OpenAI and Anthropic are feeling the pressure of a rapidly evolving AI market where Chinese open-weight models are closing the performance gap while application firms hunt for cheaper, more controllable solutions. The situation resembles a three-body problem in physics: each player can alter the others' trajectory, but none can dictate the final outcome.
The three-body AI economy
In this analogy, the three bodies are the closed-source frontier labs led by OpenAI and Anthropic, the open-weight models emerging mainly from China, and the application companies that layer services on top of both. Each is powerful enough to reshape the others, yet the system remains unstable.
Rising pressure on frontier labs
Recent months have amplified the instability. Frontier labs have seen unprecedented demand and revenue growth, expanding AI's reach across the economy. At the same time, the focus on measurable return on investment has sharpened. Spending on AI in the United States now accounts for roughly half to one percent of all white-collar salaries, a scale that invites scrutiny.
"Something has gone completely wrong" with how the labs sell their product, and enterprises are "tokenmaxxing", spending furiously on tokens with no matching gain in productivity,
said Alex Karp, chief executive of Palantir, in a July interview with CNBC. Competition at the frontier has also intensified, with Meta (Muse Spark 1.1) and xAI (Grok 4.5) fielding increasingly capable models alongside the established players.
Chinese open-weight models gain ground
Open-weight models from China are now approaching frontier performance. Zhipu's GLM 5.2 and Moonshot's Kimi K3 perform at or near the top of several key benchmarks, yet they are priced at a fraction of comparable closed models. In the United States, domestic open-weight offerings such as Thinking Machines' Inkling and Nvidia's Nemotron 3 provide credible alternatives, though they have not yet matched the very latest frontier scores.
Implications for the second half of 2026
Three broad trends are likely to shape the remainder of the year. First, the discomfort with frontier pricing should ease as competition pushes prices down and the productivity gains from AI become clearer. Historically, mass adoption of technologies such as automobiles and mobile phones followed price declines; AI adoption has outpaced utility, but the payoff is expected to arrive.
Second, a multi-model landscape will continue to develop, driven by competition and genuine differentiation in each model's strengths. Third, U.S. open-weight models are poised to become genuine alternatives to Chinese offerings, gaining adoption thanks to clearer business models that allow customers to make longer-term bets.
In parallel, frontier labs are likely to deepen their involvement in the product stack to protect margins, while application firms will embed more of the model stack to build their own moats. Software companies typically enjoy gross margins above 70 percent, and customers tend to stay when they perceive value.
The three-body system remains unsettled, but the current noise, debates over open versus closed models, concerns about Chinese dominance, and worries about return on investment, may prove temporary. The real question is not whether AI will deliver value, but which group, the frontier labs, the open-weight challengers, or the application companies, will capture the most of that value.