Frontier vs Open-Source: When AI Revenue and Token Volume Decouple

90% of AI revenue accrues to the frontier. ~80% of tokens consumed are open source. Both are true at once — and most people only hold one of them at a time.

On the latest BG2 Pod, Brad Gerstner sat down with Gavin Baker and Andrew Fox of Atreides Management and Altimeter's Clark Tang for the clearest frame on AI economics we've heard this year: value and volume have decoupled.

Our mid-year research adds one correction that sharpens their conclusion rather than breaking it. The panel placed open-weight models roughly six months behind the closed frontier. Our work, The Frontier Compresses, puts that gap at ~3 months and shrinking — down from ~12 months a year ago. On coding it's effectively gone: open weights score 80.2% on SWE-bench Verified versus 80.8% for a frontier model.

That single change matters. A 90% revenue lead resting on a three-month, fast-closing capability gap is not a lead built on capability — there isn't a chasm there. It's built on reliability, distribution, and enterprise trust. A more contestable asset than raw intelligence.

For capital allocation, that points one direction: own the stack and the compute, and avoid the squeezed middle — the sub-scale closed model with neither a reliability edge nor a cost advantage over good open weights.

The single number to track is whether the frontier holds ~90% of revenue as its capability lead keeps shrinking. When those two lines finally diverge, the moat was never capability — and the repricing follows.

Read “The Frontier Compresses” · BG2 — “The SpaceX IPO, Fable 5, AI Capex Update & Market Check”