The conversation had reached the question of who, if anyone, would call a halt. One speaker described the situation as "this sort of global economic race for the most advanced frontier intelligence" and compared it to climate change: nobody is going to stop.
Steve Keen offered a different brake. "I think what might cost it to stop it is the sheer cost of it and whether we actually have the physical resources to be able to implement that sort of future we're talking about," he said on The Diary of a CEO. His supporting observation was financial: look at the current profitability of the companies building these systems, he said, and "they're losing money hand over fist."
The frontier he is talking about is the top end of AI development — the largest models, trained on enormous amounts of data using vast quantities of specialised computer chips housed in data centres. Building and running them costs money before it earns any. Keen's argument is that this bill, rather than any treaty or moment of collective restraint, is the thing most likely to set a limit.
Two reasons the bill might not matter
The other people in the room did not accept that costs would act as a brake.
The first counterargument was military. If your side lacks "the missiles that are smarter, or the satellites, whatever, that are smarter, or the torpedoes, whatever," one speaker said, then there is a national security risk. On that logic, a government facing a rival's smarter weapons does not consult the profit and loss account; spending continues because losing is unacceptable.
The second was commercial history. How long, a speaker asked, were Facebook and Amazon unprofitable? Someone in the discussion put Amazon's unprofitable stretch at 20 years — a figure offered from memory in conversation, not checked on air. The point behind it was that investors have tolerated long periods of losses before, when the prize was a business large enough to repay them.
Keen took that as confirmation rather than refutation. There is "one survivor out of each of those companies and they're doing very well, thanks very much," a speaker observed. "Well, that's my point," Keen replied. "So one will survive."
From one survivor to a Chinese survivor
The leap came next. "With the AI who's going to survive and looking at the odds, I think it's going to be a Chinese company that survives, not the American, given the sheer cost of those data sets."
This is a forecast about consolidation, not a description of current market positions: an expectation that the money-losing phase ends with a small number of winners, and a bet on which country those winners will come from. Keen did not set out costs for particular firms on either side, and the reasoning stayed at the level of odds.
His companions had noticed a pattern. "I think we've established you're a fan of China in this," one said. "Yes, I am," Keen answered. "I'm not even making fun of you," came the reply. "I'm just saying with you all roads lead to China."
Kimi as the example he reached for
At that, Keen reached for a recent release. "I mean, looking at the Kimi system that just came out recently, the fact that it's an – the ironic thing is the communists are talking about open source and the Americans are talking about closed and secretive." He called the situation "rather ridiculous in some ways."
Keen did not name a version. The most recent Kimi release before this episode was published came on 27 July 2026, when Moonshot announced Kimi K3 by publishing the model's weights, a technical report and supporting training infrastructure together.
Weights are the learned numerical settings that make a trained model work. Publishing them means anyone with suitable hardware can download the model, run it on their own machines and adapt it, subject to the licence it is released under. That is the practical difference Keen is gesturing at: a model you can only reach through someone else's service, versus one you can take away. Moonshot gives as its reasons for the approach wider deployment, more innovation and giving users control over their own data.
The announcement describes a mixture-of-experts model with 2.8 trillion parameters, native vision and a context window of a million tokens. In that design each token activates 16 of 896 routed experts, so the headline parameter count is much larger than the computation used for any single token — a figure about the model's size and structure, not a measure of how well it performs. Downloadable weights also do not settle every question about openness: availability under a licence is not the same as a full open-source release, and where a lab is based says nothing by itself about how open its model is.
The physical side of the bill
The other constraint in Keen's argument was material rather than financial. Energy had come up earlier in the conversation as the infrastructure that lets a private sector build anything at all; where it is scarce, as one speaker put it, any new facilities have "got to be used at the expense of some other activities." The example offered in the discussion: "we're seeing Americans protesting about rising energy prices because of data centers being established and rising problems with the water supply for the same reason." Large data centres draw electricity from local grids and use water for cooling, which is where a global race turns into a neighbourhood dispute.
None of this settles who wins. It sets out the two things Keen thinks decide it — the bill and the resources behind it — against a room that thinks militaries and patient investors will pay either way.