Physical AI, Chapter 9: Industries Transformed by Physical AI

This is Part 9 of a series walking through my book Physical AI: Engineering Embodied Intelligence for the Real World. In the previous chapter, we covered what it takes to deploy safely. This one asks who has actually done it.

The most useful way to evaluate Physical AI in 2026 is not by counting demos but by tracking which industries are paying for the technology and what they are paying for it to do. The distribution is uneven and it is not random. Warehouse logistics, surgical robotics, restricted-condition autonomous trucking and precision agriculture have moved to production scale with paying customers and repeatable economics. Household robotics, unrestricted urban robotaxis and general-purpose humanoid labour remain partially demonstrated and substantially unproven.

The chapter goes segment by segment, deliberately concrete — named companies, named products, named deployment sites, published economics where they exist. Diligent Robotics’ Moxi fetching supplies and delivering specimens across more than two dozen US hospital systems, taking back on the order of one to four hours of nursing time per nurse per shift. Symbotic past a billion dollars of revenue on the back of the Walmart partnership. Boston Dynamics’ Stretch unloading trailers for DHL, Maersk and FedEx — a clearer commercial path than Atlas, which is still a research platform. John Deere’s See & Spray cutting herbicide use 50–80% by spraying individual weeds rather than broadcasting. Carbon Robotics killing weeds with a fibre laser and no chemistry at all. Intuitive Surgical past 9,000 da Vinci installations and 14 million cumulative procedures. And, on the other side of the ledger, Cruise: suspended in October 2023, wound down as a robotaxi business by GM in December 2024.

Five more patterns worth carrying away

The second is that the segments scaling fastest do not require full autonomy to be useful. Locus pairs robots with human pickers. Stretch does trailer unloading while humans handle the rest. Moxi runs errands while nurses do clinical care. Tesla’s FSD operates with a supervising driver. Partial autonomy delivers most of the economic value, and an architecture that only pays off at full autonomy is at a structural disadvantage — because the deployment cannot begin, and the data flywheel cannot start turning.

The third is customer concentration: a handful of large warehouse operators, hospital systems with disciplined procurement, equipment dealer channels in agriculture. The fourth is regulatory maturity, where surgical robotics has the slowest but most durable trajectory and consumer humanoids face the most unsettled environment. The fifth is that scaled segments all have a small number of clearly quantifiable benefits — labour hours, injury rates, herbicide volume, throughput — that let a customer build a credible business case. And the sixth, which new entrants consistently underrate, is ecosystem: integrators, training programmes, dealer networks. A technology dropped into a mature ecosystem scales through existing channels; one that has to build its own faces a harder path than the technology alone implies.

The chapter also works through the AV 2.0 transition in detail — Wayve’s GAIA world models and LINGO language-grounded policies, Tesla’s end-to-end V12 architecture, Waymo’s sixth-generation Driver, Mobileye’s tier-one position — and finds the end-to-end-versus-modular debate has resolved less as a winner than as a division of labour: planning benefits most from end-to-end learning, perception still benefits from specialised architectures and sensor diversity.

Tomorrow, the last one: the road to embodied AGI — the convergence moment, what the market projections are worth, the geopolitics of capability, and the frontier problems that are genuinely unsolved.