In July, Dean Ball, who runs strategic foresight at OpenAI, warned that China's promotion of open AI models could lead to full AI communism. He meant it as a description of a dystopia. Paul Mason, the broadcaster and author of Reds: A Global History of Communism, has taken it as an opening. In a new essay he says Ball's anxiety amuses him, that Kimi, Qwen and DeepSeek are not the end of capitalism but one bid for global dominance answering another, and that the left should stop arguing about sovereign national models and start building a model nobody can own.
Mason's framing of the two strategies is the clearest part of the argument. The American approach accumulates compute and lets users reach it only over the internet, which forces the rest of the world onto American software, American cloud services and American regulatory rules. China answers with open models that users can run on their own servers, keeping their data away from the Chinese supplier. Ball's own account, as Mason relates it, is that Beijing treats AI as a public good to be delivered eventually by the state as digital public infrastructure.
The bet underneath the Chinese position is that AI turns out to be a general-purpose technology, in which case the competition that matters is not over models but over what gets built on top of them: brain-computer interfaces, bioengineering, and the low-altitude economy of drone delivery and related services. Whoever wins that determines how information technology develops for the rest of the century. Mason's move is to reject the premise that those are the only two entries.
His case that AI is different from previous technological revolutions rests on who it displaces. Large language models already have some of the properties Marx described in 1858 as the general intellect — built to hold and work with essentially all the knowledge humanity has produced — and frontier models already outperform some trained, experienced knowledge workers. Everyone can see tasks and whole functions being taken over. What Mason argues has not been thought through is that the exposed population includes capitalists. If a machine can replace a junior lawyer or a web designer, it can replace the entrepreneur, the innovator and the currency trader.
The economist Philippe Aghion is the load-bearing citation here: AI can automate entrepreneurial skill and, more damagingly, sever the link between innovation and profit. If the blueprint for an invention can be immediately reproduced and improved by a machine, who bothers to patent a discovery? That is the sharpest idea in the essay and it does not require agreeing with anything else in it. Patent economics assume a defensible interval between invention and imitation. Remove the interval and the reward structure that funds private research goes with it — a problem for the venture model well before it becomes a problem for capitalism as such.
From there Mason sets out what to do: break the link between work and wages through universal basic services and income, grow cooperative and non-profit business models, and use AI to do what Soviet planning could not — replace the market with a rational mechanism for allocating resources. He concedes the underlying technology would have to be different, because American AI looks aggressive, extractive, closed and mercantilist for the reason that it is designed to reproduce a particular social reality, and Chinese models, open or not, are built to serve a project of global dominance. A green, social-democratic model would economise on energy and compute, prefer altruism to competition and individualism, and start from the interests of a community rather than an individual. If you accept running six months behind Claude or Kimi, he argues, you can use them to build a model that cannot be owned.
The six-month figure is where the argument stops being serious, and it is worth saying plainly. You do not choose how far behind the frontier you run. That distance is set by capital, chips and data, and a cooperative sector has none of the three at the scale required. Mason's own analogy gives the game away: he asks why alternative systems could not take share from Silicon Valley the way open Chinese models have. The answer is in his own reporting. The Chinese models are open because a state with a global strategy and firms with enormous balance sheets decided openness was the better weapon. Openness there is a funded strategic choice, not a low-cost entry route.
The pricing argument has the same shape. Mason notes the alternative would be free while today's services depend on $20-a-month subscriptions paid by millions of users. But that subscription is the revenue line, not the cost line. The reason nobody can undercut it by charging nothing is that the expense sits in training and serving, and it does not fall because the owner is a cooperative. Being free is the easy part of the plan; staying switched on is the hard one.
The tension Mason does not resolve is between his diagnosis and his prescription. He expects AI to trigger a large-scale crisis in the existing model — capitalism without wage workers, innovators or entrepreneurs, he writes, stops operating by the old rules. Universal basic services and income are paid for out of a surplus that the same crisis would be destroying. Marx described nineteenth-century industry as pushing people to the side of production, turning machine operators into machine watchers, and Mason argues AI is doing the same thing to intellectual labour at a very high level less than five years after large language models went mainstream. If that is right, the window for building the alternative closes at roughly the speed of the crisis that makes it necessary.