Theses›Robotics & physical AI›Robot foundation models

Robot foundation models

Robot foundation models (VLA) are strategically central but value-capture is contested; only a proprietary self-reinforcing data loop earns durable margin

  • Bear
  • Shut
  • $21B by 2035

The callOne model will run many robot bodies, but the majors give it away to sell compute, so the money goes to whoever holds the data loop, and I don't know who that is yet.

D1D2R1R2R3C1foragainst D1D2R1R2R3C1foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the restraints.
  1. D1One model turns camera frames and a spoken instruction into motor commands across many robot bodies, which classical per-cell programming cannot follow.
  2. D2Open-source commoditisation pulls more robots onto these models, with about 11 commercial deployments by Q1 2026 in manufacturing, logistics and vehicles.
  3. R1The majors give the model away to sell compute, so a standalone model company has nothing durable; Covariant's exit to Amazon was about $400M as a licence.
  4. R2The leaders are already priced as winners, Skild at $14B and Physical Intelligence above $11B.
  5. R3The data loop is expensive: XDOF, Mecka and Tacta each raised $68-110M before revenue.
  6. C1NVIDIA releases GR00T N2 as a free download at about 2x the task success of leading models, commoditising the model layer and moving margin to the data loop.end-2026
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
202720282029today1234predictionswatch 202720282029today1234predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1GR00T N2 ships free at the previewed task success, confirming the model layer is commoditised.Dec 2026 · open
  2. 2Through end-2027 the leading deployed robot foundation / VLA models remain controlled by integrated robot-or-AI majors; no standalone merchant-VLA vendor has a durable independent commercial footprint.Dec 2027 · open
  3. 3Whether any standalone model vendor has a durable commercial footprint.Dec 2027 · open
  4. 4A data-scaling law where open data alone reaches leading-model capability, which would move value back to the model.Dec 2028 · open
Expressionslisted, private, technologies
  • GOOGLGoogle
    • cheap
  • NVDANVIDIA
    • cheap
     Dominant share of AI training compute.
  • TSLATesla
    • expensive
     Trains its own end-to-end robot and driving models in-house for the Optimus humanoid and its cars; not sold to third parties.
  • AMZNAmazon
    • fair
     Licensed Covariant's robotic foundation models and hired its founders for in-house warehouse robotics.
  • METAMeta Platforms
    • cheap
     Releases the open V-JEPA 2 video world model, which it demonstrates for zero-shot robot planning and control.
Private · 8 private companies, 6 of them in the US.

Wayve · Physical Intelligence · Skild AI · Figure AI · FieldAI · Generalist AI · Covariant · mimic

Technologies
Who has already bet on itprivate rounds, from public sources

$6.6B raised by 8 companies, $6.4B of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: Physical Intelligence, Generalist AI, FieldAI, mimic. Backed by a strategic investor: Figure AI, Skild AI, Wayve, Physical Intelligence, Generalist AI, FieldAI.

What changedthe record, newest first
  • 22 Jun 2026Robot foundation models scale on the same data laws as LLMs, and the bottleneck is training data not compute, which is why the robot-data supply stack is where to look.
Reading
Show the reading