EN
Back to the archive

The archive · Hardware & Devices · Strategic decision · 2026

Dulo: Thrun's stealth bet that AI designs hardware at lightspeed

Sebastian Thrun's stealth startup Dulo bets foundation models can accelerate hardware design and manufacturing — 'at lightspeed.'

Dulo

The betThat foundation models can compress hardware design and manufacturing — 'lightspeed' — making Dulo a model company for the physical world, not another robot builder.Building

What the business is

Dulo is a stealth robotics startup led by Sebastian Thrun, building foundation models for hardware design — AI systems meant to dramatically accelerate how machines and their parts get designed and manufactured.

How it started

Thrun's Stanford team won the 2005 DARPA Grand Challenge, he went on to lead Google's self-driving project (later Waymo), co-found Google Brain, Google X and Udacity, and run the flying-car startup Kitty Hawk. At the end of his Actuate keynote on 2026-08-18 he revealed Dulo: 'I am not speaking about the company yet. It's under stealth, it's very small. But it's in robotics.' The only public artifact is a Stanford-hosted page: 'Foundation models for hardware design' aimed at 'manufacturing at lightspeed,' built by leaders from Waymo, Google Brain and SAIL.

No ending yet — it is still running.

Background

Dulo is a stealth robotics startup revealed by Sebastian Thrun — the engineer behind Google's self-driving project that became Waymo — on 2026-08-18, at the end of his keynote at the Actuate robotics conference in San Francisco. 'I am not speaking about the company yet,' he told the audience. 'It's under stealth, it's very small. But it's in robotics.'

The only public detail is a Stanford-hosted page: Dulo is developing 'foundation models for hardware design' with the goal of 'manufacturing at lightspeed,' built by leaders from Waymo, Google Brain and Stanford's AI Lab (SAIL), which Thrun once directed. The positioning sits a step earlier than most physical-AI startups: instead of building robots, Dulo would sell AI to the people who design and manufacture machines.

Thrun's track record explains the attention: his Stanford team won the 2005 DARPA Grand Challenge, which led to him leading Google's self-driving car project, and he co-founded Google Brain, Google X and Udacity before running the flying-car startup Kitty Hawk. The reveal came amid a physical-AI funding boom — PitchBook data cited by Business Insider put Q1 2026 physical-AI investment at a record $16.3 billion across 492 deals.

As of September 2026, Dulo has disclosed no funding figure, product, customers or timeline. The bet is that foundation models — the same approach that reshaped text, images and code — can finally speed up the slow world of physical hardware design and manufacturing.

What has to be true

  • The thesis targets a real asymmetry: software iteration accelerated exponentially while hardware design and manufacturing stayed slow, and Dulo goes at that bottleneck directly.
  • Thrun's pedigree (DARPA 2005, Waymo, Google Brain, Udacity) made a two-line announcement national tech news and opened an unusually easy path to capital.
  • The company positions itself in the design-and-manufacturing layer instead of competing with humanoid and warehouse-robot makers, a differentiated lane in a crowded boom.
  • The timing bet is explicit: launch during a record physical-AI funding cycle, when reshoring and labor shortages push manufacturers to automate.

What can be applied

Fame raises attention with nothing disclosed, but attention without product or timeline is short-lived; the bet stands or falls on compressing hardware's design loop.

Aftermath

As of 2026-09-02 Dulo remains in stealth with no public funding figure, product, timeline or customers. Its Stanford-hosted page still describes 'foundation models for hardware design' aimed at 'manufacturing at lightspeed,' with a team of leaders from Waymo, Google Brain and SAIL. Thrun has said only that the company is very small. The launch itself drew immediate press — Business Insider exclusive, plus coverage from The Next Web, Benzinga, Dataconomy and Digital Today — but the case will be judged on whether the model layer actually compresses the hardware design-and-manufacturing loop.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases