LG Rents Nvidia's Robot Brain

Nvidia gave away the mind and kept the silicon. LG took the mind and kept the data.

The photograph from Santa Clara on August 13 shows two chief executives holding a toy. LG Corp's Kwang Mo Koo and Nvidia's Jensen Huang are posing with a miniature humanoid robot they have both signed, which is the sort of image that usually accompanies a memorandum of understanding worth nothing at all.

This one is worth reading closely. Under the agreement, LG will unveil a bipedal humanoid robot in the first quarter of 2027 running Nvidia's Isaac GR00T foundation model, build an AI factory reference site on Nvidia Vera Rubin in the first half of 2027, and complete an 80-megawatt AI factory in Cheonan, South Korea, by the first half of 2028. LG described the moment as a shift from strategic blueprints into full execution.

The interesting part is not the robot. It is the division of labor the deal encodes — and what it says about where value is settling in physical AI.

What each side is actually contributing

LG is bringing the body, and it is bringing all of it in-house. As Robotics & Automation News summarized the split, LG Electronics supplies the actuators, LG Innotek the sensors, LG Energy Solution the batteries. That vertical stack is unusual. Most humanoid startups assemble from a global supply chain of harmonic drives, motors, and cells, and discover that unit economics are hostage to suppliers they cannot influence. LG owns each link, which is the accumulated advantage of a conglomerate that has spent decades manufacturing appliances at volume.

Nvidia is bringing the mind: Isaac GR00T, its open reasoning humanoid foundation model, running on Jetson Thor compute, with Halos handling robotics safety. Nvidia describes GR00T as an open reference platform spanning data pipelines, the foundation model, simulation, middleware, and the Jetson Thor runtime for real-time inference and control. Nvidia is not building the robot. It is supplying the layer that makes the robot worth building.

The rest of the agreement is where the strategy gets legible. Within the year, LG will deploy CLOiD — a wheel-based robot, not a humanoid — onto the washing machine line at LG Electronics' Tennessee plant for validation in a live production environment. The companies will also build a robot data factory on LG CNS PhysicalWorks, LG's robot data platform, to handle continuous data collection, synthetic data generation, training, and verification. And the partnership extends to vehicles, with an AI-defined vehicle platform built on Nvidia DRIVE Hyperion — an expansion UPI characterized as an alliance spanning robots, AI factories and mobility.

Read in order, the sequence is: get boring robots into a real factory now, capture the data, use the data to train the interesting robot later, and build the compute to do the training on.

The data is the point

Humanoid robotics has a specific, well-understood bottleneck, and it is not mechanical. Actuators work. Balance is largely solved. Hands remain hard but are improving. What nobody has is enough real-world manipulation data — footage and telemetry of physical tasks performed in messy, variable, unstaged environments — to train models that generalize past a demo.

Language models had the internet. Robotics has no equivalent corpus, and the reason is that physical data has to be generated, one task at a time, in a real place, at real cost.

This is why the Tennessee deployment is the load-bearing part of the announcement and the humanoid reveal is the headline. A washing machine assembly line is a controlled environment with genuine variability, repeated operations, and a business that pays for itself while the data accrues. LG is not running a research pilot. It is putting robots where its own products get made, which means the robots have to be useful enough to justify their place on the floor — and every hour they spend there produces training data that a lab renting a warehouse cannot cheaply replicate.

The PhysicalWorks platform makes the intent explicit: continuous collection, synthetic augmentation, training, verification. That is a flywheel description, not a product description. LG is also developing its own Robot Foundation Model, and says the data and validation experience from this collaboration will strengthen it — while continuing to use Isaac GR00T in the meantime.

That last clause is the most strategically revealing thing in the release. LG is adopting Nvidia's brain today and building its own for later.

Why Nvidia gives the brain away

Isaac GR00T is open. Nvidia is handing the hardest software component of humanoid robotics to a manufacturer that has openly stated it is developing a competing model. This looks like a mistake and is not one.

Nvidia does not need to own the robot foundation model. It needs robot foundation models to exist, proliferate, and be expensive to train. Every one of them — GR00T, LG's RFM, Google DeepMind's Gemini Robotics 2, whatever a hundred startups produce — consumes Nvidia silicon during training and, in most cases, ships on Nvidia silicon at the edge. The 80-megawatt Cheonan facility is the tell. LG is not just licensing a model; it is committing to a data center whose purpose is training physical AI, built on Vera Rubin.

The model is the loss leader. The compute is the business, at both ends: the training cluster and the Jetson module inside every unit. It is the logic that has already turned Nvidia's silicon into something closer to an asset class than a component.

This is the same structure Nvidia has used since CUDA — give away the layer that makes the hardware useful, charge for the hardware, and let the ecosystem's success compound into demand. It worked in graphics, then in deep learning, and it is being run again in robotics with a decade of practice behind it.

Who this lands on

For humanoid startups, the competitive picture just got worse in a way that is easy to underestimate. A venture-funded humanoid company competes on engineering talent and capital. It cannot easily compete against a manufacturer that already owns actuator, sensor, and battery production, already operates factories to deploy into, and can absorb years of losses because the robotics division is not the company. The advantage startups retain is focus and speed. Against LG's balance sheet, focus has to be worth a lot — and public-market enthusiasm for pure-play robotics listings may not survive contact with a conglomerate that can lose money for a decade.

For Nvidia, this extends a pattern that has become the company's defining move: partner deeply with an industrial incumbent, supply the software free, and lock in the compute on both the training and inference sides. Robotics is the third act after graphics and data-center AI, and it is the one with the largest addressable physical footprint if it works at all. The scale of Nvidia's existing data-center machine is what makes a speculative third act affordable.

For LG, the risk is timing rather than capability. The company has the manufacturing base and now has the model. What it does not control is whether general-purpose humanoids reach commercial viability on anything close to a 2027–2028 schedule. The CLOiD deployment is a sensible hedge — wheeled robots on a factory line are a real business today regardless of what happens to bipeds — but the 80-megawatt facility is a capital commitment sized for a market that does not yet exist.

For the physical AI field generally, the LG deal points at the shape of the winners. Not the best model and not the best mechanism, but whoever controls a continuous source of real-world task data. That favors incumbents with factories, warehouses, logistics networks, and retail floors — companies whose operations generate the training corpus as a byproduct of doing what they already do.

What could go wrong

The 2027 unveiling is a demo, not a product. Every humanoid program in the world can produce an impressive video; almost none can produce a robot that does useful work, unsupervised, at a cost that beats the alternative. Nothing in this announcement changes that arithmetic, and the gap between "unveiled in Q1 2027" and "deployed at scale" is where most robotics timelines have historically gone to die.

The 80-megawatt Cheonan facility is also a real bet on an unproven premise — that robot foundation models will require training compute on the scale that language models did. That may be right. It may also turn out that physical AI is bottlenecked by data quality rather than parameter count, in which case a very large cluster is an expensive way to discover you needed more robots instead.

And LG's dual-track approach — use GR00T now, build an RFM later — is coherent but not costless. Switching foundation models mid-program means rebuilding the tooling, the evaluation harness, and much of the training pipeline around them. Companies that plan to swap out a dependency later often find the dependency has grown roots.

The zoom-out

The instructive thing about this deal is which layer each party fought for.

Nvidia gave away the model — the thing that looks most like intelligence, the part that gets the press — and kept the silicon underneath it. LG committed to a foundation model it does not own while quietly building its own and, more importantly, securing the factory floor that generates the data neither model can be trained without.

Both are correct about where durable advantage sits in physical AI, and neither thinks it is the robot. The model will commoditize; it always does. The compute is defensible while Nvidia's lead holds. The data is defensible as long as you own the place where work actually happens.

That is the asset worth watching. In a field where every demo looks the same, the companies that win will be the ones that already had somewhere to put the robot.

Anyone can build a robot. Very few people own a floor to teach it on.

Frequently Asked Questions

What did LG and Nvidia agree to?

The two companies signed a memorandum of understanding on August 13, 2026 at Nvidia's Santa Clara headquarters. LG will unveil a bipedal humanoid robot in Q1 2027 built on Nvidia's Isaac GR00T foundation model, deploy wheel-based CLOiD robots at its Tennessee plant, build an AI factory reference site on Nvidia Vera Rubin, and construct an 80MW AI factory in Cheonan by the first half of 2028.

What is Nvidia Isaac GR00T?

Isaac GR00T is Nvidia's open reasoning foundation model for humanoid robots. It provides the high-level reasoning and control layer, running on Nvidia's Jetson Thor embedded computing platform with the Halos system handling robotics safety. Nvidia makes the model openly available rather than charging for it directly, monetizing the underlying compute instead.

Why is LG building its own robot foundation model too?

LG says the data and validation experience from the Nvidia collaboration will strengthen its in-house Robot Foundation Model, which is currently under development. The company plans to use Isaac GR00T actively while continuing to advance its own capabilities — a dual-track approach that captures near-term progress without permanently ceding the software layer.

Why does robot training data matter so much?

Unlike language models, which trained on existing internet text, robotics has no equivalent ready-made corpus. Real-world manipulation data must be generated task by task in physical environments, at real cost. That makes access to operating factories, warehouses, and production lines a durable advantage that capital alone cannot easily replicate.

Editor's note — sources:

• PR Newswire, LG press release — https://www.prnewswire.com/news-releases/lg-to-unveil-its-next-gen-humanoid-robot-built-on-nvidia-isaac-gr00t-302851652.html
• NVIDIA Developer, Isaac GR00T — https://developer.nvidia.com/isaac/gr00t
• Google DeepMind, Gemini Robotics 2 — https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
• UPI — https://www.upi.com/Top_News/World-News/2026/08/14/lg-nvidia-alliance-robots-physical-ai-partnership/7361786742323/
• Robotics & Automation News — https://roboticsandautomationnews.com/2026/08/14/lg-to-unveil-new-nvidia-powered-humanoid-robot-in-early-2027/104173/