US vs China: Inside the Race for AI Dominance — and the Safety Problem Neither Side Can Solve Alone
Business Two governments now treat artificial intelligence the way earlier generations treated nuclear capability: as the single technology most likely to decide who sets the rules for the rest of the century. One of them has the best models and the deepest capital markets. The other has the manufacturing base, the energy, and a state apparatus willing to spend for a decade without asking for quarterly returns.
And here’s the uncomfortable part. The leaders of both countries are meeting against a backdrop of rising concern about what rogue AI could do to everyone — while the competition between them keeps getting in the way of any common ground on safety.
This article breaks down where the race actually stands in 2026, which metrics matter, which ones are noise, and why the “who’s winning” framing may be the wrong question entirely.
Where the Race Actually Stands in 2026
The scoreboard has changed dramatically in three years, and most public commentary is still running on 2023 assumptions.
The Capability Gap Has Nearly Closed
This is the headline number. The US spent roughly 23 times more than China on private AI investment in 2025 — about $285.9 billion versus $12.4 billion — and yet the performance gap between the best American and Chinese models had narrowed to under three percentage points by early 2026. In May 2023, that gap sat somewhere between 17.5 and 31.6 points. Digital in Asia
Read that twice. A gap that looked structural in 2023 is now within rounding distance, despite a spending mismatch of more than an order of magnitude.
American labs still hold the absolute frontier. OpenAI, Anthropic, and Google DeepMind remain ahead on raw capability, but DeepSeek, Alibaba’s Qwen, and Baidu’s ERNIE have rewritten expectations on a fraction of the budget — while cut off from the world’s best chips. Digital in Asia
Compute Still Favors Washington
Money and silicon are where the US lead is real.
One analyst at the Institute for Progress estimates US computing power sits at roughly ten times China’s. Between them, the two countries control about 90% of global computing power and attract 70% to 80% of worldwide AI investment. BloombergRest of World
But Beijing is buying its way out of the constraint. China is preparing to spend roughly 2 trillion yuan — about $295 billion — over five years building data centers, boosting domestic chipmaking, and reducing dependence on Nvidia and AMD, while simultaneously ramping up power generation heavily weighted toward wind and solar. Bloomberg
The Chip Story Is Changing Fast
Export controls were meant to freeze China’s hardware progress. They slowed it and accelerated substitution at the same time.
Domestic Chinese chips made up nearly 41% of China’s market in 2025, with roughly half of those sales from Huawei — a sharp reversal from Nvidia’s 90%-plus share before 2023. Huawei’s Ascend 950PR chips are expected to reach 750,000 units this year with a more CUDA-compatible architecture, making migration easier for Chinese developers, while Cambricon plans to deliver 500,000 domestically manufactured accelerators in 2026. Brookings
CUDA compatibility matters more than the raw specs. It lowers switching costs for an entire developer ecosystem.
Two Strategies, Not One Race
Calling this “a race” implies both runners are on the same track. They aren’t.
The American Model: Private Capital and Frontier Research
Washington leads on frontier model innovation, powered by private-sector dynamism and research depth. The strength is speed, risk tolerance, and an investor base willing to fund experiments that may never return capital. Yahoo Finance
The weakness is coordination. There’s no central plan for energy buildout, grid capacity, or chip fabrication timelines — those get negotiated company by company, state by state.
The Chinese Model: Full-Stack and State-Directed
China is pursuing a full-stack approach spanning chips, compute infrastructure, foundation models, and applications. Beijing leverages state-led scaling and manufacturing control, benefiting from lower energy costs and a centralized grip on the critical minerals that hardware production depends on. BrookingsYahoo Finance
The strength is sequencing — you can plan a fab, a grid upgrade, and a model release as one program. The weakness is that centralized bets fail centrally.
The Stacks Are Drifting Apart
Each country is building an AI technology stack designed to reduce dependence on the other, which also makes those stacks increasingly incompatible — meaning CEOs and policymakers worldwide may have to pick a side sooner than they expect. BCG
For businesses outside both countries, that’s the practical consequence. Not who wins, but which stack you’re locked into. If you’re evaluating where to build, our guides on technology trends and global business strategy are worth a read.
The Metrics Everyone Misreads
Three numbers get quoted constantly and mean less than people assume.
- Total investment. Useful for measuring intent, poor at predicting output. The 23:1 spending ratio produced a sub-3-point capability gap.
- Patent counts. Chinese entities filed 69.7% of all AI patents worldwide, and China leads on publication volume — 23.2% of global AI papers versus 12.6% for the US. Volume isn’t the same as commercial impact. Digital in Asia
- Benchmark leaderboards. They shift monthly and measure narrow tasks, not deployed economic value.
The Metric That Should Worry Washington
Talent flow. The number of AI researchers moving to the US has dropped 89% since 2017, with an 80% decline in the last year alone. Nearly all the researchers behind DeepSeek’s five foundational papers were educated and trained in China. Digital in Asia
The brain drain that built Silicon Valley’s advantage is running in reverse. That’s a slow variable with a very long lag — and it doesn’t show up on any benchmark.
The Safety Problem Neither Side Can Solve Alone
Here’s the paradox at the center of the story.
Both governments privately recognize that a badly aligned or badly deployed AI system is a shared risk. Neither can address it unilaterally, because a safety restriction adopted by one side and ignored by the other simply transfers advantage.
As one Chinese official framed it, meaningful cooperation on AI safety becomes impossible if each side keeps treating the other’s technological progress as a national security threat — which is exactly why dialogue matters more, not less, when disagreements deepen.
Carnegie’s Scott Singer describes the problem plainly: the AI conversation between the two countries keeps getting pulled along by the broader ups and downs of the relationship.
Safety cooperation is technically straightforward and politically almost impossible. That gap is the actual story.
What Realistic Cooperation Could Look Like
- Incident reporting channels — a shared mechanism for disclosing serious model failures, modeled loosely on aviation safety reporting
- Evaluation standard alignment — agreeing on how dangerous capabilities get tested, even without agreeing on what to do about them
- Military AI red lines — narrow, verifiable commitments around autonomous weapons and nuclear command systems
- Track II dialogue — sustained researcher-to-researcher contact that survives diplomatic freezes
None of these require trust. They require that both sides recognize a category of outcome that’s bad for everyone.
What This Means for the Rest of the World
Most countries aren’t competing. They’re choosing.
Chinese AI offerings are winning developers and businesses on aggressive pricing. That matters enormously for cost-sensitive markets across Asia, Africa, and Latin America — where the deciding factor isn’t benchmark scores but price per token and deployment cost. Bloomberg
Meanwhile, the hardware spending is creating real winners elsewhere. Analysts have maintained an 8% GDP forecast for Taiwan in 2026 on the back of AI sector expansion, while Mexico and Korea are gaining structurally as they integrate deeper into the global AI hardware supply chain. Yahoo Finance
Practical Takeaways for Business Leaders
- Assume stack fragmentation is permanent. Build with portability in mind rather than betting on one ecosystem.
- Separate capability from cost. For most commercial workloads, a model three points behind the frontier at a fifth of the price is the better business decision.
- Watch export control changes, not headlines. Policy shifts move supply chains; summit photo-ops rarely do.
- Track talent, not just benchmarks. Where researchers move is a five-year leading indicator.
- Build compliance flexibility now. Regulatory divergence between the two blocs is widening, not converging.
Frequently Asked Questions
Who is actually winning the AI race right now?
The US leads on frontier capability, compute, and capital. China leads on research volume, patents, manufacturing integration, and cost efficiency. They’re winning different races that happen to share a name.
Did export controls work?
Partially. They constrained China’s access to top-tier chips while accelerating domestic substitution. Whether that trade was worth it is still genuinely contested.
Can the two countries cooperate on AI safety?
Technically yes, politically it’s very hard. The most realistic path is narrow, verifiable agreements rather than broad frameworks.
Will other countries build competitive frontier models?
A few may build strong regional or specialized models. Matching the capital and compute of the two leaders is a different proposition entirely.
Final Thoughts
The instinct to score this as a horse race is understandable and mostly unhelpful. The US and China have built genuinely different systems for producing AI, each with structural advantages the other can’t easily copy, and the capability gap between them is now narrow enough that neither can count on a durable lead.
What’s left is a competition where both sides are moving fast, neither fully trusts the other, and the risks that would hurt them equally are the ones getting the least attention.
That’s not a race with a finish line. It’s a relationship that has to be managed — and right now, it isn’t being managed well.
