This time, Old Huang has extended AI infrastructure into the model layer.

Author and source: AI New Era

Last week, NVIDIA landed two blockbuster deals in succession.

Just after signing an acquisition agreement with Hugging Face, the world’s largest open-source model hub, news emerged that the company is preparing to offer Mira Murati, OpenAI’s former CTO and one of Silicon Valley’s most prominent figures, a check worth nearly $3 billion.

Old Huang bet on Murati’s founded Thinking Machines Lab.

Pre-money valuation soared to at least $40 billion, with plans to raise $5 to $6 billion, potentially with NVIDIA alone taking about half.

This is not NVIDIA's initial investment, but rather an ongoing increase in commitment.

Looking back at the growth rate of Thinking Machines Lab, it felt like it was on steroids.

The previous round in July 2025 secured one of the largest seed rounds in history:

$2 billion seed round, post-money valuation of $12 billion.

Just 14 months later, the pre-investment valuation soared to $40 billion, more than tripling.

If this round of $5 billion to $6 billion is successfully fully subscribed, the post-money valuation will approach $45 billion to $46 billion.

Even crazier things are yet to come.

With such an astonishing valuation, the company's annual revenue has just surpassed $100 million.

A $40 billion valuation corresponds to approximately a 400x revenue multiple.

What makes a company, founded just 19 months ago and already having lost several co-founders, worth having NVIDIA willingly invest another $2.5 to $3 billion?

Left OpenAI two years ago, now valued at $40 billion

Murati went from 1.2 billion to 4 billion in just 14 months.

In September 2024, Murati announced her departure from OpenAI. At the time, she was CTO and served as interim CEO during the days when Altman was removed by the board.

Approximately five months later, Thinking Machines Lab emerged.

The list of startups reads almost like a directory of former OpenAI employees, including big names such as Lilian Weng, John Schulman, Barret Zoph, and Luke Metz among the co-founders.

In July 2025, $2 billion in seed funding was received.

a16z led the round, with Accel, NVIDIA, AMD, and Jane Street participating as follow-on investors. At the time, Murati didn’t even have a single polished product—investors were betting on her personally.

Soon, the product delivered its results.

In October 2025, the first product, Tinker, launched, with full availability in December.

This is not a model, but a fine-tuning platform: developers upload their own data to customize open-weight models and pay based on compute usage.

Next, an even bigger move is coming.

On March 10, 2026, Thinking Machines and NVIDIA announced a multi-year strategic partnership: deploying at least 1 GW of Vera Rubin systems, alongside a significant, undisclosed investment by NVIDIA.

Jensen Huang (left) and Murati (right)

On March 10, 2026, both parties officially announced a 1 GW Vera Rubin collaboration, with NVIDIA simultaneously completing an undisclosed investment.

Four months later, on July 15, 2026, they unveiled their first proprietary flagship model, Inkling:

975 billion total parameters, 41 billion activated, up to 1 million token context length, supports text, images, and audio input, with full weights open.

Then there are the ongoing negotiations for this $5 billion to $6 billion funding round.

The cost is this high due to the 1 GW partnership mentioned above.

1 GW is computing power at the cutting-edge laboratory level. Chips, networks, power, and data centers are all configured at the gigawatt scale, each representing an astronomical figure.

From an initial rumored $1 billion to the current $5 to $6 billion, the funding has grown five to six times over to support this 1 GW ambition.

Where will this money come from?

In the previous round, NVIDIA was only a co-investor; the lead investor was a16z.

In this round, Accel has become the lead investor, and NVIDIA may now cover approximately half, transitioning from a follow-on investor to the largest contributor.

Thus, the familiar scenario reappears: NVIDIA is both the largest investor in this round and the supplier of the 1 GW system.

The money entered Thinking Machines' account, and a large portion will likely circle back and become an order for NVIDIA.

Investment, procurement, and revenue are closed within the same company.

The direction is clear enough.

NVIDIA is no longer content with merely supplying chips to established labs like OpenAI and Anthropic—it is now using equity stakes to secure early partnerships with the next generation of frontier players.

The March announcement also included a lesser-known statement: both parties will jointly design training and inference systems tailored for NVIDIA's architecture.

This means that Thinking Machines' future models will be fundamentally built around Vera Rubin.

Murati's approach is not about competing for model supremacy.

Through Murati’s 19-month strategy, you’ll notice she has consistently avoided directly competing with OpenAI and Anthropic over which model is stronger.

Even outright acknowledge in Inkling's release blog:

Inkling is not the strongest model currently available, whether open-source or proprietary.

Thinking Machines' business model was established from the day Tinker launched—it sells the idea of "making the model yours."

Enterprises bring their own data and business processes to its platform to train, fine-tune, and run inference models, and they are charged based on their compute usage.

In Murati’s own words: build AI that people can shape and claim as their own.

This statement places Thinking Machines on a different path from OpenAI and Anthropic.

The latter two are competing for the strongest general-purpose closed-source models, but it positions itself one level below: not vying for the top model, but building the infrastructure to help everyone customize their own models.

Isn't this exactly the AI infrastructure that Old Huang has mentioned multiple times?

NVIDIA has been talking for two years about AI factories, AI infrastructure, and token economies—all aiming for a business model where each new customer consumes additional computing power.

At the GTC 2026 keynote on March 16, Jensen Huang stated that computing demand has increased by a factor of one million, with revenue between 2025 and 2027 expected to reach at least $1 trillion.

Understanding this, it becomes clear why NVIDIA is willing to spend money.

Stacking Hugging Face, NVIDIA holds two cards.

This isn't NVIDIA's only major move this month.

On September 3, NVIDIA officially announced the acquisition of Hugging Face for $12.9303 billion, and the agreement has been signed.

Look at the two transactions together:

Hugging Face controls the distribution channel for open models and the developer community; Thinking Machines holds cutting-edge models, a fine-tuning platform, and a research team.

One end is the entry point, the other is the model. NVIDIA has bet on both ends of the open-weights spectrum.

Together, the two transactions amount to nearly $16 billion, occurring just a few days apart.

The full weight of Inkling is available on Hugging Face, with a dedicated NVFP4 version optimized for NVIDIA Blackwell.

Developers download Inkling from Hugging Face, customize it on Tinker, and run it on Vera Rubin.

The entry point, model, fine-tuning platform, and chip—all are part of one ecosystem.

This is exactly what NVIDIA wants: locking in model companies with capital upfront, thereby securing their future computing power demands.