The gap is closing, and closing fast
Three to five months now separate the best model you can rent from the best you can run yourself. Two years ago it was six to nine.
The classic argument against self-hosting was quality: open-weight models trailed badly, and the gap looked structural. That argument eroded faster than almost anyone expected.
Kimi K3, released open-weight on July 27, 2026, ranks third on Artificial Analysis's intelligence index — behind only Claude Fable 5 and GPT-5.6 Sol Max — and first on the Frontend Code Arena. A single generation closed thirteen index points in three months.
What matters for an infrastructure decision isn't the instantaneous ranking but the speed of convergence. A three-to-five-month gap means a self-hosted deployment catches up to the frontier on every release cycle, with no renegotiation and no provider migration.
The side effect is economic. When several open models of comparable quality are available, the cost of switching models collapses, and pricing power shifts toward whoever owns — or reserves — the compute.
That's the underlying bet behind dedicated infrastructure: you aren't betting on any particular model, you're betting that there will always be an excellent one to run on the machine.