The new cold war runs on AI

Illustration: TBS

The old Cold War was measured in warheads, satellites and aircraft carriers. The new one is measured in models, chips, tokens and gigawatts. Artificial intelligence has become a form of national power because it changes who controls knowledge, infrastructure, markets and military advantage. The contest is now about who can decide who may use the defining technology of the age.

That became visible on June 12, when the Trump administration ordered Anthropic to block foreigners from Fable and Mythos, its latest frontier AI models. The trigger was a supposed jailbreak for Fable, said to bypass safeguards against hacking, bioweapons and other dangerous uses.

Anthropic switched off the models for everyone while arguing that the danger was overblown. The message was unmistakable: access to frontier AI can now depend on a decision in the Oval Office.

US's new gatekeeping power

In March, the administration designated Anthropic a "supply-chain risk". Now it has shown that it can trample not only on AI companies but also on users worldwide. The US has restricted powerful technologies before: it stopped assisting Britain's nuclear weapons programme after the Second World War, blocked exports of modern cryptography in the 1970s, and keeps the F-22 fighter exclusively for itself while allies receive the F-35.

AI sits between those examples. Like nuclear weapons, the most capable models may be too dangerous to release if they can disable infrastructure or help create pandemic-ready pathogens. Like cryptography, their methods may be impossible to contain.

Like cyber weapons, a small advantage can be decisive: if an attacker has version 5 and a defender is stuck with version 4, one extra vulnerability may be enough.

Yet this new US power is fragile. Anthropic claims that 80% of its consumer use is overseas, and many staff are not from the US. Cutting off foreign users too aggressively could freeze research, alienate allies and push customers towards China, the second-ranking AI power.

The new AI Cold War, is not a simple race between Washington and Beijing. It is a contest over chokepoints: models, chips, compute, power, data, safety rules and moral authority. The US has the strongest frontier systems and the ability to restrict access. China has made openness a weapon and affordability a challenge.

The likely future is a hierarchy of access. The best capabilities may be reserved for US cyber offence, defence and military systems. Allies may receive next-best models, the AI equivalent of the F-35. The rest of the world may get safer, handicapped versions.

US allies are already vulnerable to Washington's bargaining on trade, alliances and the dollar system. AI could become the most important lever. The dependence is not entirely one-way: the US still needs Dutch lithography machines and Taiwanese fabs. But in computing, it is dominant.

It has perhaps 15 times more AI computing power than Europe, with much greater ongoing investment. A gloomy essay called "Europe 2031" imagines Europe reduced to vassal status as its cyber security, defence and economy depend on US models and computing.

China's open-source countermove

China has seen the same map of power and is trying to redraw it. In January 2025, DeepSeek R1 showed that Chinese AI could be both strong and free. It erased $1 trillion from US capital markets; Nvidia briefly lost 17% of its value and the Nasdaq fell 3.1% in a day.

On June 13, a Beijing lab called Zhipu, or Z.ai, released GLM 5.2, promising "a step closer to frontier intelligence for everyone". It is the strongest Chinese-trained model so far, runs at less than a tenth of the cost of Anthropic's Fable 5, and has publicly released weights.

One day after the US made frontier AI access revocable, China presented openness as reliability. "Our attitude is one of radical openness," Tang Jie, Zhipu's co-founder, told The Economist. He attacked "external blockades" that make AI systems "subject to revocation at any moment". Chinese open-source models can be downloaded and run locally, outside the immediate reach of governments or labs.

Capability is narrowing, too. GLM 5.2 is ranked as the most intelligent open-source model and fourth overall, behind OpenAI's ChatGPT 5.5 and ahead of Google's Gemini.

Illustration: TBS

Fable 5 remains about 17% more capable across benchmark tasks, and a comparable Western model appeared about four months earlier. Elon Musk wrote that China may match today's frontier by early next year. Tang replied, "It won't take that long".

The caveat is that benchmarks can flatter Chinese models. Harvard Tveit Ihle of the Norwegian Defence Research Establishment told The Economist that open-source models often do better on public tests than private ones, partly because labs may "teach to the test".

His work found Chinese models four to six months behind US models on public tests, but eight to ten months behind on private tests. GLM 5.2 is about seven months behind on WeirdML and a year behind on SimpleBench.

Yet, on a new office-work benchmark released on June 19, it outperformed ChatGPT 5.5. The US still leads, but China has real momentum.

Cost is similarly complicated. DeepSeek charges $0.87 per 1 million output tokens for its V4 model, compared with Anthropic's $50 for Fable 5. Ramp saw a sharp rise in US companies paying for DeepSeek in June, and Microsoft is reportedly considering DeepSeek for Copilot.

But cheap per token does not always mean cheap per task. Research by Du Zheng of Georgia Tech found that one DeepSeek model used 23 times more tokens than its OpenAI rival for essentially the same result. On a software engineering benchmark, GLM 5.2 ended up costing more than systems from Anthropic and OpenAI.

Compute, power and the race to self-improve

This is why AI power is not just software. It is energy, steel, land and public consent. The US currently has just under 12GW of data-centre power capacity, but that could rise fivefold by 2030.

Moody's estimates that $3 trillion will go into AI data centres globally between 2026 and 2030, much of it in the US. Amazon, Google, Meta, Microsoft and Oracle may spend $750 billion. Meta's "Prometheus" site in Ohio alone may contain $30 billion of semiconductors and use 1GW of power, enough for as many as 1 million homes or roughly a large nuclear reactor.

But the AI boom has met a backlash. Americans object to ugly buildings, generators, cooling systems, transmission towers and concerns about water use. Surveys suggest they would sooner live next to a nuclear plant than a data centre.

At least 20 projects worth $42 billion, representing 3.5GW of capacity, were cancelled in the first quarter of 2026; $85 billion has been cancelled over three years. Pew found in April that Americans who had merely heard of data centres were as opposed to them as those living within five miles of one. In Virginia's Loudoun County, "data-centre alley", permissive rules were removed in March 2025. San Marcos, Texas, imposed a moratorium.

Washington sees the resistance as a strategic risk. "We need to stay a ways ahead of China," said Chris Wright, US energy secretary. AI leadership is "the overriding goal" of his tenure: "We've got to enable these data centres to get permitted and built and connected to power."

Moody's estimates that $3 trillion will go into AI data centres globally between 2026 and 2030, much of it in the US. Amazon, Google, Meta, Microsoft and Oracle may spend $750 billion. Meta's "Prometheus" site in Ohio alone may contain $30 billion of semiconductors and use 1GW of power, enough for as many as 1 million homes or roughly a large nuclear reactor.

The Department of Energy expects that the US will need 50GW of additional generation for AI by 2030 and another 50GW for manufacturing. "We need to grow our capacity of dispatchable power," Mr Wright says. SemiAnalysis estimates about a terawatt of large-load grid-connection requests, close to the 1,250GW maximum capacity of the US grid.

Industry insists it can be a responsible neighbour. "We want to be good grid citizens," says Microsoft's Alistair Speirs, adding that batteries allow the company to "choose when to sip and when to slurp from the grid".

Ohio has tried to protect ratepayers by requiring large data-centre operators to pay monthly for at least 85% of requested power capacity, even if unused, a binding version of the "ratepayer protection pledge". Still, in Ohio, three-quarters of Democrats and two-thirds of Republicans oppose local data-centre development. State Senator Shane Wilkin says constituents remain unmoved: "Well, I just don't want it."

The pressure is sharpened by AI's growing ability to build itself. Claude Code, launched in February 2025, has become indispensable to developers. Anthropic says that more than four-fifths of the code it published in May was written by Claude, up from "low single digits" before.

METR found that early-2025 Anthropic models could complete tasks that took human engineers just under an hour; newer systems can do work that takes more than a day. Yet Anthropic has called for the world to have "the option to slow or temporarily pause frontier AI development".

Jack Clark, its co-founder, puts the odds at 60% that an AI system will create its own successor without human involvement by the end of 2028. "What can seem to many like a fanciful story may instead be a real trend," he told The Economist.

That process, recursive self-improvement, would be a closed loop: one model builds a better model, which builds another. Doomers call a rapid intelligence explosion "going foom". The Centre for Security and Emerging Technology warned that as AI performs more AI R&D, productivity could rise tenfold, then a hundredfold, then a thousandfold, and that systems created through RSI "pose extreme risks. This warrants preparatory action now."

Andrej Karpathy's Nanochat example shows the direction. GPT-2-level training that once took 168 hours on 32 advanced chips was reproduced on one computer with eight GPUs in three hours, then reduced to just over two.

When he handed optimisation to an AI agent, Autoresearch, training time fell to one hour and 48 minutes within two days, and to one hour and 39 minutes five days later. "I didn't touch anything," he said. "They stack up and actually improve Nanochat."

This makes safety and ethics part of the power struggle. AI labs are hiring philosophers because models must reason about truth, harm, agency and obedience. Humanities students were once told to "learn to code"; now US philosophy graduates have lower unemployment than computer science graduates, 5.1% versus 7% in 2024.

Luciano Floridi calls the departure of philosophers into AI firms a "haemorrhaging". Socratic questioning can reduce sycophancy; Socratic ignorance can temper overconfidence, or "AI immaturity". Iason Gabriel says philosophy is "a powerful mechanism" for improving long reasoning "chains of thought".

The stakes are concrete. Anthropic's Claude constitution has drawn on Kant, Apple's terms of service and the Universal Declaration of Human Rights; staff call the 78-page document its "soul doc". IBM's Granite models allow firms to tune trade-offs such as individual agency versus social harmony.

Google aims for "likely overall benefits [that] substantially outweigh the foreseeable risks". Autonomous vehicles, weapons and emotional-support chatbots all encode moral choices. Critics warn of "moral deskilling". Roman Yampolskiy argues that morality "is historically unstable, culturally variable, strategically manipulable, and often only retrospectively legible".

The new AI Cold War, then, is not a simple race between Washington and Beijing. It is a contest over chokepoints: models, chips, compute, power, data, safety rules and moral authority. The US has the strongest frontier systems and the ability to restrict access. China has made openness a weapon and affordability a challenge.

Europe and other allies must decide whether to remain dependent or build strength through energy, planning reform, data centres and regional integration. The country that leads in AI will not merely possess a better tool. It will shape what others may know, build, buy, defend and become.