Type “cold war in AI” into Google and the AI Overview obligingly describes a rivalry between the US and China over AI, advanced chips and digital infrastructure. The phrase is everywhere now. It gets treated as a description of the world when it’s really an analogy, and analogies are good at some things and quietly terrible at others.
So what does the label mean in practice, and how does it stack up against the real Cold War, the one that ran from 1947 to 1991? Mostly it doesn’t. The similarities are real, but the differences matter more.
How we got here
The term took off after 2017, when China published a national plan to lead the world in AI by 2030, according to the Institute for Economics and Peace. Washington answered with chip restrictions. The US announced controls on China’s semiconductor and supercomputing sectors in August 2022 and tightened them in October 2023. Since then the policy has lurched around. A sweeping Biden-era rule that would have set up licensing tiers for countries worldwide was rescinded in May 2025, and the Trump administration later approved sales of Nvidia’s H200 chip to China.
The results are a mess. sources, citing IDC data, reported that Nvidia’s share of China’s data-center AI chip market dropped to 34% in the first half of 2026, with shipments down 44% from a year earlier. Several sources reported that Washington cleared the H200 for export but Chinese regulators blocked its import. Jensen Huang has said the restrictions handed a market worth around $50 billion to Huawei.
Then there’s DeepSeek, which released its V4 model in April 2026 as open source, reportedly trained on Huawei Ascend and Cambricon chips. That’s the interesting part. It’s a test of whether China can build frontier AI without Nvidia at all.
What the analogy gets right
The core resemblance is structural. Two superpowers, one technology, and each side thinks it could decide the century. One analysis going around this year points to The Economist’s phrase “cold-war-style dilemma” and puts the US and China together at roughly 90% of the world’s frontier computing power. That number comes from a newsletter, so take it as an estimate.
Smaller countries are getting pulled toward one camp or the other, too. In the 20th century, newly independent states were pressured to line up behind Washington or Moscow. Now developing countries are being courted by both the US and China with funding, technology transfers and training programs, and they have to pick whose AI ecosystem to build on.
And the money is huge. The Institute for Economics and Peace, citing sources, lists announced AI infrastructure commitments of more than $23 billion from Europe, over $114 billion from Saudi Arabia, and somewhere between $1.14 trillion and $2.28 trillion from the United States. Those are pledges, not money already spent. Still, governments clearly treat AI as national power now, not just another tech sector.
Where it falls apart
The weapon isn’t a weapon: The Cold War was about nuclear warheads. The scarce thing now is advanced chips, and the technology built on them runs everything from customer service bots to targeting software. Foreign Affairs makes the point that nuclear weapons require specific isotopes of two elements, which can be traced and counted, while AI cuts across nearly every sector. Try imagining a nuclear-style inspection regime for that. I can’t.
The money comes from companies: Governments paid for most of the first Cold War. The leading AI systems today come from private firms: Nvidia, Google, OpenAI, Anthropic, Meta. Governments still push in through export rules, equity stakes and revenue-sharing deals, so you get a hybrid that looks different in each country. The Modern War Institute notes that private-sector and open-source development has no good historical precedent.
Secrecy flipped: Soviet and American programs were hidden. AI research is mostly published, and plenty of models can be downloaded by anyone. Everyone can see roughly what everyone else is building, so the contest moves to what’s hard to copy: chips, talent, electricity and data.
The two sides are tangled together: The US and the USSR barely traded. The US and China are deeply entwined. A Forbes analysis, relayed by the Institute for Economics and Peace, says AI rests on five layers (energy, hardware, data, models and talent) and that no country controls all five, the US and China included. China holds about 70% of global rare earth production and nearly 90% of processing, according to Down To Earth, and those minerals go into chips and defense hardware. After the October 2025 Busan meeting, Beijing paused its broadest rare earth restrictions for a year. The US-China trade truce was recently extended only to January 10, 2027, which is shorter than markets wanted.
There was no Hiroshima: This might be the biggest gap. Foreign Affairs argues the nuclear order grew out of a concrete demonstration: the bombings of Hiroshima and Nagasaki in 1945, then hundreds of tests. The Non-Proliferation Treaty of 1968 and the hotline set up after the 1962 Cuban Missile Crisis came after leaders had seen what the weapons did. AI’s worst scenarios are still hypothetical. With no shared shock, there’s less reason to build brakes, and both sides keep accelerating.
Does the label cause the race?
Some scholars wonder whether “Cold War” describes the situation or helps create it. A 2025 paper by Yujia He and Richard Heeks in St Antony’s International Review looks at the narrative itself. The Bulletin of the Atomic Scientists made a related warning in 2024: comparing AI to nuclear war pushes people into all-or-nothing thinking, the same mindset behind the most dangerous stretches of the arms race. If every advance reads as a move in a zero-sum game, cooperation on safety gets harder to justify.
Then there’s the problem of measuring who’s winning. Estimates of how far China trails the US swing widely depending on what you count. A Stanford-linked figure reported this year put the performance gap between top US and Chinese models at 2.7% as of March, while another analysis estimated an average lag of about seven months. The US leads on capital and infrastructure spending. China’s strength is cheap, widely deployed and often free models. Who’s ahead depends on the scoreboard you pick.
What it means for regular people
For most of us the rivalry shows up less as missiles and more as choices baked into everyday technology. The Institute for Economics and Peace notes that American firms mostly sell subscriptions, like $20 a month for ChatGPT Plus, while Chinese firms such as DeepSeek, Baidu and ByteDance give consumers free access and make money through ads, enterprise services and cloud bundles.
Countries that can’t make their own chips or models have to buy from one bloc, or both. A fight over chips or rare earths can raise costs and delay products a long way from Washington or Beijing. And whoever writes the standards for AI may shape what systems treat as normal, true or harmful, according to analysts cited by the Institute for Economics and Peace.
Not everyone is picking a side
France, the UAE and Singapore are described as taking a “third way”: buying computing power and tools from the leaders while building their own models and datasets. The Institute for Economics and Peace goes further, suggesting some countries could form something like a new Non-Aligned Movement for AI, building for local needs and working selectively with both powers. Whether real AI sovereignty is even possible is disputed, since every country leans on someone for at least one of those five layers.
The verdict
The Cold War comparison is good for one thing: it captures the fear that falling behind could be permanent. After that it misleads. This is a contest between interdependent economies, run largely by private companies, over a technology with no clean military use and no agreed catastrophe to fear. The original Cold War ended in arms control treaties and the collapse of one side. Nothing like that is in view here.
The better picture is a competition that’s also a dependency, each side holding something the other needs. That makes it less dramatic than a standoff and harder to resolve. And the question the analogy tends to bury may be the most important one: is anyone running this race toward something other than the race itself?
