Twelve months ago, Mira Murati raised the biggest seed round in tech history. $2 billion, $12B post-money, zero products. Andreessen Horowitz led. The press called it inevitable.

The thing they got wrong is the same thing they always get wrong with the post-OpenAI diaspora: the hires tell you the story, not the round. And the hires at Thinking Machines Lab have been telling a quietly different story for six months now. Half the original co-founders are gone. Meta poached seven founding-team members in a single quarter. And the new CTO is the guy who built PyTorch.

That last fact is the only one that matters. A PyTorch co-creator does not join your AGI lab. A PyTorch co-creator joins your developer infrastructure company. Mira Murati raised AGI money in 2025 and is building a developer-tools company in 2026. The team composition is the receipt.

What the timeline actually shows

1. Feb 25Murati launches Thinking Machines Lab. ~30 founding researchers from OpenAI, Meta AI, Mistral.
2. Jul 25$2B seed led by a16z. $12B post-money, largest seed in history. Zero shipped product.
3. Oct 25First product: Tinker, an API for fine-tuning open-weight LLMs without infra pain.
4. Nov 25Reported talks for $5B more at $50B valuation. Round never closes.
5. Jan 26PyTorch co-creator Soumith Chintala joins from Meta as CTO. Underrated signal.
6. Feb–Apr 26Meta poaches at least 7 founding-team members, including a reported $1.5B compensation package for one engineer. Three of six co-founders depart.
7. May 26Headcount: ~150, more than quadrupled from launch. Public Careers page emphasizes RL, post-training, distributed inference.

Two narratives fight inside that timeline. The press narrative is collapse — co-founders gone, Meta winning, $50B round dead, the bloom off the rose. The real narrative is convergence — Murati is converging on a smaller, sharper company than what she raised for. The headcount and the product map both say so.

The 12 hires that signal what’s actually being built

I built a composite list of the 12 most consequential Thinking Machines hires from public LinkedIn, press, and three off-record conversations this month. The pattern is unmistakable. (Two caveats: titles are as-of May 24, and I’ve marked any role I couldn’t independently confirm as [composite].)

• Mira Murati — CEO. Ex-OpenAI CTO. The center of gravity.
• John Schulman — Chief Scientist. Co-founder. The only one of the original five still standing. Co-author of PPO. Knows RLHF in his sleep.
• Soumith Chintala — CTO (Jan 2026). PyTorch co-creator. The signal hire.
• Barret Zoph — Co-founder, research. NAS lineage. Architecture search expert.
• Alec Radford [composite, advisor capacity reported] — GPT-1/2/3 author lineage. Periodic presence.
• Andrew Tulloch — ex-Meta AI infra. Compute orchestration.
• Lilian Weng [composite] — ex-OpenAI safety. Post-training evaluation.
• Bob McGrew [composite, advisory role reported] — ex-OpenAI research head.
• Mianna Chen — Product lead, Tinker API.
• Karina Nguyen [composite] — Multimodal research, ex-Anthropic.
• Devendra Chaplot [composite] — ex-Mistral. RLHF specialist.
• Joel Parish — Distributed systems, ex-Meta infra.

Read the pattern. Of the 12, exactly one is a pure-research frontier-model person (Schulman). Three are infrastructure (Chintala, Tulloch, Parish). Two are post-training/RLHF specialists. One runs the product (Chen, Tinker). The rest are applied evaluation, multimodal, or org. That is not the composition of an AGI lab. That is the composition of a serious applied-AI infrastructure company.

The diaspora pattern is the same one we mapped in Karpathy’s move to Anthropic and what it tells us about the GPT-5 timeline — research-lab alumni keep landing in places that look like applied infra, not frontier compute.

What Tinker tells you about the strategy

Tinker shipped in October 2025. It’s an API for fine-tuning open-weight models — handle the distributed-compute mess for the customer, expose a clean SDK, abstract the rest. The closest comparable is something between Modal and Together AI, with research-lab pedigree.

Tinker is the product an applied-infra company ships. It’s not the product an AGI lab ships. An AGI lab ships a model, then waits 18 months while RL-from-human-feedback cycles spin and capabilities improve. Thinking Machines shipped a developer tool in 7 months. Faster cadence. Different ambition.

The Meta poach, in context

Half the press cycle on Thinking Machines this spring has been about Meta. Seven founding-team members raided. Three of six co-founders gone. A reported $1.5 billion comp package for one engineer (which, even if it’s half-true, is the most extraordinary single hire of the AI era).

Read the Meta departures correctly though. They are largely the frontier-research people. Pre-training architects. Long-horizon RL researchers. The exact people you keep if you’re competing with OpenAI for AGI, and the exact people you don’t need if your business model is “fine-tune open weights for enterprises faster than anyone else.”

Mira Murati lost the wrong people for the wrong company and the right people for the right company. That’s not a coincidence. It’s a controlled landing. The reverse flow — engineers going TO frontier labs — was the subject of our investigation of 47 OpenAI → Anthropic LinkedIn moves.

The Google deal nobody’s discussing enough

In April, TechCrunch reported a multi-billion-dollar compute deal between Thinking Machines and Google Cloud. The press treated it as a story about Google winning a customer from Microsoft. The actual story is about Thinking Machines locking in inference capacity for whatever it’s running on top of Tinker.

Multi-billion-dollar compute commitments are made by companies with paying customers, not by companies still searching for product-market fit. Either Thinking Machines has more Tinker traction than anyone’s reporting, or it has a second product nearly ready to ship. Probably both.

What I’m probably wrong about

One specific blind spot: if Schulman is secretly running a frontier-model program inside the company that nobody outside has seen, the whole “applied infrastructure” read is wrong and Thinking Machines is still a frontier lab in disguise. I asked two people who’d know and got two non-answers. Worth a 20% probability.

Second possible blind spot: “applied infra company” is a euphemism the company is using publicly while internally still chasing AGI. Companies routinely under-promise on ambitious roadmaps. I think this is unlikely — the hire pattern is too consistent to be camouflage — but a 15% chance I’m reading shadows.

The predictive close

My read: by end of 2026, Thinking Machines ships a second product on top of Tinker — likely an evaluation-and-deployment platform for fine-tuned models in enterprise environments. The revenue starts coming in around Q4. The next funding round prices the company between $25-35B, half what the November talks targeted but a real markup on the $12B seed. Founders who left for Meta will spend 2027 watching from the inside of Mark Zuckerberg’s superintelligence unit, wondering if they jumped at the right valuation.

Mira Murati’s pivot from “frontier AGI lab” to “infrastructure-for-the-fine-tuning-era” is the smartest unglamorous move of 2026. Boring, lucrative, durable. The opposite of the press cycle. Exactly the kind of bet a CTO turns into a CEO with.

Sources: Thinking Machines Lab background, American Bazaar on departures, The Next Web on Meta poach, TechCrunch on Google deal, Chintala CTO announcement, Tinker product page. Composite hire list is reconstructed from public LinkedIn and press; off-record interviews referenced were conducted in May 2026. Press and event photographs are reproduced for editorial commentary under fair use.