Nvidia's $12.9bn wake-up call on physical AI

What happened

On 2 September, Nvidia signed a definitive agreement to acquire Hugging Face for approximately $12.9bn ($11.9bn to shareholders plus up to $1bn in employee retention), with closing expected in H1 2027 subject to antitrust clearance. Hugging Face is where open AI lives: roughly 3 million models and 1 million applications, used by 18 million developers and 200,000 companies. Nvidia pledged that its compute will not be required to build on or deploy through the platform, which stays open to AMD, Intel and other silicon.

After the roughly $20bn Groq and $6bn Poolside licensing deals, this is Nvidia's third move in nine months along the same axis: inference silicon, model tooling, and now model distribution.

Impact on our Investment Case

The consensus read is right, and already priced

Nvidia's structural risk is concentration: its largest buyers, Microsoft, Meta, Amazon, Alphabet and OpenAI, are the only ones that can justify designing their own accelerators. Open-weight models are the counterweight, diffusing AI capability to organizations that will never design a chip and that standardize on CUDA by default; that is why the hardware-agnostic pledge costs Nvidia little. And as we argued in July around Kimi K3, when frontier capability is free to download, the model stops being the asset. Nvidia did not buy a model; it bought the repository, the tooling and the developer habit.

The read-through spans our portfolio's two largest clusters: semiconductors, which gain from a larger open-weight ecosystem regardless of whose accelerator runs it (e.g. Broadcom, Marvell), and data and cloud platforms, where more open-weight workloads mean more demand for data and monitoring tools (e.g. Snowflake, Datadog). All of this is the part the market saw within a day. The part it has not priced is physical.

Physical AI is becoming more software-like

Coverage of the deal has been almost entirely about language models. Yet robot intelligence now runs on the same recipe. Vision-language-action models like Nvidia's GR00T are trained on demonstration data: shown enough examples of a task, real or simulated, they learn policies that transfer across robot bodies and tasks. It is the shift that produced ChatGPT, applied to movement: robot capability stops being a bespoke engineering project and becomes a model checkpoint, priced in downloads and fine-tuning runs. Once intelligence is a downloadable artefact, the place where those artefacts are shared, ranked and improved becomes strategic infrastructure.

It also explains why robotics is Nvidia's logical next frontier. Every robot generates compute demand three times over, world-model training, simulation, and on-board inference, and the buyers, robot makers, industrial groups and integrators, are far too fragmented to ever design their own silicon: the opposite of the hyperscaler concentration problem.

The robotics value chain has four layers: components and actuators, where China's scale advantage is decisive; robot makers, where hardware differentiates less each year as intelligence migrates into software; foundation models and data, where the real scarcity sits; and vertical integrators, who own the customer relationships and domain data. As in generative AI, the model layer commoditizes, and durable value accrues below it, in compute and simulation, and above it, in deployment and integration.

Hugging Face fills a missing layer in Nvidia's stack

Nvidia has spent years assembling the stack for this market, from Isaac simulation tools to Cosmos world models to Jetson on-robot computers. What it lacked was the community and data layer where robot intelligence actually gets built and shared. That is what comes with the acquisition: LeRobot, the open physical-AI platform the two companies have built together since 2025, carries Isaac GR00T, Nvidia's open vision-language-action model, and the largest open physical-AI dataset, more than 350,000 trajectories downloaded over 15 million times. Robotics has been held back by data scarcity and fragmented tooling far more than by hardware cost; a single open repository where trajectories, policies and evaluation converge is the closest thing to a solution the field has produced. Owning it puts Nvidia at the centre of physical AI on the terms it holds in generative AI: not by selling the models, but by being where they are built.

The competitive backdrop explains the timing. Robotics is the one AI domain where China is not chasing: 54% of global industrial robot installations in 2024 according to the International Federation of Robotics, humanoid makers (e.g. Yushu Technology Co Ltd, UBTech) shipping at price points Western rivals cannot match, and open robot-data initiatives of its own, AgiBot World being the most visible. The two scarcest inputs in physical AI, hardware and deployment data, are being generated at scale in China.

Nvidia cannot control where robots are built, so it is buying the place where robot intelligence is built. If LeRobot becomes the default repository, Chinese hardware feeds a stack anchored to Nvidia tooling; if access is ever restricted, the effort splinters into a domestic equivalent, accelerating the vertical integration of the Chinese automation names our portfolio already holds.

None of this is in the price because the revenue is years out: downloadable robot intelligence expands the addressable market for the entire automation cluster (e.g. Estun, Inovance), and slower still for medical technology, where regulatory cycles dominate. It is a change to the shape of the opportunity, and the market does not pay for shape until it has to.

The China question is the one real risk

Chinese technology and automation names are a substantial part of our AI & Robotics portfolio, cross-cutting the clusters above, and this is where the deal cuts against us. Hugging Face is today the main distribution channel for Chinese open-weight models (Qwen, DeepSeek, GLM, Kimi). Under US ownership it becomes a US-controlled asset in a sector where export policy has repeatedly been applied to software as well as silicon; Nvidia's own filing flags the risk. Access conditions imposed through the review are a plausible outcome, not a tail scenario.

The effect is not uniformly negative: restricted Western distribution would accelerate the domestic substitution already driving names like Naura, Cambricon, and Alibaba's ModelScope operates at scale. Baidu is the most exposed on the distribution side. The near-term signal is the Trump-Xi meeting of 24 September, which will set the tone for how technology restrictions are handled into the antitrust review.

Our Takeaway

When the company with the best visibility into AI demand redirects close to $39bn toward a single premise, the premise is the news: value in AI accrues to compute below the model layer and to deployment and applications above it, and the next place this logic plays out is the physical world.

Our portfolio is exposed to capture this shift. Semiconductors (35.1%) and data and cloud platforms (22.8%), the two layers the deal reinforces, together represent roughly 58% of the portfolio. Enterprise software, where the effect is second-order, accounts for 23.3%, and automation, the slow-burn beneficiary, for 8.96%. Chinese technology and automation names, where the risk sits, represent 24.6%, a figure that cross-cuts the other clusters. Nvidia itself is held at 3.65%.

This is also precisely why we are not adding to Nvidia. The deal will not close before H1 2027, with antitrust review in between. More fundamentally, its benefits accrue to the entire open-weight and physical-AI ecosystem, which our portfolio is already built to capture across the value chain; leveraging the single name would add concentration risk, not exposure to the thesis. What does warrant attention is the Chinese sleeve, which now carries a new risk: not to demand, which domestic substitution supports, but to the distribution channel behind Chinese open-weight models' international footprint. That is what we are monitoring into 24 September.