The AI economy has a stability problem, and it is not the one most analysts are watching.
Saurabh Gupta, Managing Partner and Co-Founder of DST Global, laid out the dynamic in a Fortune commentary published August 22: three distinct forces — closed-source frontier labs led by OpenAI and Anthropic, open-weight models mostly out of China, and the application companies built on top of both — are each powerful enough to bend the others' trajectory, yet none can dictate where the system finally settles. He calls it AI's Three-Body Problem.
The frontier labs are winning revenue and losing the narrative
Anthropic and its peers have posted what Gupta describes as 'unprecedented demand' and 'extraordinary revenue growth.' But scale has a cost: it has sharpened enterprise focus on demonstrable ROI and the search for cheaper alternatives. Gupta estimates that spending on AI now runs 'somewhere between 0.5 and 1 percent of all white-collar salaries in the United States.'
At that level, scrutiny is inevitable. In July, Palantir's Karp told CNBC that 'something has gone completely wrong' with how the labs sell their product, arguing that enterprises are 'tokenmaxxing' — spending furiously on tokens with no matching gain in productivity. Competition at the frontier has also intensified, with Meta fielding Muse Spark 1.1 and xAI releasing Grok 4.5 alongside Anthropic, OpenAI and Google.
Open-weight models close the gap
The second body in the system is moving fast. Zhipu's GLM 5.2 and Moonshot's Kimi K3 now 'perform at or near the frontier on several important benchmarks,' according to Gupta, and are priced at a fraction of comparable closed models. On the domestic side, Thinking Machines' Inkling and Nvidia's Nemotron 3 are adding credibility to U.S. open-weight alternatives, even if they have not yet reached the frontier.
Application companies are responding rationally: many have ramped up efforts to build on open-weight models, targeting lower costs and greater control.
Three trajectories for the second half of 2026
Gupta offers three broad forecasts. First, discomfort with frontier pricing will ease as competition pushes prices down and returns on AI spend begin to materialize — what he calls a timing mismatch between adoption and utility. Second, the shift toward a multi-model world will continue. Third, U.S. open-weight models will become genuine alternatives to Chinese releases and carry a clearer business model, making longer-term customer bets easier.
He also anticipates convergence: frontier labs going deeper into the product stack to protect margins, and application companies going deeper into the model stack to build their own moats. Software companies, he notes, 'typically enjoy 70%+ gross margins while customers feel they get their money's worth from the product.'
The market question that actually matters
The debate over open versus closed models, the China panic, and the hand-wringing over returns all look temporary from a capital-allocation standpoint. The genuinely interesting question — the one that will determine where enterprise value accumulates — is not whether AI pays off, but who captures that value: the frontier labs, the open models competing on price, or the application layer that owns the customer relationship.
For investors and enterprises alike, that is the only scorecard worth keeping. Free enterprise rewards whoever solves the ROI equation first, and right now, the race is wide open.
