Mira Murati's $50B Lab Just Gave Its First Model Away

Fifty billion dollars. That was the valuation before a single product shipped. That’s Thinking Machines Lab, founded by Mira Murati after she left the CTO seat at OpenAI. The lab has finally shipped something. And the delivery method is the interesting part: instead of hiding it behind an API, they released the weights. It’s called Inkling.

One caveat up front. There isn’t much community reaction to read here yet — the usual venues have been quiet, so this isn’t a temperature check on developer sentiment. It’s an argument about what the decision means structurally, given where the industry sits right now.

What Open Weights Actually Means

Quick vocabulary check, because the terms get sloppy. Open weights means the trained parameters — the weight files — are downloadable by anyone. That is not open source. Open source would mean publishing the training data and the training code too. Open weights hands you the finished brain and says nothing about how it was built.

That distinction matters less than you’d think in practice. With the weights, you run the model on your own hardware. No per-token API bill. No data leaving your network. Fine-tune it however you like. From the company’s side, this is handing over control, permanently and irrevocably. You can’t un-release a weight file.

The Idealism Case

Murati’s founding pitch was never subtle: AI shouldn’t be locked inside the black boxes of a handful of companies. If researchers can’t look inside, safety claims are unfalsifiable. If startups can’t self-host, they’re hostages to whatever pricing the big labs decide on next quarter.

By that logic, shipping Inkling as open weights is the founding thesis in executable form. Plenty of people say the words. Actually publishing the file is the message. And the symbolism lands harder given the context — OpenAI has spent years absorbing the joke about its own name, and now its former CTO walked out and went the other direction on her first move.

The Survival Case

Now the cold read. The situation facing any late-arriving frontier lab is brutal.

First, the API market is already carved up. OpenAI, Anthropic, and Google are entrenched. A new lab walking in and saying “please use our API instead” needs a reason. Switching costs are real — prompts get tuned, evals get built, contracts get signed. Nobody migrates for a marginal gain.

Second, there’s no distribution. Google has Search and Android. Microsoft has Office and a sales force that’s been in every enterprise since the Clinton administration. A new lab has none of it. Open weights is the cheapest distribution strategy in existence. Developers download it themselves, benchmark it themselves, and argue about it on Hacker News for free. That’s a marketing channel with a budget of zero.

Third, hiring. Top researchers want to publish. They want their names on things. A lab that closes everything is a hard sell against a lab that ships weights. An open-weights release doubles as a recruiting post.

Fourth, the Chinese labs already proved the playbook works. DeepSeek and the Qwen family used open weights to embed themselves in the Western developer ecosystem while everyone was still debating whether it was strategically wise. It was.

These Two Stories Don’t Actually Conflict

Here’s my take: separating the idealism from the survival math is a waste of energy. Good strategy is what happens when your principles and your incentives point the same direction.

Murati can genuinely believe openness is right. It can also be the only card her company can profitably play right now. Both can be true, and the fact that both are true is exactly what makes it durable. Companies that lose money defending an ideal don’t get to defend it for long.

The real thing to watch is the second act. Inkling was the first model and it was open. What about model number two? That’s the fork in the road. If the pattern becomes flagship-closed, one-generation-behind-open — the Meta-with-Llama, Google-with-Gemma approach — then the idealism narrative quietly deflates into standard tiered marketing. If the flagship keeps shipping with weights attached, then this is genuinely a different kind of company.

Where Does the Money Come From

Open weights means the model itself doesn’t generate revenue. So what justifies $50 billion?

A few paths exist. Sell hosting and convenience. Sell enterprise customization and support contracts. Or build a product on top and monetize that. In every version, the model is the loss leader and the money comes from somewhere adjacent. Red Hat ran this play on Linux and it worked.

The problem is that the formula hasn’t been proven in AI. Red Hat existed because packaging, distributing, and supporting Linux was genuinely hard. Model serving isn’t. Once the weights are public, AWS serves it, Together serves it, some team in a garage serves it on spot instances at cost. Where’s the moat that lets the original authors hold a margin? I don’t have a good answer, and I’m not sure they do either.

The Bet

Inkling isn’t a model release. It’s a bet: capture the ecosystem first, figure out monetization later. If it works, it rewrites the default playbook for frontier labs. If it doesn’t, it becomes a case study titled “The Ideals Were Nice.”

The checkpoint is a single question. Do they open the next one? That’s the only signal I’m watching. And it’s worth asking which answer you’d actually bet on — conviction, or the best available move from a company that doesn’t have another one.