Physical AI and Compute: Vendor Reports on Robots, Chips and GPU Markets

A robotic arm at rest beside server racks in a dark industrial hall

Six scored vendor reports on the physical side of AI: humanoid robots in the US and China, robot training data, AI chip design tools, the markets forming around GPU compute and the nuclear reactors aiming to power data centers.

AI is leaving the data center for factories and homes, and the supply chain behind it is being rebuilt. This hub groups our scored vendor reports on robots, the data and chips they need, and the market for compute.

What holds physical AI back?

Software AI scaled on text that already existed. Robots have no such archive, so data has to be recorded task by task. The models then need GPU capacity that is expensive and bought years ahead, and the chips themselves take years to design.

The reports below cover each constraint, with separate reports on Chinese humanoid makers and on the advanced nuclear reactors aiming to power data centers. Each scores vendors on four criteria using what they publish.

Where does the humanoid robot market stand?

Two humanoids have a public price. Unitree lists the G1 from $13,500, and 1X takes orders for its home robot NEO at $20,000 or $499 a month. The industrial makers sell through pilots instead. Agility says Digit has 65,000 hours of real production experience, and Figure says it has built more than 350 Figure 03 robots.

Published specifications are still modest. Figure 03 is rated for a 20 kg payload, NEO carries 25 kg and Atlas sustains 30 kg, and battery life runs four to five hours. That limits them to light handling for now. Ask any vendor for hours worked per robot per week on a customer site, not the lab demo.

Why is robot training data scarce?

Language models trained on text the internet had already produced. Robots learn from recorded demonstrations, and no archive of that size exists. A group of vendors now collects the data for robotics labs. The main types are egocentric video of people doing tasks, teleoperation of real robots, demonstrations with handheld devices, simulation and logs from robots already deployed.

Each type trades cost against realism. Egocentric video is cheap and plentiful but shows a human body, not a robot's. Teleoperation is the closest match to what a robot will do and the most expensive to collect. PrismaX is the only vendor in our report with self-serve plans and a data console. Mecka, Scale AI and Parallax Worlds sell through a conversation.

What is AI changing in chip design?

Chip design is slow because verification is slow, so the new tools aim first at writing and checking RTL and testbenches. Three approaches compete. Incumbents such as Synopsys add agents to the EDA tools teams already use. Startups such as ChipAgents and Cognichip build agents that sit on top of those tools. Labs such as Ricursive Intelligence train foundation models for chip design. None publishes a price, and Ricursive is the only one with an open-weight model you can try.

How are GPU hours turning into a market?

GPU hours are starting to trade like a commodity. Several companies now publish daily rental price indexes for H100, H200, B200 and B300 chips. Regulated trading is early: Architect describes its exchange as a CFTC-regulated designated contract market, and Liquid Compute says its exchange is pending CFTC designation. For most buyers the practical use today is the data, because a public index gives a reference price when negotiating a GPU contract.

The report scores six platforms on four things: the quality of their price data, whether a buyer can hedge, their regulatory status and whether a trade ends in real GPU capacity being delivered. Ornn, Compute Desk and Silicon Data score highest on data. Ornn, Liquid Compute and SF Compute score highest on delivery, and Architect leads on hedging and regulation.

The reports in this cluster

How do the four constraints connect?

Each constraint sets the pace for the next. More robot data means larger robot models, and larger models need more training compute. Compute supply depends on chips, and chip supply depends on design cycles that AI tools are trying to shorten. A price shock in GPUs changes what a robotics lab can afford to train. A breakthrough in data collection raises demand for compute. Anyone planning a physical AI project should track all four, even when buying only one.

What should you ask physical AI vendors?

Most numbers in this market come from the vendors themselves, so the questions matter more than the brochure. These five separate a working product from a demo.

  • Deployment hours. How many hours has the product run on customer sites, and how many on the largest single site?
  • Intervention rate. How often does a person have to step in, take over remotely or reset the robot during a shift?
  • Total cost. What does one robot, one dataset or one GPU hour cost once support, integration and delivery are included?
  • Data rights. Who owns the data recorded on your site, and can the vendor use it to train models it sells to others?
  • Delivery dates. For hardware and compute, what is the committed delivery date, and what happens if it slips?

Which report should you read first?

  • Operations and manufacturing leaders weighing a robot pilot: start with the humanoid robots report, then the data report to see what training a new task involves.
  • Robotics and AI research teams: start with the robotics data report, then GPU compute markets for training cost.
  • Semiconductor teams: the chip design report compares the incumbent and startup approaches to verification.
  • Infrastructure and finance buyers: GPU compute markets covers price indexes, exchanges and financing, and pairs with our neocloud and FinOps reports below.

Physical AI changes faster than most software categories, and much of the evidence is still vendor claims. We update each report when vendors publish new prices, specifications or deployment numbers.

How to use these reports

  1. Start from your problem. Pick the report closest to the work you want done, not the vendor you have heard of.
  2. Read the criteria first. Each report scores four things from 1 to 5. Check they are the four you care about.
  3. Treat scores as a shortlist. They reflect what vendors publish. Run your own trial before you buy.
  4. Check the date. Prices and plans change monthly. Each report says when its sources were read.

What is physical AI?

Physical AI is AI that senses and acts in the real world, mainly robots and autonomous machines. It also covers the supply chain behind them: the data they learn from, the chips they run on and the compute used to train them.

Why group robots with chips and compute?

They constrain each other. Robot models need large amounts of compute and recorded data to train, compute depends on chip supply, and chip design is itself being sped up by AI. A buyer or investor looking at one usually needs to understand the others.

Can a company buy a humanoid robot today?

Yes, in two ways. Unitree lists the G1 from $13,500 and 1X takes orders for NEO at $20,000 or $499 a month. Industrial humanoids such as Agility Digit, Figure 03, Boston Dynamics Atlas and Apptronik Apollo are sold through paid pilots with a quote, as of November 2026.

What is a GPU price index used for?

A GPU price index publishes a daily reference rate for renting a given chip, such as an H100 or B200, for an hour. Buyers use it to check a quote when negotiating a compute contract. As regulated GPU futures appear, the same indexes can settle contracts that let a buyer fix a future price.

Explore More

Ready to Find the Right AI Tools?

Browse our data-driven rankings to find the best AI tools for your team.