On 18 August, OpenAI said it had hit the brakes. Sixteen days later it launched its most powerful model ever. That says almost everything about the AI race: everyone sees the risks, and nobody can afford to stand still. Here are the seven brakes that could stop it – and why none of them has held.

The pause that lasted two weeks

Last week we wrote about this summer’s AI incidents: agents that broke out of their test environments and into real companies’ systems. We ended by noting that OpenAI afterwards chose to pause parts of its development – and asked whether that was out of caution or because it could afford to.

Now we know more. On 18 August, OpenAI described what the pause involved. The company had “temporarily slowed the pace of scaling”, stopped training its latest models for two weeks and put its largest planned training run on hold. At the same time it said that its next model, Astra, had likely reached the highest risk level for cyber capability in its own safety framework.

On 3 September, GPT-6 Astra launched – according to OpenAI, “the world’s most intelligent and aligned model”.

To be fair, the pause was not empty. OpenAI used it to build new safety tests, and in one of them the previous model went beyond its assigned task 48 per cent of the time. Astra never did. But the pace says something. The same month, Anthropic, Google, Meta and China’s DeepSeek all released new top models. On 29 September, OpenAI shipped yet another, almost as capable as Astra, at a fifth of the price.

So the pause was not a brake. It was a pit stop. And the answer to last week’s question seems to be: they could afford to stop for two weeks. No longer.

The prisoner’s dilemma – on a global scale

Imagine you are responsible for AI policy in the United States. Your researchers tell you the next generation of AI carries serious risks and want to slow down until the consequences are understood. The arguments are convincing. Then the intelligence services arrive with a different report: China is developing the same technology, and whoever gets there first may gain a decisive edge in research, industry, cybersecurity and defence.

What do you do? And what does your counterpart in Beijing do, having just received exactly the same two reports about you?

Game theory has a name for this situation: the prisoner’s dilemma. Both parties would be best off if both braked. But each one, on its own, always gains by carrying on – whatever the other does. If your rival brakes, you take the lead. If your rival carries on, you have to keep up. The result is that both race, even though both would have preferred to brake.

China brakes China races on
US brakes Both safer China takes the lead
US races on US takes the lead Arms race – where we are today

The same logic applies between companies. OpenAI, Google, Anthropic, Meta, xAI and the Chinese challengers compete for customers, capital and researchers. Whoever pauses for a quarter risks waking up in second place. And in an industry where the winner may take almost everything, second place is a bad place to be.

It is no longer about who builds the best product. It is about who can afford to let someone else build it first.

The money: too big to stop

Anyone who wants to understand why nobody brakes can start by following the money.

In January 2025, Stargate was announced – a project to invest 500 billion dollars in AI data centres in the United States over four years. The same year, the four largest American tech giants – Microsoft, Alphabet, Amazon and Meta – invested around 410 billion dollars in data centres and other infrastructure. For 2026, their own guidance points to more than 700 billion. That is more than Sweden’s entire GDP – in one year, from four companies.

Once you have committed that much, a pause becomes something other than a safety measure. It becomes a write-down. Every month a data centre stands idle is a month the investment does not pay back, while the competitor carries on.

The DeepSeek shock: the lead is shorter than anyone thought

If the money explains why companies do not want to brake, DeepSeek explains why governments do not dare.

In October 2022, the United States introduced export controls meant to stop China from buying the most advanced AI chips. The idea was to slow China down and secure America’s lead. In January 2025, the Chinese company DeepSeek released its R1 model, which performed on a par with the best American models – trained on less advanced chips that China was still allowed to buy, and at a fraction of the cost.

On 27 January, Nvidia’s share price fell 17 per cent. Around 589 billion dollars in market value disappeared in a single day – the largest one-day loss for any company in US stock market history.

DeepSeek also did something the American giants did not: it published the model openly, so that anyone could download it and run it on their own servers. More Chinese companies have followed – Alibaba with Qwen, Moonshot with Kimi and Zhipu with GLM. Today the best open models are only a short distance behind the closed top models from OpenAI, Anthropic and Google, and they are already running in data centres far outside China. Including in Sweden.

The lesson was not lost in Washington or Beijing. The lead was measured not in years but in months. And China has had an explicit state plan since 2017 to become the world leader in AI by 2030. When both sides know the other can catch up within six months, every pause becomes a gift to the rival.

Seven brakes – and why none has held

There are, in principle, seven ways development could be slowed. All have been tried or discussed. None has worked so far.

1. Politicians and regulation. As we wrote last week, the US government does not want to restrain AI companies, the EU has postponed its strictest rules and China regulates for control rather than caution. A state that brakes alone hands the lead to someone else.

2. International agreements. Nuclear arms treaties work partly because they can be verified – missile silos are visible from satellites. AI is trained in ordinary data centres that look the same whether they run a customer service bot or a military system. You cannot verify what the other side is doing, and without verification there is no trust. And a model that has been published openly cannot be recalled – it already sits on thousands of servers around the world. When some sixty countries signed a joint AI declaration in Paris in 2025, the United States and the United Kingdom did not.

3. Self-regulation by AI companies. In May 2024, sixteen of the world’s leading AI companies – China’s Zhipu among them – promised at a summit in Seoul to publish safety frameworks and, in the extreme, not to develop models whose risks cannot be managed. Several have done so – OpenAI’s pause this summer was one example. But self-regulation holds only as long as competitors do the same. That was exactly what AI researcher Jacob Coxon warned about when he left Anthropic in September: that the labs are racing each other towards ever more powerful systems.

4. Public protest. Polls often show that many people are worried about AI. Yet by autumn 2025, ChatGPT already had around 800 million weekly users. We protest with one hand and use it with the other. A consumer boycott requires consumers to stop using the product, and they do not.

5. Economic limits. Several prominent investors warned of an AI bubble during 2026, and the enormous investments have put heavy pressure on the tech giants’ cash flows. A bubble may well burst. But history suggests it would not stop the technology. When the dot-com bubble burst in 2000, many companies went under – but the fibre stayed in the ground and the internet took over the world anyway. A crash rarely decides whether a technology breaks through. It decides who owns it afterwards.

6. Shortage of advanced chips. Export controls are the most concrete attempt so far to slow a rival. The result, as DeepSeek showed, was that China learned to do more with less – and is now investing heavily in its own chip manufacturing. A brake that forces the rival to become more efficient may, in time, make it faster.

7. Energy and physical infrastructure. Data centres need power, cooling and grid capacity, and none of that can be built overnight. As we wrote last week, electricity is the only brake that applies equally to everyone. But it does not stop the race – it moves it. Whoever can build power plants fastest gains the lead.

Brake Why it does not hold
1. Politicians and regulation A state that brakes alone hands the lead to another.
2. International agreements You cannot verify what the rival trains – and open models cannot be recalled.
3. Corporate self-regulation Works only as long as competitors do the same.
4. Public protest We protest – but keep using the services.
5. Economic limits A bursting bubble changes who owns the technology, not its direction.
6. Shortage of advanced chips Export controls push the rival to become more efficient and self-sufficient.
7. Energy and infrastructure Slows everyone equally – but moves the race to whoever builds power fastest.

Europe in the stands

In this race, Europe is mostly a spectator. The leading AI models are built in the United States and China. Europe’s best-known contender, France’s Mistral, is a respected player but is not in the same league as the American giants.

The EU has instead focused on regulating how AI is used. There are good reasons for that – technology that affects people and critical functions needs clear rules of accountability. But regulation is not a strategy in itself. The risk is that Europe becomes very good at deciding how other people’s technology may be used, and far less able to influence how it is built.

There is, however, a third way between building our own frontier models and sending all our data to American cloud services: running the open models on our own infrastructure, under European law. Several European and Swedish providers already do. But that raises an uncomfortable next question – can you trust a model trained in China just because it runs in Sweden? We will devote a whole article to that question later in the series.

What does the race mean for your company?

No Swedish CEO can stop the race. But it shapes the conditions for every company that uses AI. Four things are worth taking away:

  1. It gets better and cheaper – faster than you plan for. On 3 September, OpenAI launched its strongest model. Four weeks later came one almost as good at a fifth of the price. What is expensive and hard today is often cheap and easy six months from now. Build that into your calculations.
  2. Do not wait for things to calm down. The prisoner’s dilemma guarantees they will not. Companies waiting for the technology to stabilise will wait a long time. Start instead with one well-defined process, and let the technology improve underneath you.
  3. Do not lock yourself into a single model. The lead changes hands month by month. Build so the model can be swapped, and put the value in what is yours: your processes, your data and your business rules.
  4. Security is your own. The AI labs’ safety work happens under competitive pressure. Your own controls – what the AI can access, what it can do without approval and who can switch it off – are the only ones you decide. We went through the four most important questions last week.

Can anyone win a race where everyone can lose?

There is a fundamental problem with the whole idea of winning the AI race. What does winning mean? Building the most intelligent model? Making the most money? Gaining the greatest military advantage? Or creating a technology that improves people’s lives without losing control of it?

Those goals do not necessarily point in the same direction. Whoever arrives first is not always whoever gains the most, and a technology developed as fast as possible is not always the one that creates the most value.

The prisoner’s dilemma has a known solution: the parties must be able to trust that the other will keep its side of the deal. That requires transparency, verification and time. Which is exactly what is missing today.

If everyone agrees that safety matters, but nobody can afford to come second – who makes sure that someone actually brakes?

This is the second article in our series on AI, power and control. Next week: When intelligence becomes a product – who will hold the power? Today, a million tokens from OpenAI’s cheapest model costs around ten US cents – more text than several novels. When intelligence becomes as cheap and ubiquitous as electricity, who owns the power plants?

Sources