Chipmakers are now buying the AI model makers

In the space of four weeks, the two leading suppliers of AI accelerators each agreed to buy a company whose product is models, and the people who build them, rather than silicon. Nvidia announced on 3 September that it would acquire Hugging Face, the open-source model hub, in a deal valued at about $13bn. On 28 September AMD said it would buy World Labs, the spatial-intelligence lab founded by Fei-Fei Li, for about $8.2bn in stock.

The two targets do different things, but the deals rest on a shared premise: a chip supplier now needs first-hand knowledge of the models that will run on its hardware, and the quickest way to get it is to own a team that builds them.

What AMD is paying for

World Labs develops models that generate, reconstruct and simulate interactive 3D environments from text, image and video inputs. Its first product, Marble, is aimed at entertainment content and at simulated environments for training robots. The company launched publicly in September 2024 with $230m of funding at a $1bn valuation, and in February raised a $1bn round that included $200m from Autodesk and participation from Nvidia, Fidelity, Emerson Collective and AMD itself. It has not confirmed its valuation after that round; earlier reports said it was seeking about $5bn.

AMD's price is therefore a large step up even on recent private marks. The number of shares it will issue is not fixed and will be set by its 10-day volume-weighted average price shortly before closing, which is expected by the end of 2026. Li will become executive vice-president and chief scientist, reporting to chief executive Lisa Su. According to CNBC, as cited by CRN Asia, it is AMD's second-largest acquisition, behind the roughly $50bn purchase of Xilinx in 2022. "Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving," Su said.

Nvidia's move up the stack

Nvidia's deal is larger and different in kind. Hugging Face has about 18 million users and runs an open-source hub for collaborative machine learning where developers share models and data. Semafor put the price at $12.9bn; according to Nvidia's SEC filing, as reported by CRN Asia, about $11.9bn is payable to Hugging Face stockholders, with an equity-based retention programme of up to $1bn for staff who join Nvidia. Closing is expected in the first half of 2027.

Jensen Huang said Hugging Face "will remain an open platform for the entire AI ecosystem", and Nvidia has framed the deal as protecting open-source developers. It is Nvidia's second-largest transaction after a $20bn non-exclusive agreement last December to buy assets from chipmaker Groq, and follows a $6bn structured agreement in August with Poolside, which builds open-weight models. Valuations at the model layer are not confined to chipmakers: in August Stripe agreed to buy OpenRouter, a model aggregator, for $7.5bn.

Why hardware firms want model makers

Analysts frame the AMD deal less as a product purchase than as an early-warning system for chip design. "This isn't about AMD buying a model company," Matt Kimball of Moor Insights & Strategy told Data Center Knowledge, arguing that the aim is to understand future compute and memory requirements early enough to shape low-level architecture decisions. Brendan Burke of the Futurum Group pointed to talent: "You can't hire a more accomplished researcher than Dr. Fei-Fei Li."

There is also a competitive gap to close. Nvidia already offers open-weight world models through its Cosmos family, while AMD has so far released only text and video models publicly. AMD estimates the physical-AI market could reach $200bn by 2035, and world models matter there because real-world robot training data is scarce and synthetic data must fill the gap.

The questions buyers inherit

The strategic logic is coherent, but both deals carry obvious tensions. Hugging Face's value lies in neutrality across hardware and clouds, which an owner selling accelerators must take care to preserve. World Labs is a research organisation that AMD says will continue to focus on model research after closing, so the return depends on research insight reaching product roadmaps rather than on revenue. Both deals still need regulatory approval. If they succeed, the boundary between building AI hardware and building AI models is likely to blur further, and the remaining independent labs will find chipmakers among their most motivated bidders.