For most of the last cycle of frontier-model competition, the working frame was that everyone was running roughly the same business: train models, host inference, capture revenue. The differences read as tactical — pricing, distribution, model quality. The shape of the underlying companies looked the same.
That frame is breaking. What's emerging is two distinct businesses that happen to share a product surface. Calling them by what they actually are: the build side and the rent side.
The build side owns its silicon. It signs decade-class compute commitments, builds custom training clusters, and runs the model lab as one half of an integrated capital structure where the other half is a multi-year capacity bet. OpenAI's Stargate posture, Anthropic's compute commitments to Amazon and Google, xAI's Tesla-flavoured supply chain — these are not procurement decisions. They are bets that the next model generation will require capacity that has to be locked in twenty-four months early or it doesn't exist.
The rent side does not own its silicon. It rents capacity from whoever has megawatts free that month. It optimises for capital efficiency over capacity certainty. Mistral's posture, most of the application-layer companies, and increasingly the second tier of frontier labs, sit here. The rent side is faster to start, cheaper to iterate, and structurally exposed to the build side's ability to set effective marginal compute prices.
The structural consequence is that the two sides are now running incompatible business models inside the same product category. The build side's economics only work if frontier capability stays on a steep enough curve that the locked-in capacity becomes load-bearing. The rent side's economics only work if frontier capability flattens enough that renting on the spot is sustainably cheaper than owning. They are betting in opposite directions on the same question.
Operators evaluating any twelve-month commitment to a model vendor right now are not, despite appearances, choosing between two model lineups. They are choosing which side of that bet they want their counterparty to be on, because the failure modes of each side look completely different. A build-side counterparty that misjudges the capability curve has stranded capex; a rent-side counterparty that misjudges it has stranded contract terms when their compute supplier reprices.
The next confirmation that this split is real, not narrative: the next major frontier-model lab to announce a multi-generation custom-silicon program. The build side will continue to commit publicly to compute through 2027. The rent side will continue to argue that capability is flattening and that renting is rational. Both will keep shipping models that look comparable on benchmarks; the divergence will only be visible in the contract structure underneath.
Before any twelve-month vendor commit this week, ask which side they sit on. The build side has capital through 2027. The rent side rides whoever has megawatts free that month.