Originally published on andrew.ooo
TL;DR
Thinking Machines Lab just secured one of the most significant AI infrastructure deals of 2026. Nvidia is making a "significant investment" and committing 1 gigawatt of Vera Rubin compute—enough power for 750,000 homes and worth approximately $50 billion in infrastructure.
The kicker? Mira Murati built this leverage with just 100 employees and a $12 billion valuation. That's $120 million in valuation per employee—making it one of the most capital-efficient AI companies ever built.
The Numbers That Matter
| Metric | Value |
|---|---|
| Valuation | $12 billion (potentially $50B+ after Nvidia deal) |
| Total Raised | $2 billion |
| Employees | ~100 |
| Valuation Per Employee | $120 million |
| Funding Per Employee | $20 million |
| Nvidia Compute Commitment | 1 gigawatt |
| Infrastructure Value | ~$50 billion |
| Time to $12B Valuation | 5 months |
The Mira Murati Playbook: From OpenAI CTO to $12B Founder
The Exit That Launched a Competitor
In September 2024, Mira Murati quietly departed OpenAI where she served as Chief Technology Officer. By February 2025, she launched Thinking Machines Lab with a clear mission: make AI "more widely understood, customizable, and generally capable."
Poaching the Best Talent
Before raising a single dollar, Murati recruited approximately 30 researchers from OpenAI, Meta AI, and Mistral AI:
- John Schulman (Chief Scientist) - OpenAI co-founder
- Barret Zoph - Former OpenAI VP of Research
- Lilian Weng - Former OpenAI VP
- Advisors: Bob McGrew (ex-OpenAI CRO) and Alec Radford
The Record-Breaking Seed Round
In July 2025, Andreessen Horowitz led a $2 billion seed round at a $12 billion valuation. Other investors included Nvidia, AMD, Cisco, Jane Street, and even the Government of Albania ($10 million).
The Nvidia Gigawatt Deal: What It Actually Means
The Scale is Staggering
On March 10, 2026, Nvidia and Thinking Machines Lab announced:
- "Significant investment" from Nvidia (amount undisclosed)
- 1 gigawatt of Vera Rubin systems—Nvidia's most advanced chips
- Multi-year compute commitment for training and inference
Putting a Gigawatt in Perspective
- 1 gigawatt = Power for approximately 750,000 U.S. homes
- $50 billion = Estimated cost to build and operate infrastructure at this scale
- Vera Rubin = Nvidia's next-gen AI systems, successor to Blackwell
$120M Valuation Per Employee: The AI Efficiency Thesis
| Company | Valuation | Employees | Per-Employee Valuation |
|---|---|---|---|
| Thinking Machines Lab | $12B | ~100 | $120M |
| Cursor (Anysphere) | $13B | ~100 | $130M |
| OpenAI | $840B | ~3,500 | $240M |
| Anthropic | $76B | ~1,100 | $69M |
What Makes This Possible?
- Elite talent density - 100 researchers who each worked on billion-dollar projects
- Infrastructure partnerships - Why build data centers when Nvidia will supply compute?
- API-first business model - Tinker launched October 2025 as a fine-tuning API
- Public benefit corporation - Lower pressure for rapid monetization
Key Takeaways for Founders
- Elite talent trumps headcount - 100 exceptional people beat 1,000 average ones
- Infrastructure partnerships are strategic - Don't build when you can partner
- Governance matters - Murati's voting rights prevent hostile board actions
- Speed to valuation - 5 months from founding to $12B
- Mission focus - "Customizable AI" is a clear, differentiating vision
Read the full analysis with sources and FAQ at andrew.ooo
Top comments (0)