Three bodies, one price sheet: why no single AI winner is worth betting on

Substitution is the force doing the work in the metaphor, and it carries a price tag. Zhipu's GLM 5.2 and Moonshot's Kimi K3 are useful to a procurement team not because they top a leaderboard but because they sit at or near the frontier on several benchmarks while costing a fraction of the closed models they get compared against ^[5]. Thinking Machines' Inkling and Nvidia's Nemotron 3 do a narrower job: they are described as highly capable but not quite at the frontier, which is still enough to hand a US buyer a domestic version of the same negotiating line ^[6].

The number worth sitting with is the author's own sizing of the market. AI spend is put at somewhere between 0.5 and 1 percent of all white-collar salaries in the United States ^[4]. That band is a factor of two ^[13]. Anyone building a five-year model off it is compounding an estimate that is uncertain by 100 percent at the starting line, and the piece presents the figure as the writer's estimate rather than a measured total ^[4].

The margin arithmetic is where the convergence story gets sharp. Frontier labs are described as going deeper into the product stack to widen moats and sustain high margins, while application companies go deeper into the model stack to build moats of their own, against a software norm of 70 percent-plus gross margins ^[11]. Both directions chase the same points of margin inside the same account. An application company that moves workloads onto open weights swaps a variable token bill for fixed engineering cost ^[8]. A lab that ships product features starts competing with the customers reselling its tokens ^[11].

Demand is not the disputed part. The source reports unprecedented demand and extraordinary revenue growth at the frontier labs, led by Anthropic, and says that growth is what sharpened the focus on demonstrable ROI and on cheaper alternatives ^[12].

The forecast for the second half of 2026 is really two bets stacked: that competition pushes prices down, and that the returns on AI spend begin to show ^[9], with the open versus closed distinction fading as labs support customer personalization ^[10]. The first shows up on a published price sheet. The second shows up in a customer's own accounts, if anywhere. One is auditable and one is inferred, which is the sharper reason to size exposure by where value lands rather than by which body currently looks largest.