If you’re working in physical AI today, a new report out from the Bank of America Institute deserves your full attention. The analysis, issued this month under the title Physical AI, part 2: Humanoid robots, contends that the robotics industry has concluded its long era of incremental progress and is at a genuine inflection point — one driven by the convergence of generative AI and rapidly declining hardware costs.
The headline number is impressive—BofA predicts a global “population” of three billion humanoid robots in operation by 2060, which would eventually put humanoids ahead of cars on a per-capita basis. The growth curve isn’t linear, either: annual shipments are expected to jump from about 20,000 units in 2025 to 10 million by 2035, with 2026 alone expected to bring about 90,000 sales as early manufacturers begin to ramp up their first commercial-grade fleets. Industrial and automotive use cases will dominate the early years – tapping into existing EV supply chains and manufacturing know-how. But the report expects household assistants to eventually take over the market, accounting for 62% of the total robot population by 2060.
The Mind, Body & Soul Paradigm:
A good mental model in the report is BofA’s three layers of what makes a modern humanoid work: the *Mind* (the AI system that handles cognition, planning, and interaction, powered by LLM breakthroughs), the *Body* (structural hardware, energy systems and sensing), and the *Soul* (motion control — the layer that handles balance and physical execution of movement). Combined, these layers enable “embodied intelligence” – robots that learn from the world, instead of simply executing pre-written scripts. Early versions of this are already happening, from Physical Intelligence’s memory systems that allow robots to learn from past mistakes, to Sharpa Robotics’ tactile, vision-based hands that recently demonstrated apple-peeling dexterity.
Why the Economics Work Finally:
BofA lists five accelerants: AI maturity, hardware economics, industrial overlap with EV supply chains, demographic pressure from labor shortages and improving deployment readiness. The clearest signal is the cost curve. Pilot-stage humanoids now cost $90,000-$100,000 to produce, but BofA expects standardized specs to bring that down to around $50,000 by 2030. In China, where companies like Unitree are ramping toward 20,000 units a year, bill-of-materials costs are expected to fall below $17,000 — still high compared with consumer kits like the $1,445 Noetix Bumi, but a big cut for full-spec industrial machines. In particular, actuators (linear and rotary combined) account for more than half of projected hardware costs in 2030, with dexterous hands close behind at 19%—a strong signal for where component innovation and cost-cutting will matter most.
The Cold Hard Truth:
The report is bullish on the long-term arc, but candid that the industry is still in a difficult “Phase One.” Today’s humanoids remain too fragile for continuous industrial work, and there is a real gap between the impressive lab demo and an actually productive worker. Companies are approaching this from various angles: Boston Dynamics’ production Atlas favors utility over human-like form with 360-degree joints, Xiaomi is validating narrow, specific tasks like nut installation at 90% success rates rather than pursuing general-purpose labor, and startups like Sunday Robotics — fresh off a $165 million raise — are explicitly pivoting from viral demos toward a shippable, $10,000 consumer appliance.
But the biggest long-term constraint in the report isn’t compute or capital — it’s data. BofA calls out this industry’s core “data bottleneck”: gathering enough high-quality manipulation data to duplicate human-level improvisation. The solution, with either skill-capture gloves or shared-autonomy systems, is emerging as the defining race in physical AI over the next decade.
The takeaway for founders:
It’s no longer about the long-term market size — the report treats that as a given. The real opportunity is in the narrow unglamorous problems of reliability, data collection and cost-down manufacturing that will decide who really survives Phase One.
You can find the original Humanoids Daily article here.
