AI's Three-Body Problem: no single force can dictate the outcome

The AI market is starting to look less like a clean race and more like a messy orbit. Closed-source frontier labs such as OpenAI and Anthropic still matter most, but open-weight models, especially from China, and the application companies built on top of them are now tugging just as hard on the system. A small shift in one place is being felt everywhere else.

The pressure is coming from the top of the stack. Anthropic has seen unusual demand and big revenue growth, while the broader frontier group has become more competitive, with Meta, xAI, Anthropic, OpenAI and Google all pushing out stronger models. That has helped spread AI tools through more of the economy. It has also sharpened the complaint that the bill is rising faster than the payoff. One estimate in the piece puts AI spending at somewhere between 0.5 and 1 percent of all white-collar salaries in the United States.

That is why the ROI argument has gotten louder. In July, Palantir’s Alex Karp told CNBC that something had gone wrong with how the labs sell their product, arguing that enterprises are spending heavily on tokens without seeing matching productivity gains. The term he used was “tokenmaxxing.” Whatever the label, the basic point is simple: companies are asking whether the current price of frontier AI still makes sense.

At the same time, the open-weight side is becoming harder to ignore. Zhipu’s GLM 5.2 and Moonshot’s Kimi K3 are said to be at or near the frontier on several important benchmarks, and they do it at a fraction of the price of comparable closed models. US open-weight efforts are building around that pressure too, with Thinking Machines’ Inkling and Nvidia’s Nemotron 3 offering a domestic alternative, if not yet a full match for the frontier. Application companies are responding by building more on open-weight models to cut costs and gain control.

The shape of the rest of 2026, then, is likely to be less about one winner than about movement in all three directions. Frontier pricing should ease as competition brings prices down and value starts to show up. Multi-model setups should keep spreading. And the line between open and closed may blur if the frontier labs themselves start supporting more personalization for specific customer needs. The real question is not whether AI works. It is who gets paid when it does.