How much does NVIDIA's Isaac GR00T Reference Humanoid Robot cost?

The full reference design — including a Unitree H2 Plus body ($29,900 list), dual Sharpa Wave tactile hands, and NVIDIA Jetson AGX Thor T5000 compute module ($2,999 at production volume) — is estimated at approximately $95,000. The Jetson Thor T4000, a lower-cost option, starts at $1,999 per module.

Which humanoid robot companies use NVIDIA's Jetson platform?

NVIDIA lists Amazon Robotics, Boston Dynamics, Agility Robotics, Figure AI, Meta, Caterpillar, Medtronic, and Hexagon among the 2 million+ developers using its robotics stack. Tesla is the only major humanoid maker building its own compute silicon (the FSD chip), making it the sole exception to the NVIDIA compute dependency.

NVIDIA Sells a $2,999 Robot Brain to Every Major Humanoid Maker Except One. The Module Revenue Is 0.13% of Sales. The Lock-In Is the Point.

$2,999. That is what NVIDIA charges for the Jetson Thor T5000 production module at volume — the Blackwell-based compute brain delivering 2,070 FP4 teraflops of AI performance from a package that draws between 40 and 130 watts, the kind of on-device reasoning power that lets a robot distinguish a coffee mug from a wine glass by touch alone. Buy a thousand or more and you get the wholesale rate. It ships inside humanoid robots built by Amazon Robotics, Boston Dynamics, Agility Robotics, Figure AI, Caterpillar, and at least two million other developers working on some form of physical AI.

One company is conspicuously absent from that list. Tesla builds its own compute silicon for Optimus.

Everyone else pays the NVIDIA tax.

On June 1, Jensen Huang walked onto the GTC Taipei stage and unveiled what he called the first open humanoid robot reference design: the Isaac GR00T Reference Humanoid. It combines a Unitree H2 Plus body — six feet tall, 150 pounds, 31 degrees of freedom — with dual Sharpa Wave five-fingered hands adding another 44 degrees of freedom, and the Jetson Thor T5000 as the onboard brain. Stanford, ETH Zurich, Ai2, and UC San Diego signed on as the first research partners. Unitree, which is simultaneously pursuing a $610–620 million IPO on Shanghai's STAR Market, will sell the complete system in late 2026.

The estimated price: $95,000.

Two days earlier, Sam Altman had posted on X that OpenAI was launching a dedicated robotics division — humanoid hardware, controls, machine learning — aimed at building robots that help skilled workers first, then eventually a personal robot for everyone. Tesla's market cap dropped $75 billion in a single session.

Seventy-two hours. That is how long it took for three companies to announce three fundamentally different strategies for winning humanoid robots, and the quiet arithmetic underneath all three points in exactly one direction.

The module math doesn't matter

Analysts expect somewhere between 50,000 and 100,000 humanoid robot shipments in 2026, so take the aggressive end and assume every single unit ships with a Jetson Thor T5000 at the $2,999 production price. That yields $299.9 million in module revenue.

NVIDIA's most recent quarter brought in $56.3 billion; annualized, the company generates roughly $224 billion in revenue, which means robot brain modules at the most optimistic volume projection represent 0.13% of that number. Huang could lose the entire humanoid compute market and his CFO wouldn't notice until the next board deck.

This is not a revenue play, and it never was.

NVIDIA's stock gained 6.26% on the day of the GR00T announcement, adding approximately $317 billion in market capitalization — a gain larger than the entire market value of 95% of the Fortune 500, bolted on in a single trading session because of a robot that doesn't ship until late 2026. Wall Street did not model $300 million in Jetson Thor sales. It modeled the ecosystem logic that the module revenue obscures.

The razor and the blade

Every humanoid robot is a training-data factory. Not metaphorically. Literally a machine whose primary economic output is the data it generates while doing its job. Each one walks through real environments, grasps physical objects, recovers from slips, adjusts force on irregular surfaces, and generates the proprioceptive, visual, and tactile data that locomotion and manipulation models need. That data is worthless without compute to process it, and this is where the economics shift away from module sales toward something the balance sheet doesn't make obvious.

Training a single locomotion skill through simulation-to-real transfer — the dominant pipeline for humanoid dexterity — requires on the order of 500 to 1,000 GPU-hours. The range is wide because complexity varies: a simple pick-and-place task might converge in 200 hours, while whole-body dynamic recovery from a push can consume 2,000 or more. For the calculation that follows, we use 500 GPU-hours as a conservative midpoint.

A production humanoid needs roughly 50 discrete skills: walking on flat ground, walking on slopes, stair ascent and descent, object recognition, grasping rigid objects, grasping deformable objects, opening doors with handles, opening doors with knobs, carrying loads, and so on, each with its own policy network. Ten major companies — Tesla, Figure AI, Boston Dynamics, Agility, 1X, Apptronik, Unitree, AGIBOT, Fourier, and Sanctuary AI — are iterating on these skill sets simultaneously, running thousands of training cycles per skill as they tune policy parameters, adjust reward functions, and retrain on newly collected real-world data.

Call it 50 skills × 10 companies × 1,000 training iterations per skill per year × 500 GPU-hours per run. That is 250 million GPU-hours of annual training compute demand, generated by an industry that barely shipped 16,000 units last year.

At prevailing cloud rates of roughly $3.50 per GPU-hour for NVIDIA H100-class hardware, that training compute is worth $875 million annually. Double the GPU-hours to account for simulation rendering in NVIDIA's own Omniverse and Isaac Sim environments, and the number approaches $1.75 billion. This estimate is deliberately rough, but the point is the order of magnitude: the training ecosystem generates three to six times more NVIDIA revenue than the modules inside the robots themselves.

The module is the razor, and the training pipeline — the simulation-to-real loop that every humanoid maker needs to run, at scale, continuously, on NVIDIA GPUs — is the blade.

The one company that escapes

Tesla runs custom silicon — its Optimus Gen 3 uses the AI5 chip, descended from the Full Self-Driving compute platform that processes camera feeds across 7 million vehicles on the road. Musk has made vertical integration a core principle: Tesla designs its own chips, trains on its own Dojo supercomputer (supplemented by NVIDIA GPUs), manufactures in its own factories, and deploys in its own operations. The FSD chip avoids the $2,999 per-unit NVIDIA module cost entirely.

At 1 million Optimus units per year — Musk's target — that saves $3 billion annually in module costs, but custom silicon R&D is not free. Apple spends an estimated $3–5 billion per year on chip design; Tesla's silicon team is smaller but still represents billions in cumulative investment. The breakeven math depends on volume: below roughly 300,000–500,000 units per year, buying Jetson Thor modules from NVIDIA is cheaper than amortizing custom silicon design across the fleet, because the fixed costs of a chip team — the RTL engineers, the verification suites, the TSMC wafer reservations, the multi-year tapeout cycles that consume hundreds of millions before a single transistor switches — don't shrink when production volume does. Above that threshold, vertical integration wins decisively.

Tesla shipped 150 Optimus robots in 2025. One hundred fifty. It is converting the Fremont assembly line that once produced the Model S and Model X — those production lines ceased in May 2026 after 14 and 11 years respectively — into an Optimus manufacturing facility targeting 1 million annual units. That is a 6,667x ramp. Tesla's own car production, from zero vehicles in 2012 to 1 million around 2020, took eight years and happened with the benefit of an existing automotive supply chain built over a century by hundreds of companies. Humanoid robots have no established supply chain, no tier-one parts catalog, and no workforce trained to assemble them.

Until Tesla crosses the 300,000-unit threshold, its vertical integration is an expense. After that threshold, it becomes a moat. The question is how many years separate those two points, and whether NVIDIA's ecosystem lock-in is irreversible by the time Tesla gets there.

OpenAI enters with the strongest AI and no hardware

Sam Altman's robotics division is the third strategy. OpenAI has, by most benchmarks, the strongest or second-strongest general AI models in the world. What it does not have: a humanoid body, a manufacturing partner, a sensor stack, a proven locomotion controller, or any track record in physical systems. Altman's post framed the effort around infrastructure work first — construction sites, skilled labor augmentation — before consumer robots.

The $75 billion that evaporated from Tesla's market cap in a single session reflects a rational repricing of exclusivity. Musk's Optimus valuation thesis (Piper Sandler prices it at $100 per share) assumed Tesla would be one of a very small number of companies capable of building general-purpose humanoids. OpenAI's entry, funded by a $40 billion round and staffed by the engineers behind GPT-5.4, makes that assumption shakier. But OpenAI will almost certainly train its robot models on NVIDIA GPUs. It may well use NVIDIA Jetson Thor modules in its prototype hardware. The new entrant that wounded Tesla's valuation simultaneously strengthened NVIDIA's platform position.

Jensen Huang does not need to pick the winner. He sells the compute substrate to every contestant except one, and if history is any guide — and in Huang's case, it is, because NVIDIA captured north of 80% of AI training GPU revenue by running this exact playbook in data centers — the physical AI market will follow the same gravitational path toward NVIDIA silicon that the software AI market already did.

The Android parallel — and its limits

NVIDIA's reference design follows the Android playbook: provide an open platform, let a thousand OEMs build on it, and capture value at the ecosystem layer rather than the device layer. Google doesn't profit from Android licensing fees — it profits from Search, Play Store, and advertising delivered through Android's 3.3 billion active devices. NVIDIA doesn't need margins on the $2,999 module. It needs the 2 million developers, the Isaac Sim training pipeline, the Omniverse simulation platform, and the DGX/HGX cluster sales that serve the training workloads those developers generate.

But the Android parallel has a sharp edge. Google won market share with openness. Apple won profits with integration. Android's global smartphone share exceeds 70%; Apple's iPhone captures roughly 85% of industry profits. If Tesla succeeds at vertical integration — builds its own chips, its own AI stack, its own manufacturing, its own deployment — it could become the Apple of humanoid robots: lower share, higher margins, deeper customer lock-in. NVIDIA's reference design may create a fragmented market of adequate robots while Tesla produces a smaller number of exceptional ones.

The counterargument is chronology, and it cuts against Tesla sharply. Apple had the iPhone for a full year before Android launched. Tesla shipped 150 robots in 2025; Unitree and AGIBOT shipped thousands. The integrated player is behind, not ahead. And the fragmented market is not standing still — Figure AI's robots sorted 250,000 packages in 200 hours with zero hardware failures, running on NVIDIA silicon. The ecosystem's best performers may close the quality gap before Tesla closes the volume gap.

What this does not prove

The training compute estimates above are order-of-magnitude, not audited figures. Actual GPU-hours per skill vary by 10–100x depending on task complexity, simulation fidelity, and algorithmic approach — a reinforcement learning policy for stair ascent in a high-fidelity physics simulator can consume twenty times the compute of a basic grasp controller trained in a simplified environment, and both figures shift further depending on whether the team uses imitation learning, model-based RL, or some hybrid nobody has published yet. Some companies train on NVIDIA hardware but supplement with custom accelerators. The $875 million to $1.75 billion range is illustrative of relative magnitude, not a forecast.

NVIDIA does not separately disclose robotics-related data center revenue, so the precise split between robotics training compute and, say, autonomous driving or industrial vision training is unknowable from public filings. The $95,000 reference design price comes from humanoid.guide, not NVIDIA's official pricing. And Unitree's IPO prospectus — which would provide the first audited financials from a pure-play humanoid robot maker — is not yet public.

We also do not know OpenAI's hardware strategy. If Altman partners with an existing robot maker running Jetson Thor, the NVIDIA tax argument strengthens. If OpenAI builds or acquires its own silicon (it has invested in chip design), the calculus changes.

The Bottom Line

Three strategies, one week. Tesla bets on vertical integration: own every layer, escape the NVIDIA dependency, and profit at million-unit scale. OpenAI bets on AI superiority: bring the world's best models to a body someone else builds. NVIDIA bets on platform infrastructure: sell the brain to everyone, capture the training ecosystem, and collect rent regardless of which robot wins.

If you're an investor, the cleanest bet is NVIDIA. It profits whether Tesla, OpenAI, Figure, or Unitree wins. But the return is capped — robot compute is a rounding error in a company already valued at $5.4 trillion. If you're a robotics engineer, the Isaac GR00T reference design is the most important product announcement of the year: a $95,000 on-ramp to frontier humanoid research that used to require millions in custom hardware integration. And if you are watching the humanoid race as a labor market signal, the relevant number is not which company ships the best robot. It is that 50,000 to 100,000 humanoid robots will ship this year regardless, and the compute infrastructure to make them useful is already in production — for $2,999 a brain.