Physical AI & Robotics

Separate the two markets. Industrial robotics is a mature deployment story: 542,000 robots installed worldwide in 2024, more than half of them in China, whose domestic manufacturers now hold a majority of their home market for the first time. Humanoids are not that story yet. Unitree shipped roughly 5,500 units in 2025, more than every Western maker combined, and most went to laboratories and showcases, not production lines. Figure's fleet at BMW Spartanburg numbers in the tens, reportedly billed by the robot-hour; Tesla concedes Optimus remains, in its own words, primarily for learning. The honest reading of 2026: humanoid volume is arriving, led from China, but productive work per unit is the unproven variable. Procure accordingly.

Foundation models changed what is possible; they have not changed what is provable. NVIDIA's Isaac GR00T line, Figure's Helix and Amazon's DeepFleet mark a genuine architectural shift: robots now learn behaviours from data rather than executing programmed motion. The synthetic-data flywheel is real: NVIDIA reports generating 780,000 training trajectories, nine months of human demonstration, in eleven hours.

But a vision-language-action policy is a statistical artefact commanding actuators that exert real force near real people. When it fails, there is no stack trace. This is exactly what our founder coined "Complex AI" for in 2019: a statistical policy exerting real force inside a high-stakes system, where the audit trail has to exist before the incident. A warehouse forgives a mis-picked tote. A gas compression plant does not.

Certification is now the binding constraint, and the deadlines are set. From 20 January 2027, the EU Machinery Regulation treats machine-learning safety components as high-risk, requiring third-party conformity assessment. ISO 25785-1, the first safety standard for dynamically stable legged robots, reached committee-draft stage in May 2026. Neither yet answers the hard question: how do you certify a policy that cannot enumerate its own behaviours? Our position is unfashionable but practical: architect the autonomy stack so the learned component never holds the safety case alone. That means deterministic envelopes, measured intervention rates and auditable logs. The firms that treat certification as an engineering discipline, not paperwork, will own regulated markets while their rivals stay stuck at pilot.

For the Gulf, the opportunity is sharper than buying humanoids. The region's real advantages are hazardous-environment use cases with unarguable economics, and the freedom to write certification regimes rather than inherit them. ADNOC's heavy-duty inspection robot at the Taweelah gas plant, and the ARGOS operator robot due by end-2026, are the right pattern: bounded tasks, hostile conditions, measurable risk removed from people. Saudi ambitions, including a stated 10,000-robot deployment programme and NEOM's automated rebar assembly, will succeed or stall on the same discipline. Our advice to ministries is consistent: contract against measured availability, task completion and human-intervention rates, never against demonstration footage; and build sovereign certification capability now, while the standards are still wet.