For years, the assumption was that AI model training would get so expensive that only a handful of labs could keep up. Everyone else would fold or get bought.

That has not happened, according to AI researcher Nathan Lambert, who writes Interconnects AI, a newsletter on AI models and research. He argues:

“We’re at a place where more companies are training strong models… and an increasing number of organizations are releasing these models openly.”

What Lambert is arguing

Lambert’s post rounds up recent open-weight model releases, models whose underlying files anyone can download and run, rather than only reach through a paid API, from labs in the US and China. His point is not just that these releases keep happening. It is that the companies making them are finding real revenue in it.

His clearest example: Thinking Machines, the AI lab that launched in February 2025, now runs what Lambert describes as “an open model finetuning service [that] is making hundreds of millions in revenue per year.” Finetuning means taking a general open model and further training it on a specific task or dataset, a service labs can charge for even when the base model itself is free to download. Lambert says Thinking Machines is now ahead of NVIDIA’s Nemotron and Arcee’s Trilogy as the strongest US-based open-weight effort.

He also points to continued releases from Chinese labs. DeepSeek shipped an updated V4 Flash model. Moonshot AI’s Kimi K3 was, in Lambert’s words, “the biggest open model release in some time.” Tencent switched its Hy3 model line to the Apache 2.0 license, a standard permissive open-source license with few restrictions on reuse, replacing more restrictive custom terms.

Lambert’s explanation for why consolidation has not arrived: demand for AI usage keeps climbing as models get more efficient and cheaper to run, so building and hosting a strong model is a path to revenue on its own, not just a competitive move that only makes sense with a bigger, richer parent company behind it.

Who Lambert is

Lambert spent years as a senior research scientist and post-training lead at the Allen Institute for AI (Ai2), where he helped build Ai2’s own open models. He left Ai2 in mid-2026 and now writes Interconnects AI independently. He is also a paid advisor to Mercor and to Arcee AI, one of the companies his post names approvingly as an early leader in US open-weight models.

What he gets right, and where the take is incomplete

Lambert is right that consolidation predictions have been wrong so far, and his examples back that up: DeepSeek, Moonshot, Tencent, and Thinking Machines are all still shipping competitive open releases, not folding into larger rivals.

The Thinking Machines revenue figure is worth reading carefully. Thinking Machines is a private company, and “hundreds of millions in revenue per year” is not a number the company has published or that BuilderWithin could independently confirm. Treat it as Lambert’s own estimate, not an audited figure.

His advisory relationship with Arcee AI is also worth knowing. It does not mean his broader argument is wrong, but readers should know he has a paid connection to one of the companies he cites as a strength in the open model market.

Why it’s notable

This adds a different angle to a story BuilderWithin has covered before, which focused on whether open-weight models are risky or whether US restrictions on frontier models are the bigger problem. Lambert’s post is about supply, not safety: the number of labs shipping strong open models is not shrinking toward a handful of players. It is getting more crowded, which is a separate question from whether any of them should be restricted.

What it means for builders

More viable open-weight options is good news if you build products on top of AI models rather than train your own. It means:

  • More competition on price for similar capability, instead of being stuck with one or two vendors.
  • Less risk of a single provider’s price hike or outage taking down your product, if your app is designed to swap models.
  • More finetuning options if your use case needs a model trained on your own data, not just a general-purpose one.

The practical step is not to chase every new open release. It is to keep your app’s model choice swappable, so a future price change or a stronger, cheaper option doesn’t require rebuilding your product to take advantage of it.

End of article