Our Forecast: Humanoid Robots Will Work for Hire in Two Markets by 2028

By the end of 2028, we expect humanoid robots to do paid daily work at real scale in two separate markets. One runs on cheap Chinese bodies. The other runs on leased US fleets priced against wages and placed mainly where workers cannot bargain over them.

Key points

• We forecast two price regimes by 2028: Chinese hardware under $10k sold as equipment, and US robots-as-a-service priced just under a human wage.
• The first US employer to run more than 1,000 humanoids will most likely be a non-union site on a per-hour contract (about 65%).
• Robots will usually work fewer hours than their pricing assumes, so vendors will close the gap with hidden teleoperation and penalties for shortfalls.
• Homes stay marginal: home robots arrive remotely piloted, and privacy limits them more than price does.
• The most likely thing to break this call is a serious injury in a shared workspace before ISO 25785-1 is published.

Our forecast

By December 2028, at least one US employer will run more than 1,000 humanoids in paid daily work at a non-union site, billed per robot-hour. By then no unionized US or Korean plant will run more than 200 under a signed agreement, and no Western maker will list a full-size humanoid below $20,000. We put this at about 65%.

Why we think so

The signal scan found the cost curve to be the strongest force. Unitree sells full-size bodies for $4,290 to $10,500, about 5 to 10 times below Western industrial models¹. China shipped about 97% of the world's humanoids in the first half of 2026 and expects output above 100,000 this year, with two firms taking about 80%²³. The FCC ban on newly authorized foreign-made humanoids keeps those bodies out of the US⁴. Our convergences piece showed this is the same bloc split already seen in AI compute. Behind that wall, a price pegged to wages can survive.

Our implications piece traced what follows from that pricing. Agility models about $8,500 a month per Digit against a burdened rate of $30.50 an hour and reports more than $300M in multi-year orders, roughly 1,000 robots⁵. That order book is enough for one large customer to pass our threshold. To beat a wage, though, a robot has to work about 3,300 hours a year. Figure's BMW run logged 1,250 hours in 11 months⁶. Vendors will close the gap with teleoperation, which is already the main source of training data⁷, and with contract clauses rather than with autonomy alone.

Union status decides where the fleets go. Hyundai's Korean union bars Atlas without an agreement and has already walked out over job guarantees, while Hyundai's first Atlas plant is in Georgia⁸. Tesla is converting a line at non-union Fremont⁹. The contradictions piece left liability open, and that gap favors employers whose workers have the least say over the risk. Cross-embodiment models such as GR00T and Qwen-VLA make any body usable¹⁰¹¹. That helps cheap hardware outside the US and helps the platforms that license the models everywhere.

How this call fails

It is 2029 and we were wrong. The most likely reason is a serious injury. A cage-free robot falls on a worker¹² while ISO 25785-1 is still in draft, and a state moratorium or insurers' exclusions freeze shared-space deployment. A second possibility is that autonomy stalls: Tesla's ramp keeps slipping⁹, teleoperation turns out to be most of the work, and the hourly economics never close at scale. Less likely, a trade deal reopens the US to Chinese bodies, wage-pegged pricing collapses, and paid work spreads as cheap equipment rather than as a service.

Who loses

• Warehouse and line workers at non-union US sites, who absorb displacement first and carry the safety risk while the standards remain unsettled.
• Western hardware makers that own no model layer, squeezed between Chinese bodies and the platforms that license the robot brains¹¹⁰.
• Offshore teleoperators, paid low wages to produce the data that removes their own role⁷.
• Early buyers of home robots, who pay a subscription for a remotely piloted device with camera access that no one has audited¹³.

First moves

• Operators: write robot-hour contracts with audited floors for hours worked and mandatory disclosure of remote-assisted hours.
• Unions outside Korea: put humanoid deployment on the bargaining agenda now, using the Hyundai language on written job guarantees as the template⁸.
• Insurers and risk teams: price shared-space cover per hour worked near people, and require fail-passive testing before going cage-free¹².
• European makers: compete on certification ahead of the Machinery Regulation in January 2027 and the 2028 high-risk AI deadline¹⁴.
• Investors: weight the model and licensing layer over Western hardware.

What would change our mind

• A unionized plant signs an agreement allowing a fleet of more than 500 humanoids before any non-union site reaches 1,000.
• Published fleet data shows robots averaging well over 3,000 hours a year with low remote assistance, which would mean autonomy, not labor arbitrage, is driving adoption.
• The FCC rule is reversed or waived, or a Western maker lists a full-size humanoid below $20,000.
• A serious injury or a privacy breach in a home triggers a federal or state pause on deployment in shared spaces.