SIGNAL
The AI lab shakeout that Nathan Lambert long predicted would arrive by 2026 or 2027 is not happening. More organizations are training and releasing strong models this year than ever, and the value path behind the proliferation is hiding in plain sight: the token machine.
A token machine runs a model primarily as a revenue-generating inference product, where the commercial service is the moat, not the weights. Open weights become a customer-acquisition channel for a paid service, and the service keeps the lights on while the lab ships the next release. Thinking Machines' first release, Inkling, a 975-billion-parameter multimodal mixture-of-experts model that handles text, images, and audio as inputs, anchors the pattern. Tinker, the company's paid fine-tuning service, is reportedly generating hundreds of millions of dollars per year (USD), with the model itself positioned as the base customers pay to adapt.
Most readers will hear 'open model' and picture a giveaway. The pattern is the opposite. Poolside's Laguna S 2.1, Moonshot's Kimi K3, and Tencent's Hy3 sit in the same release window with the same logic: monetize the service, ship the weights.
The repeatable mechanism: when a paid API is the moat, the weights can travel without subsidizing a competitor, so for now the token machine suggests a path to value that does not require consolidation — but that path is still being tested. Revenue-share licenses may not stick, NVIDIA dependence remains, and a real shakeout could still arrive on a longer timeline, but the token machine is durable enough to keep the field open.
Reported by Sky for Type0, from Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier. Read the original: interconnects.ai