UpTrajectory Review

Sebastian Thrun, the Stanford professor who convinced Google to bet on autonomous vehicles nearly two decades ago, has quietly launched another company—this time targeting not cars but the factory floor itself. Dulo, still in stealth mode, aims to build 'foundation models for hardware design' with the stated goal of enabling 'manufacturing at lightspeed.' The announcement came as an almost offhand coda to a keynote at San Francisco's Actuate robotics conference, with Thrun deliberately withholding details beyond the company's existence and its recruiting pedigree from Waymo, Google Brain, and Stanford's AI lab.

For small and midsize manufacturers, this matters because the ground beneath their competitive position is shifting faster than most realize. The current robotics boom—$16.3 billion in physical AI investment in Q1 2026 alone, per PitchBook—is driven by genuine economic pressure: reshoring mandates, persistent labor shortages, and finally-cheap-enough hardware. But the real disruption arrives when AI moves from controlling individual machines to designing the machines themselves, and then optimizing their production in closed loops. Thrun's track record suggests he is not building a niche tool for Tesla or Boeing. Foundation models in this context imply something more democratizing and more threatening: design capabilities that previously required teams of specialized engineers, compressed into software that smaller competitors might license—or find themselves competing against.

What is genuinely new here is the specific application of foundation-model architecture to hardware design itself, rather than to the operation of already-designed hardware. Most 'AI in manufacturing' coverage focuses on predictive maintenance, quality inspection, or robot arm programming. Dulo's framing suggests something more structurally ambitious: models trained on enough design and manufacturing data to generate or optimize physical objects the way large language models generate text. This is unproven at scale, and Thrun's history includes notable failures—Kittyhawk's flying-car collapse comes to mind—that should temper automatic enthusiasm. The stealth posture may signal genuine early-stage uncertainty rather than mere competitive caution.

The downstream effects split unevenly across the manufacturing landscape. Large enterprises with existing digital design pipelines and proprietary data hoards could integrate Dulo-like capabilities fastest, widening their advantage. Contract manufacturers and job shops—the backbone of American small-batch production—face a more complicated calculus. If design-to-production cycles compress from months to days, the value of their flexibility and customer relationships gets tested against competitors who can iterate faster with fewer people. Conversely, if Dulo or rivals sell access rather than build vertically integrated factories, some smaller operators could leapfrog capital-intensive tooling investments. The labor question is equally fraught: 'lightspeed manufacturing' promises efficiency, but for skilled machinists and design engineers, it may mean obsolescence or, at best, a forced transition from doing to supervising AI outputs.

Watch whether Dulo emerges with a platform model or a captive manufacturing play—Thrun's Google pedigree suggests the former, but hardware's capital demands often push toward vertical integration. For operators now, the actionable move is auditing your own design and production data posture: what is digitized, what is trapped in tribal knowledge, and what partnerships would let you pilot generative design tools before a Dulo-like platform sets the terms. The $16.3 billion quarterly figure is not merely venture capital exuberance; it reflects buyer demand that will reshape supplier expectations. Manufacturers who treat this as a distant concern may find the floor pricing and lead-time assumptions underlying their quotes have become obsolete while they were not looking.

Thrun's repeated pattern—academic breakthrough, corporate scaling, then entrepreneurial spinout—suggests Dulo will seek institutional validation before broad commercialization. That creates a narrow window for smaller manufacturers to engage with early pilots or academic partnerships, particularly through Stanford's SAIL network, rather than waiting for polished enterprise software. The risk of early adoption remains real; the risk of waiting, given Thrun's history of defining categories before others recognize them, may be greater.

Takeaway: Audit your design-to-production data now, before foundation-model manufacturing tools set terms you cannot negotiate.