Thinking Machines' bet: real-time, customizable AI — $2B seed, 1GW of Nvidia, then Inkling

Mira Murati's lab bet the frontier's next step is interaction and customization, not just scale — a record $2B seed and 1GW of Nvidia compute behind it.

Thinking Machines Lab

它在赌什么That frontier AI's next leap is real-time interaction and customization, not just scale — and that a record seed round plus 1GW of Nvidia compute could buy it.在扩

做的是什么生意

Thinking Machines Lab is an AI research lab, founded by former OpenAI CTO Mira Murati, building multimodal frontier models that work collaboratively in real time and can be fine-tuned by users on its Tinker platform.

启动资金：$2B seed led by a16z in July 2025 at a $12B post-money valuation — the largest seed round on record — with Nvidia, AMD, Accel, ServiceNow, Cisco and Jane Street participating

起因

Mira Murati left OpenAI in late 2024 after six years, including as CTO of ChatGPT, and founded Thinking Machines Lab in February 2025 with OpenAI co-founder John Schulman as chief scientist and Barret Zoph as CTO. The 29-person founding team was drawn mostly from OpenAI, Character AI and Google DeepMind. Its launch post said AI systems remain hard for people to customize, so the lab would build multimodal models that 'work with people collaboratively.'

新闻TechCrunch: launch with Schulman and Zoph

经过

In July 2025 the company closed a $2B seed led by a16z at a $12B post-money valuation — the largest seed round in history — with Nvidia, AMD, Accel, ServiceNow, Cisco and Jane Street participating, despite no revenue or products yet. It shipped its first product, the Tinker API, in October 2025. In March 2026 it signed a multi-year partnership with Nvidia to deploy at least one gigawatt of Vera Rubin systems from 2027, plus a strategic Nvidia investment; in May 2026 it previewed its first 'interaction model,' TML-Interaction-Small, which streams audio, video and text in 200-millisecond chunks instead of turn-taking. Along the way it lost four co-founders: one to Meta, three back to OpenAI.

新闻Reuters: $2B seed led by a16z 新闻TechCrunch: 1GW Vera Rubin deal, co-founder exits 报道RITS: first interaction models, May 2026

结果

In July 2026 it released Inkling, its first open-weights model — a 975B-parameter multimodal mixture-of-experts model under an Apache-2.0 licence — explicitly positioned not as the strongest model but as a base for customer fine-tuning on Tinker. The bet is still being tested: $2B+ raised, a >$12B valuation, a 1GW Nvidia supply deal, and first products shipped in 2025–2026.

报道Simon Willison: open-weights Inkling

背景

Thinking Machines Lab was founded in February 2025 by Mira Murati, who had left OpenAI in late 2024 after six years including as CTO of ChatGPT, DALL-E and Codex. OpenAI co-founder John Schulman joined as chief scientist and Barret Zoph, who had led post-training at OpenAI, as CTO; the 29-person founding team was mostly ex-OpenAI. The launch thesis: AI systems have advanced fast but remain hard for people to understand and customize, so the lab would build multimodal models that 'work with people collaboratively.'

The market bought the people before the product: in July 2025 the company closed a $2B seed round led by Andreessen Horowitz at a $12B post-money valuation — the largest seed round in history — with Nvidia, AMD, Accel, ServiceNow, Cisco and Jane Street participating, even though it had no revenue or products. Its first product, the Tinker API for fine-tuning models, shipped in October 2025.

In March 2026 the lab signed a multi-year partnership with Nvidia to deploy at least one gigawatt of Vera Rubin systems starting in 2027, alongside a strategic Nvidia investment. In May 2026 it previewed its first 'interaction model,' TML-Interaction-Small, a 276B-parameter mixture-of-experts model that streams audio, video and text in 200-millisecond chunks — replacing the turn-taking pause of voice assistants — though only to research partners. The same period exposed the talent risk: four co-founders left, one to Meta and three back to OpenAI.

In July 2026 Thinking Machines released Inkling, its first open-weights model: a 975B-parameter multimodal mixture-of-experts model under Apache-2.0, trained on 45 trillion tokens, explicitly framed as a strong base for customer fine-tuning on Tinker rather than a frontier model. As of September 2026 the company is private, valued above $12B, and has converted a record seed into compute commitments and two product lines — with the interaction-model vision still in preview.

这件事要成立，得有什么

Talent was the wedge: two-thirds of the founding team came from OpenAI, and investors bet on people before product — the $2B seed closed months before anything shipped.
The thesis was differentiation: instead of another chatbot, build models that are real-time, understandable and customizable — a direction the big labs weren't optimizing for.
Compute became the moat strategy: locking up 1GW of Vera Rubin plus a strategic Nvidia investment turned a seed-stage lab into a capital-infrastructure play.
The risk showed in the exits: four co-founders left within roughly 18 months, a reminder that a talent-concentrated lab's biggest asset is also its most liquid one.

可借鉴之处

A record round buys talent and compute, but the bet lands only if the product matches the thesis — and a talent-first lab can leak exactly what made it valuable when co-founders leave.

后续进展

As of September 2026, Thinking Machines Lab remains private, with more than $2B raised and a valuation above $12B. It has shipped two product lines since late 2025: the Tinker fine-tuning API and the open-weights Inkling model (975B parameters, Apache-2.0, 45 trillion training tokens), positioned as a base for customization rather than a frontier model. In March 2026 it signed an Nvidia partnership to deploy at least one gigawatt of Vera Rubin systems from 2027, with Nvidia also investing. Its biggest unresolved risk is people: four co-founders left, one to Meta and three back to OpenAI.

资料来源

发现哪里写错了？告诉我们。

轮到你了

你刚读完一家。说说你在做什么，看看谁在赌同一件事。

免费账号 · 3 次免费提问 · 不用绑卡