This time, Jensen Huang has extended the reach of AI infrastructure into the model layer.
In the past week, Nvidia unleashed two blockbuster deals.
First, it signed an acquisition agreement with Hugging Face, the world's largest hub for open-source models. Then, news broke that it was preparing to write a massive check of nearly $3 billion to Mira Murati, the former OpenAI CTO and one of Silicon Valley's most-watched figures.
Huang is betting on Murati's startup, Thinking Machines Lab.
The pre-money valuation has skyrocketed to at least $40 billion, with plans to raise $5 to $6 billion. Nvidia alone could cover about half of that.
This is not Nvidia's first investment but a continued doubling down.
Looking back at Thinking Machines Lab's growth speed, it's almost like it's been on steroids.
In the previous round in July 2025, it secured one of the largest seed rounds in history:
A $2 billion seed round, with a post-money valuation of $12 billion.
Just 14 months later, its pre-money valuation has surged to $40 billion, more than tripling.
If this $5 to $6 billion round is fully subscribed, the post-money valuation will approach $45 to $46 billion.
The crazier part is yet to come.
Against such a staggering valuation, the company's annualized revenue has just crossed $100 million.
A $40 billion valuation corresponds to a revenue multiple of roughly 400x.
Why would Nvidia willingly fork over another $2.5 to $3 billion for a company only 19 months old, from which several co-founders have already departed?
Two Years After Leaving OpenAI, Snagging a $40 Billion Valuation
It took Murati just 14 months to go from $12 billion to $40 billion.
In September 2024, Murati announced her departure from OpenAI. At the time, she was CTO and served as interim CEO during the days when Sam Altman was ousted by the board.
About five months later, Thinking Machines Lab burst onto the scene.
The founding roster read almost like an OpenAI alumni directory: co-founders included big names like Lilian Weng, John Schulman, Barret Zoph, and Luke Metz.
In July 2025, the $2 billion seed round landed.
a16z led the round, with Accel, Nvidia, AMD, and Jane Street participating. At the time, Murati didn't even have a substantial product to show; investors were betting on her personally.
Soon, the product delivered.
In October 2025, the first product, Tinker, launched, and opened to the public in December.
It's not a model; it's a fine-tuning platform: developers throw in their data, tune open-weight models into their own, and pay based on compute usage.
Then came an even bigger move.
On March 10, 2026, Thinking Machines and Nvidia officially announced a multi-year strategic partnership: deploying at least 1GW of Vera Rubin systems, with Nvidia also making a significant, undisclosed investment.
Jensen Huang (left) and Mira Murati (right). On March 10, 2026, the two sides announced a 1GW-scale Vera Rubin partnership, with Nvidia simultaneously completing an undisclosed investment.
Four months later, on July 15, 2026, they unveiled their first in-house flagship model, Inkling:
975 billion total parameters, 41 billion activated, up to 1 million token context, capable of processing text, images, and audio, with full open weights.
Then came the discussions for this $5 to $6 billion funding round.
The reason for needing so much money is that 1GW partnership mentioned above.
1GW is compute power at the level of frontier labs. Chips, networking, power, data centers – everything is configured at the gigawatt scale, each item costing a fortune.
The funding target has ballooned five or six times, from an initial rumored $1 billion to the current $5 to $6 billion, to sustain this 1GW ambition.
Where is this money coming from?
In the last round, Nvidia was just one of several participants, with a16z leading.
This round, Accel is the lead investor, and Nvidia may take about half, moving from a participant to the largest contributor.
Thus, the familiar picture emerges again: Nvidia is both the biggest investor in this round and the supplier of that 1GW system.
Money flows into Thinking Machines' account, and a large portion of it likely circles back to become Nvidia purchase orders.
Investment, procurement, and revenue form a closed loop within the same company.
The direction is clear enough.
Nvidia is no longer content with just supplying chips to established labs like OpenAI and Anthropic. It's starting to use equity to lock in the next generation of frontier players early.
There's another easily overlooked sentence in that March announcement: the two sides will jointly design training and inference systems optimized for Nvidia's architecture.
This means Thinking Machines' future models will be built from the ground up for Vera Rubin.
Murati's Strategy: Not Competing to be the #1 Model
Looking at Murati's moves over these 19 months, you'll see she has consistently avoided head-on competition with OpenAI and Anthropic over who has the strongest model.
She even frankly admitted in Inkling's launch blog:
Inkling is not the current strongest model, open or closed.
Thinking Machines' business model was set from the day Tinker launched; it sells the act of "making the model yours."
Companies bring their own data and business to its platform for training, fine-tuning, and inference, and it charges based on compute usage.
In Murati's own words: to create AI that people can shape and truly make their own.
This statement places Thinking Machines and OpenAI/Anthropic on two different paths.
The latter two are racing to build the strongest general-purpose closed models. Thinking Machines positions itself one layer below: not competing to be the #1 model, but building the infrastructure that helps everyone customize models.
Isn't this exactly the AI infrastructure Huang has talked about repeatedly?
For two years, Nvidia has talked about AI factories, AI infrastructure, and the token economy, aiming for precisely this kind of business where each new customer burns more compute.
At the GTC 2026 keynote on March 16, Jensen Huang stated compute demand "has grown by a million-fold," with at least $1 trillion in revenue expected from 2025 to 2027.
Understanding this makes Nvidia's willingness to pay clear.
Stacking Hugging Face, Nvidia Holds Two Cards
This isn't Nvidia's only major move this month.
On September 3, Nvidia announced its acquisition of Hugging Face for $12.9303 billion, with the agreement already signed.
Looking at the two deals together:
Hugging Face controls the distribution gateway and developer community for open models; Thinking Machines holds frontier models, a fine-tuning platform, and a research team.
One is the gateway, the other is the models. On the open-weights front, Nvidia has placed bets on both ends.
The two deals total nearly $16 billion, announced just days apart.
Inkling's full weights are hosted on Hugging Face, complete with a dedicated NVFP4 version for Nvidia's Blackwell.
Developers download Inkling from Hugging Face, fine-tune it into their own on Tinker, and then run it on Vera Rubin.
Gateway, model, fine-tuning platform, chips – every step is within one ecosystem.
This is precisely what Nvidia wants: using capital to lock in model companies early also locks in their future compute demand.
References:
https://techcrunch.com/tag/thinking-machines-lab/
https://www.theinformation.com/articles/thinking-machines-lab-talks-raise-billions-roughly-40-billion-valuation
https://thinkingmachines.ai/news/introducing-inkling/
https://blogs.nvidia.com/blog/nvidia-thinking-machines-lab/
This article is from the WeChat public account "Xin Zhi Yuan," author: ASI Revelation; Editor: Yuanyu
