Nathan Lambert, writing in Interconnects, argues that the current set of organisations releasing near-frontier open models is contracting and that a consortium structure is the only long-term mechanism capable of sustaining well-funded open development.
TL;DR: Lambert contends that frontier training costs, departures at open model labs, and competitive pressure to monetise models rather than release them will force a shift toward a funded multi-company consortium as the primary vehicle for open, near-frontier AI.
What it says
Lambert traces the argument from a conversation with Percy Liang, Stanford professor and lead of the Marin project, which Lambert says crystallised for him the structural case for a consortium. He names Nvidia’s Nemotron coalition as an existing single-company attempt to bootstrap this model, but argues it lacks the stability of a multi-funder structure.
He points to high-profile staff departures at Qwen and Ai2 as early indicators of the sustainability problem, and compares these to earlier shifts when Meta moved focus away from Llama. The Chinese mid-tier labs — Moonshot AI, MiniMax, and Z.ai — he describes as financially precarious if they maintain current release strategies, given the cost of frontier training and the revenue opportunity in keeping models proprietary.
Lambert divides future release behaviour into two groups. The first includes companies such as Arcee AI, Thinking Machines, OpenAI, and Google, which he expects to release many smaller, fine-tuneable models targeting niche downstream use cases. The second group — those willing to release fully open, near-frontier models — he expects to shrink. He writes: “there will be an ever increasing number of companies releasing models that are good for creating a lively niche of smaller, custom models, but an ever decreasing number of companies willing to release fully open, near-frontier models.”
On Nvidia’s position, Lambert lists several scenarios that could cause the company to pull back from open model investment: its open efforts becoming too competitive with its largest GPU customers; loss of cash flow to competition on its core chip business; or sufficient success on frontier closed models to justify retaining weights.
Lambert acknowledges that a consortium is inherently difficult to execute, writing that training models “is inherently a complex and high-focus endeavor” that resists governance by committee. He nonetheless concludes the structure is inevitable given the combination of rising costs, commercial pressure, and the growing number of companies dependent on open-weight access. He closes by noting that the window for true open pressure to materialise may be delayed until capital markets begin penalising less efficient spending strategies — a pressure he expects to hit Chinese startups first.