The Behemoth invoice: What Meta's $1.5B hire from Thinking Machines Lab actually paid for

The headline — "Meta raids Mira Murati's Thinking Machines Lab, hires five founders" — is wrong in its most important detail. Meta hired one founder: Andrew Tulloch. Three co-founders (Barret Zoph, Luke Metz, Sam Schoenholz) returned to OpenAI. One (Devendra Chaplot) joined xAI. This is not a Meta raid. It is a dispersal — five researchers going to three labs whose needs align with their specific expertise, in two waves. Tulloch (and founding engineer Joshua Gross) went to Meta in October 2025. The three OpenAI returns and Chaplot's move to xAI followed in the months after TML's $50 billion valuation round failed in November 2025.

That failure is the actual story, and it has been missed by nearly every outlet covering the reported package.

The Behemoth connection no coverage made

Tulloch's career is large-scale distributed training. His most cited published work: "ImageNet in 1 Hour" (demonstrating large-minibatch SGD at scale) and foundational inference optimization for data center deployment. These are not general AI credentials — they are specific capabilities in the training infrastructure domain.

Llama 4 Behemoth, Meta's flagship open model, was shelved in early 2026 due to MoE routing instability and chunked-attention reasoning blind spots at 2 trillion parameters. SemiAnalysis named Meta's "two core shortcomings" as talent and compute. The specific failure modes — MoE routing at 2T scale, distributed training stability — map precisely to the problems Tulloch's published research addresses.

Meta reportedly paid up to $1.5 billion over six years — not because the talent war drove prices to absurdity, but because Behemoth failed, the failure was technically traceable to infrastructure gaps, and Tulloch is the researcher whose published work directly addresses those gaps. The reported package is a repair bill with a name on it. Meta disputed the exact figure as "inaccurate and ridiculous" but did not deny substantial compensation; sources close to negotiations told reporters the numbers "weren't far off."

Notably, Tulloch's departure in October 2025 predated the November 2025 failed raise by one month. His recruitment was driven by Meta's specific technical need, not by TML's subsequent fundraising difficulties.

The $50B failure is why Thinking Machines Lab lost five co-founders

Zoph, Metz, and Schoenholz — the three co-founders who returned to OpenAI — departed in January 2026. Fortune reported their departures had "more to do with money than otherwise" and that some were being offered "insane packages to return to OpenAI."

The timing is not coincidental. TML's $50 billion valuation round collapsed in November 2025 — the same period during which these three founders would have been negotiating their departure terms. When you raise $2 billion on a founding team and that team's next raise fails at $50B, the co-founders who hold the most leverage begin receiving competing offers.

I think the "Meta aggression" framing has dominated because it's a better story than "startup valuation failed and founders left." The aggression framing lets Meta be the villain and Murati be the survivor. The actual story is more ordinary: a startup that could not maintain compensation parity with public companies after a failed fundraise lost one-third of its founding team — approximately 13 of ~40 founding members — in a six-month window.

Zoph's role is the unreported detail

Barret Zoph co-pioneered Neural Architecture Search at Google Brain. At OpenAI from 2022 to 2024, he led post-training inference — one of the most technically demanding research roles at the frontier lab. He co-founded TML as its CTO.

He returned to OpenAI to lead enterprise go-to-market, reporting to Fidji Simo (CEO of Applications), not to research leadership.

A researcher who built some of the foundational scaling techniques in modern AI is now in a sales leadership role. Most coverage missed this entirely — leading with his Google Brain and OpenAI technical credentials, implying his return is a research win for OpenAI. It is a commercial-division hire.

Two readings: (a) Zoph concluded TML's technical frontier work was not distinctive enough to justify the compensation gap, and took money and distribution over research independence; (b) Zoph believes enterprise distribution is now the actual leverage point at OpenAI's scale — that who wins deployment wins AI, and the most important role right now is getting OpenAI into enterprises, not building new models. I don't have enough information to choose between them, but either reading is more interesting than "TML CTO returns to former employer."

Each researcher went exactly where their expertise was most needed

The dispersal is strategically coherent. Metz (training infrastructure, contributions to GPT-4 training) went to OpenAI's training infrastructure team. Schoenholz (μTransfer, hyperparameter transfer for large models) reports to Zoph at OpenAI — his scaling methodology has direct application to OpenAI's production training pipeline. Chaplot (embodied AI, navigation research from CMU and Meta AI) joined xAI, whose physical AI roadmap through SpaceX makes him an obvious fit. Tulloch went to Meta to fix Behemoth.

The talent market is sorting by technical specificity, not by prestige or compensation ceiling alone. Sam Altman confirmed Meta offered "$100M signing bonuses and even more" to OpenAI researchers — but the Tulloch package is a different category. Once a reported $1.5B figure enters the public record (even if disputed), every elite researcher now knows the ceiling. The Tulloch case does not just reflect the talent war; it will accelerate it.

Can Thinking Machines Lab survive?

TML is weakened but not terminal. It still has $2 billion in seed capital from July 2025, a Google Cloud multibillion-dollar deal signed in April 2026, an NVIDIA strategic investment with a 1 gigawatt Vera Rubin compute agreement, and approximately 140 to 183 employees. Murati has been building a counter-raid: Soumith Chintala (PyTorch co-creator, 11 years at Meta) replaced Zoph as CTO; Piotr Dollár (Segment Anything co-author) was recruited from Meta FAIR. At the Bloomberg Tech Conference on June 9, Murati previewed "interaction models" processing continuous audio, text, and video streams at 200 millisecond intervals — significant if real, but a preview without a model release timeline.

I think TML is in the position of having premium infrastructure and financial backing but an uncertain technical thesis now that the founders who held the scaling methodology (Tulloch, Metz, Schoenholz) have left. Chintala and Dollár are serious researchers replacing departed training and scaling expertise — but they are replacements for what was lost, not additions above it.

TML still has enough runway to find out whether Murati's technical vision is differentiated enough to justify frontier lab status. The founders who left did not kill it. But the failed $50B raise signals what investors thought the founding team was worth — and one-third of that team is now at three of TML's direct competitors.