussels effect.” That rulebook, however, has just become less rigid. In May 2026, EU lawmakers reached a political agreement to delay many of the Act’s toughest high-risk compliance deadlines by up to two years, pushing most obligations to 2027 and 2028.22 Some rules, like transparency and watermarking requirements for AI-generated content, still arrive on the original 2026 schedule.23 The net effect is a moving compliance target that both American and Chinese labs must design around — a cost and a complication for everyone, favoring no one in particular.
The second is export controls, and this is where America’s strategy may be quietly undermining itself. The United States restricts the sale of its most advanced chips to China, hoping to freeze Chinese AI a generation behind. The policy has not worked as intended, for two reasons. Enforcement leaks — chips are smuggled, and third countries serve as grey markets.24 More importantly, the controls handed China a powerful incentive to build its own chip industry, and it did. Huawei’s AI chip revenue is on track for roughly $12 billion in 2026, up more than 60 percent year over year, and its latest processors are being adopted at scale by Chinese firms that can no longer reliably buy from Nvidia.25 Nvidia’s chief executive has acknowledged that the company’s market share in China, once dominant, has fallen to essentially zero.26 DeepSeek’s latest model was tuned to run on Huawei hardware.27 A Washington think tank summarized the outcome in a report title that needs no commentary: “Backfire.”28 The controls did slow China’s frontier work at the top end. They also gave China a reason to stop depending on America at all.
The third is energy, and it constrains everyone. Data center moratoriums and restrictions have spread well beyond the United States. Ireland effectively froze new data center connections around Dublin after the industry’s electricity use passed a fifth of the entire country’s consumption; Singapore, expected to devote nearly 20 percent of its national grid to data centers in 2026, banned new ones for years before reopening under strict efficiency rules.29 Gigawatts, not algorithms, increasingly set the ceiling on both teams. China’s answer has been to move its compute to its power: a strategy of building data centers in the west of the country, in Xinjiang and Inner Mongolia, where coal and renewable energy are cheap, and routing the data to them.30 It is willing to accept some efficiency loss to get scale and cheap electricity — a trade the American grid is not currently set up to offer.
The fourth is the sovereign buyer. The wealthy Gulf states — the United Arab Emirates through G42 and Saudi Arabia through its Humain venture — are now buying compute at a scale that rivals the hyperscalers, building national AI infrastructure and bidding up the global price of chips and talent for everyone else.31 They are not contestants in the two-runner race, but their checkbooks change the track for both.
The Current Standings
Put the runners and the track together, and a picture emerges that is more interesting than “America is winning.”
On compute and frontier capability, America leads, and it is not close. The $725 billion in spending and the benchmark-topping models are real.32 But that lead is expensive, energy-constrained, and increasingly questioned by the people paying for it. It is a lead, but it is a lead with a meter running.
On integration and adoption, China leads, and — this is the crucial part — its lead compounds. There is a flywheel hidden in the integration strategy. Cheap, open models get adopted widely. Wide adoption generates enormous amounts of real-world usage data. That data makes the next round of fine-tuning better. Better, cheaper models get adopted even more widely. Each turn of the wheel makes the next turn easier. A frontier lead, by contrast, has to be re-won with every model generation, at ever-higher cost. One advantage feeds itself; the other has to be repurchased.
The signs of this are already visible at the edges of the enterprise market. Even inside the United States, some of the fastest-growing AI vendors are inference platforms that give companies cheap access to open-source models — a way to get “good enough” AI at a fraction of the cost of the frontier, especially for routine work.33 If that pattern is appearing in the most lucrative, frontier-friendly market in the world, it is appearing far faster everywhere else.
The metrics that actually capture this race are not the ones on the leaderboards. They are quieter: how many daily AI queries are running in Lagos, São Paulo, and Jakarta; what share of small and medium businesses are quietly running Qwen or ERNIE instead of GPT or Claude; how many factory floors are using Chinese computer-vision models to check products coming off the line. These numbers are harder to find and harder to headline. They are also the ones that determine whose technology becomes the default.
The Spectator Decides
Here is where the “spectator” in the title turns out to be the most important figure in the whole race — because the spectator is not actually watching. The spectator is choosing.
The countries that will decide this contest are not the United States and China. They are the large, non-aligned nations watching from the stands: India, Indonesia, Brazil, Nigeria, and dozens of others. Their AI futures are not yet locked in. They are shopping. And their choice is not ideological — it is brutally practical. They are not asking which country shares their values. They are asking which AI is cheap enough to deploy, open enough to build on, and available without a permission slip from a foreign government.
On every one of those questions, China’s integration strategy is currently the better fit. A startup in Nairobi can download an open Chinese model today and ship a product tomorrow, with no license, no export-control review, and no State Department sign-off.34 The American offering, by contrast, increasingly comes wrapped in conditions: export licenses, chip restrictions, and alliances built around shared values. Those conditions work fine for close allies. They are precisely the wrong sales pitch for a cost-sensitive entrepreneur in the Global South who simply wants something that works and does not cost much.
This is the quiet victory in a single image. In a marathon, the runner who is handed water at every station, by every country along the route, has a structural advantage over the runner who can only drink from the cups his allies hold out. America’s strategy relies on control. China’s relies on cooperation, or at least on the absence of friction. Control works on the people who already agree with you. Cooperation works on everyone else, and everyone else is most of the world.
There is a real cost hidden in this choice, and it would be dishonest to ignore it. If the Global South standardizes on Chinese models because they are eight times cheaper, those countries inherit a technology stack with long-term dependencies baked in — a phenomenon some call digital colonialism. The data, the defaults, and the architecture would tilt toward Beijing in ways that are hard to reverse later.35 But this is a 2035 problem, and the spectators are making a 2026 decision. When you are a small company trying to ship a product this quarter, the dependency you might regret in a decade loses every argument with the price you can afford today. Shipping beats sovereignty, at least at the moment of purchase. That is exactly how lock-in always works.
Three Futures
None of this is settled. It is worth laying out, plainly, the three ways the next several years could go.
In the first future, America’s compute advantage holds. Small modular reactors and grid upgrades finally unlock the power the data centers need; the enormous capital spending starts to pay off; and the frontier lead hardens into a genuine economic moat that integration eventually follows, because the best model becomes cheap enough and easy enough to deploy that the rest of the world adopts it after all. This is the future the American strategy is betting on.
In the second future, integration wins. Cost, ease of deployment, and the absence of strings turn Chinese models into the world’s default infrastructure. The frontier lead becomes a curiosity — genuinely impressive, and genuinely irrelevant, if 80 percent of the world’s AI users never once touch a frontier American model because a cheaper open one already does what they need. This is the quiet-victory scenario, and the flywheel of compounding adoption is what makes it plausible.
In the third future, nobody wins cleanly. A regulatory shock, a safety incident, or the bursting of the capital-spending bubble freezes both giants at once. In that world, the open-source collectives — the loose international communities that build and share models freely — inherit the market by default. It is worth noting that even in this messy outcome, China’s strategy is better positioned, because it is already built around open weights and cheap deployment rather than around a single, expensive, proprietary frontier.
Two of these three futures favor integration. That is not a coincidence. It is a reflection of which kind of advantage survives contact with the real world.
Conclusion
The race is not over, and it would be foolish to declare a winner. The United States genuinely leads on the metric it has chosen to compete on. Its models are the best, its labs are the most advanced, and its capital is the deepest.
But it is worth being precise about what that lead is, and what it is not. America is ahead on frontier capability — a metric that may not decide the race. China is ahead on the metrics that quietly accumulate into permanence: users, developers, and countries shipped. The American strategy produces the most impressive trophy. The Chinese strategy fills the stadium. And a trophy for the best model means very little if the stands are empty and everyone else is already watching a different game on a screen that runs on someone else’s stack.
The deepest insight in this whole contest is also the simplest. Frontier capability without integration is a trophy with no audience. The country that builds the smartest AI and the country that gets AI into the most hands are running two different races — and only one of those races, the unglamorous one about plumbing and price and access, builds the kind of advantage that feeds itself and is hard to take away. America is winning the race everyone is watching. China may be winning the one that counts.
Notes
Citations and source links, in order of appearance. Figures reflect reporting as of late May 2026.
1. “AI Trends (May 2026),” LLM Stats, May 26, 2026, https://llm-stats.com/ai-trends; “AI Models in 2026: Which One Should You Actually Use?,” GuruSup, May 2, 2026, https://gurusup.com/blog/ai-comparisons.
2. Britney Nguyen, “Big Tech Is Spending up to $725 Billion on AI This Year,” Business Insider, April 29, 2026, https://www.businessinsider.com/big-tech-earnings-microsoft-ai-investment-capex-plan-2026-4; “Big Tech’s AI Spending Plans Reach $725 Billion,” Tom’s Hardware, April 30, 2026, https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion.
3. “‘Magnificent 7’ Earnings Rush Reveals AI Spending Surge, With Hyperscaler Capex Set to Reach $725 Billion in 2026,” Yahoo Finance, April 29, 2026, https://finance.yahoo.com/markets/article/magnificent-7-earnings-rush-reveals-ai-spending-surge-with-hyperscaler-capex-set-to-reach-725-billion-in-2026-224901707.html.
4. “AI Trends (May 2026),” LLM Stats, https://llm-stats.com/ai-trends; “GPT-5 in 2026: Features, Benchmarks, Pricing,” Runbear, May 15, 2026, https://runbear.io/posts/gpt-5-explained.
5. “DeepSWE Blows Up the AI Coding Leaderboard, Crowns GPT-5.5,” VentureBeat, May 26, 2026, https://venturebeat.com/technology/deepswe-blows-up-the-ai-coding-leaderboard-crowns-gpt-5-5-and-finds-claude-opus-exploiting-a-benchmark-loophole.
6. “Artificial Intelligence (AI) in Military Market,” MarketsandMarkets, May 19, 2026, https://www.marketsandmarkets.com/ResearchInsight/ai-in-military-driven-by-ai-enabled-isr.asp.
7. Ibid.
8. Drawn from the author’s working notes; consistent with reporting on single-site training power demand. See “How AI Data Centers Are Reshaping Electronic Component Supply,” Accuris, May 27, 2026, https://accuristech.com/blog/ai-data-center-electronic-component-supply/.
9. “The Laws That Stop a Data Center From Coming to Your Town (or State),” The Existentialist Republic (Substack), May 12, 2026, https://cmarmitage.substack.com/p/the-laws-that-stop-a-data-center.
10. “Can $600 Billion in Capital Expenditure Support the AI Narrative?,” FuTu News, April 28, 2026, https://news.futunn.com/en/post/72168312/mag-7-earnings-week-ahead-can-600-billion-in-capital; “Big Tech’s $700 Billion AI Spending Spree Has No Clear End in Sight,” Fortune, April 30, 2026, https://fortune.com/2026/04/30/big-tech-hyperscalers-will-spend-700-billion-on-ai-infrastructure-this-year-with-no-clear-end-in-sight-eye-on-ai/.
11. “China’s AI Talent Returns Home, Accelerating Global Competition,” LinkedIn, May 14, 2026, https://www.linkedin.com/posts/keith-king-03a172128_chinas-reverse-brain-drain-is-accelerating-activity-7460654745662418945-Mb_o.
12. “Anthropic Finally Beat OpenAI in Business AI Adoption,” VentureBeat, May 13, 2026, https://venturebeat.com/technology/anthropic-finally-beat-openai-in-business-ai-adoption-but-3-big-threats-could-erase-its-lead; U.S. Census Bureau Business Trends and Outlook Survey, December 2025–May 2026.
13. “9 Best Open-Source AI LLMs in 2026, Ranked for Real Work,” Taskade, May 23, 2026, https://www.taskade.com/blog/open-source-llms; “Best Open Source LLMs in 2026: Benchmarks, Licenses,” AceCloud, May 13, 2026, https://acecloud.ai/blog/best-open-source-llms/.
14. “LLM API Pricing Comparison & Cost Guide (May 2026),” CostGoat, May 28, 2026, https://costgoat.com/compare/llm-api; “LLM Pricing: Top 15+ Providers Compared,” AIMultiple, May 8, 2026, https://aimultiple.com/llm-pricing.
15. “9 Best Open-Source AI LLMs in 2026,” Taskade, https://www.taskade.com/blog/open-source-llms.
16. “US AI Models Outperform China’s in Latest Benchmark,” LinkedIn (summary of analyst commentary), May 1, 2026, https://www.linkedin.com/posts/ziaukhan_1-united-states-has-roughly-a-seven-month-activity-7455871271780716544-iLRZ; “The Gap Closes Again—and This Time It’s on Chinese Silicon,” Trilogy AI Center of Excellence (Substack), April 29, 2026, https://trilogyai.substack.com/p/the-gap-closes-again-and-this-time.
17. “9 Best Open-Source AI LLMs in 2026,” Taskade, https://www.taskade.com/blog/open-source-llms; “The Gap Closes Again,” Trilogy AI Center of Excellence, https://trilogyai.substack.com/p/the-gap-closes-again-and-this-time.
18. Drawn from the author’s working notes on Chinese industrial policy and state coordination; consistent with reporting on Beijing’s push toward domestic chips and coordinated investment. See “Big Chinese Tech Firms Scramble to Secure Huawei AI Chips After DeepSeek V4 Launch” (Reuters), summarized at https://www.linkedin.com/posts/nicolaschaillan_there-you-have-it-huawei-just-announced-activity-7456694720359067648-KpCy.
19. “Artificial Intelligence (AI) in Military Market,” MarketsandMarkets, https://www.marketsandmarkets.com/ResearchInsight/ai-in-military-driven-by-ai-enabled-isr.asp; concept of civil-military fusion drawn from the author’s working notes.
20. “ByteDance Weighs Capex of As Much As $70 Billion in AI Push,” Bloomberg, May 27, 2026, https://www.bloomberg.com/news/articles/2026-05-27/bytedance-weighs-capex-of-as-much-as-70-billion-in-ai-push.
21. “AI Export Controls Are Not the Best Bargaining Chip,” Chatham House, April 28, 2026, https://www.chathamhouse.org/2026/04/ai-export-controls-are-not-best-bargaining-chip.
22. “Brussels Slows Down the AI Act—and Makes It Sharper Where It Counts,” EU Insider, May 17, 2026, https://euinsider.eu/news/eu-ai-act-digital-omnibus-deal-2026; “Breaking: The EU AI Act Just Changed,” LinkedIn, May 8, 2026, https://www.linkedin.com/pulse/breaking-eu-ai-act-just-changed-heres-what-compliance-shivendra-yadav-v77sc.
23. “Breaking: The EU AI Act Just Changed,” LinkedIn, May 8, 2026, https://www.linkedin.com/pulse/breaking-eu-ai-act-just-changed-heres-what-compliance-shivendra-yadav-v77sc.
24. “AI Export Controls Are Not the Best Bargaining Chip,” Chatham House, https://www.chathamhouse.org/2026/04/ai-export-controls-are-not-best-bargaining-chip.
25. “Huawei Surpasses $12B in AI Chip Revenue” (citing Financial Times reporting), LinkedIn, May 3, 2026, https://www.linkedin.com/posts/nicolaschaillan_there-you-have-it-huawei-just-announced-activity-7456694720359067648-KpCy; “AI: Huawei’s AI Gains in China Over Nvidia,” AI: Reset to Zero (Substack), May 4, 2026, https://michaelparekh.substack.com/p/ai-huaweis-ai-gains-in-china-over.
26. “Nvidia Says It Has ‘Largely Conceded’ China’s AI Chip Market to Huawei,” TechBuzz, May 20, 2026, https://www.techbuzz.ai/articles/nvidia-says-it-has-largely-conceded-china-s-ai-chip-market-to-huawei.
27. “AI: Huawei’s AI Gains in China Over Nvidia,” AI: Reset to Zero, https://michaelparekh.substack.com/p/ai-huaweis-ai-gains-in-china-over.
28. Information Technology and Innovation Foundation, “Backfire: Export Controls Helped Huawei and Hurt U.S. Firms,” cited in “Huawei Surpasses $12B in AI Chip Revenue,” LinkedIn, https://www.linkedin.com/posts/nicolaschaillan_there-you-have-it-huawei-just-announced-activity-7456694720359067648-KpCy.
29. “The Laws That Stop a Data Center From Coming to Your Town (or State),” The Existentialist Republic, https://cmarmitage.substack.com/p/the-laws-that-stop-a-data-center; “Singapore Is Expected to Use Nearly 20% of Its National Grid to Power Data Centers in 2026,” Nikkei Asia.
30. Drawn from the author’s working notes on China’s “East Data West Compute” strategy; consistent with reporting on direct green-power supply to Chinese data centers. See “Direct Green Power to Data Centers Gains Traction,” LinkedIn, May 7, 2026, https://www.linkedin.com/posts/howard-h-4b9757297_greenpower-datacenters-solarenergy-activity-7458111607089369088-ONYj.
31. “UAE Receives Nvidia AI Chips for Large-Scale Deployment,” LinkedIn, May 9, 2026, https://www.linkedin.com/posts/smashi-business_uae-activity-7458820254392102912-qqlg; reporting on Saudi Arabia’s Humain venture, May 25, 2026, https://www.instagram.com/p/DYwmljViGst/.
32. Nguyen, “Big Tech Is Spending up to $725 Billion”; “AI Trends (May 2026),” LLM Stats, https://llm-stats.com/ai-trends.
33. “Anthropic Finally Beat OpenAI in Business AI Adoption,” VentureBeat, https://venturebeat.com/technology/anthropic-finally-beat-openai-in-business-ai-adoption-but-3-big-threats-could-erase-its-lead.
34. Reasoning drawn from the author’s working notes; supported by the open-weight, permissive-license analysis in “9 Best Open-Source AI LLMs in 2026,” Taskade, https://www.taskade.com/blog/open-source-llms.35. Concept of “digital colonialism” and stack lock-in drawn from the author’s working not