Thinking Machines Hits $50B Valuation: Mira Murati’s Lab Secures 1GW of Vera Rubin Compute from NVIDIA

The $50 Billion Signal: Compute Is the New Moat

Thinking Machines Lab, founded in February 2025 by former OpenAI CTO Mira Murati, reached a reported $50 billion valuation in early 2026 — roughly a 4x jump from its $12 billion seed valuation booked in July 2025. On March 10, 2026, NVIDIA announced a significant investment in Thinking Machines alongside a 1-gigawatt compute deal based on its next-generation Vera Rubin architecture, with the first capacity scheduled to come online in early 2027.

The combination — a $50B valuation and a gigawatt-scale compute reservation — positions Thinking Machines as one of a small number of AI labs operating at frontier scale outside the incumbents (OpenAI, Anthropic, Google DeepMind, xAI). It also crystallizes a thesis that is now unmistakable: in 2026, the moat is compute contracts, not model weights.

The Murati Lab’s First Year, By the Numbers

The company’s first twelve months are a compressed case study in how fast frontier AI funding moves when the founder is credentialed:

• February 2025: Founded by Mira Murati, with a core team of around 30 researchers and engineers pulled from OpenAI, Meta, Mistral, and other labs. Founding group includes Barret Zoph (former OpenAI VP Research, Post-Training), Lilian Weng (former OpenAI VP), John Schulman (OpenAI co-founder, briefly at Anthropic), Andrew Tulloch, and Luke Metz.
• July 2025: Closed a $2 billion seed round at a $12 billion valuation, led by Andreessen Horowitz with NVIDIA, AMD, Cisco, Accel, and Jane Street participating. It was reported as the largest seed round in Silicon Valley history — roughly 4x the prior record.
• October 2025: Launched Tinker, the company’s first product — a Python API for distributed LLM fine-tuning, in private beta. The launch came with the Tinker Cookbook, an open-source library.
• March 2026: NVIDIA announces significant additional investment plus a 1GW Vera Rubin compute commitment; valuation reported approaching $50 billion.

What the NVIDIA Deal Actually Contains

NVIDIA’s Vera Rubin architecture is the company’s next major generation, pairing high-performance Rubin GPUs with the new Vera CPU engineered for data orchestration in large AI workloads. Industry executives have estimated that a 1-gigawatt AI data center built on Vera Rubin would cost roughly $50 billion to construct and operate — a number that is, not coincidentally, the same order of magnitude as Thinking Machines’ new valuation.

Three features of the deal matter beyond the headline number:

1. Multi-year compute reservation. A gigawatt commitment is not a spot purchase — it is a structured contract ensuring capacity over years, with NVIDIA as both vendor and equity investor.
2. Early 2027 starting point. Compute capacity comes online over time; the early 2027 start gives Thinking Machines a compute runway lined up for frontier-scale training runs in 2027 and 2028.
3. NVIDIA as investor-and-supplier. The same entity financing a meaningful portion of the round is the one selling the chips, which aligns incentives tightly but also concentrates supply-chain and strategic risk.

Why This Is About Infrastructure, Not Model Architecture

Thinking Machines’ public positioning emphasizes open research and collaborative AI. Tinker — its first product — is an API for fine-tuning open-weight LLMs, not a closed frontier model. The company has published research on open-source techniques rather than locking capability behind a single product.

Against that backdrop, a 1GW compute reservation reads differently than it would for a closed-model competitor. It signals that Thinking Machines is building for scale regardless of whether its commercial model converges on frontier training, inference-heavy products, or scientific research infrastructure. The compute is the moat; what gets built on top can evolve.

What This Means for the Global AI Market

Three downstream implications flow from the deal:

1. Compute scarcity is repricing equity. A startup’s ability to lock in frontier-scale compute is now a valuation input in its own right. Labs without multi-year reservations are exposed to spot-market pricing that is already constraining OpenAI, Anthropic, and others.
2. NVIDIA’s investor-supplier model is consolidating. NVIDIA has now taken direct stakes in multiple compute-hungry AI companies, including Nscale (the European AI infrastructure player that just closed a $2 billion Series C at $14.6B valuation). The pattern is deliberate — NVIDIA is using equity to anchor long-term GPU demand.
3. Frontier lab count is narrowing. Fewer than a dozen entities globally can now credibly claim gigawatt-scale compute access. Thinking Machines joins that shortlist after roughly 13 months of existence.

How This Looks From Emerging Markets

For markets like Algeria, Morocco, or Egypt where sovereign AI strategies are still being built, the signal from Thinking Machines is not that local labs should attempt frontier-scale compute. It is that the global gap between compute-haves and compute-have-nots is widening, fast. Local AI strategies in 2026 should concentrate on applied AI, vertical fine-tuning, and inference-grade deployment — the layer where Tinker-style tooling, not gigawatt compute, is the relevant unit.

Singapore’s sovereign compute investments, discussed often in North African policy circles as a benchmark for small-country tech strategy, share the same logic: do not compete on raw scale; concentrate on deployable layers built on top of global infrastructure. Thinking Machines’ trajectory validates the direction.

Sources & Further Reading

• Nvidia invests in Mira Murati’s Thinking Machines Lab, plans 1GW Vera Rubin AI compute deployment — The Tech Portal
• Thinking Machines Lab’s Mira Murati commits to a gigawatt of Nvidia compute — Axios
• Nvidia partners with Mira Murati’s Thinking Machines Lab for 1 gigawatt of Rubin GPUs — Sherwood News
• Mira Murati’s Thinking Machines Lab is worth $12B in seed round — TechCrunch
• Thinking Machines’ first official product is here: meet Tinker — VentureBeat