Everything About AI
NextFin News - Marc Raibert, the founder of Boston Dynamics, has spent three decades arguing that the hard part of robotics is not the demo — it is the deployment. His appearance at the Barclays Asia Forum in Singapore on Oct. 22, 2026, where he discussed the development of intelligent robotics technology, lands at the moment that thesis is being stress-tested: a handful of humanoid platforms are finally leaving the lab, while the capital committed to them has outrun the number of machines actually at work.
The tension defines the sector. As of May 2026, approximately 7,000 to 8,000 commercial humanoid robots were operational worldwide — led by Unitree's cumulative shipments of more than 5,500 units, Tesla's internal Optimus deployments of over 1,000, and roughly 250 Agility Digit units. Yet Figure AI closed a Series C round in September 2025 at a $39 billion post-money valuation, and in September 2026 shares of Unitree Robotics fell about 53% from their IPO debut, erasing roughly $35 billion from the peak. Raibert's long-stated position — that robots will earn their keep in controlled industrial settings long before they enter homes — is a reminder that the physics of deployment, not the charisma of demonstrations, will decide who survives.
The Deployment Gap: What the Numbers Actually Show
The first question any physical-AI investor should ask is simple: how many machines are doing real work? The answer is far smaller than the market narrative implies, and the distribution is uneven.
The global installed base of commercial humanoid robots stood at roughly 7,000 to 8,000 units as of May 2026. Unitree accounts for the bulk, with more than 5,500 G1 and H1 units shipped to date, mostly to research labs and Chinese automotive OEMs. Tesla has deployed more than 1,000 Optimus Gen 3 units across its Fremont, Austin, and Berlin factories. Agility Robotics has placed about 250 Digit units in logistics deployments, including Amazon and GXO Logistics sites. Figure, Apptronik, 1X, Sanctuary, and Boston Dynamics' own Atlas sit between roughly 40 and 150 units each — pilot scale, not production scale. These are the machines that define the installed base; everything else is a prototype or a press release.
Against that base, the capital is enormous. Figure AI exceeded $1 billion in committed capital in its September 2025 Series C at a $39 billion post-money valuation, with total funding of approximately $1.9 billion. Tesla has stated capacity targets of roughly 100,000 Optimus units in 2026 and one million per year by 2028. Goldman Sachs Research projects the market for humanoid robots could reach $38 billion by 2035, up more than sixfold from an earlier $6 billion estimate, with shipments of 1.4 million units over the same period. Morgan Stanley sees the addressable market — including supply chains, repair, maintenance, and support — surpassing $5 trillion by 2050, with more than one billion humanoids deployed. The gap between a 7,000-robot installed base and a $39 billion private valuation is not a rounding error. It is the central risk of the trade.
Price is the second hinge, and it is where the story splits. Unitree's G1 starts around $16,000, roughly an order of magnitude below full-size Western platforms. The Unitree H1 sits near $90,000. Commercial-grade systems from Figure, Apptronik, and Sanctuary run $200,000 or more per unit, with Sanctuary's Phoenix estimated near $250,000. Tesla targets $20,000 to $30,000 for Optimus Gen 3 but is internal-only in 2026. At those prices, the economics only work where labor is expensive, the task is repetitive, and the environment is controlled. That describes the factory. It does not describe the home.
Raibert has made this sequencing explicit in public remarks outside the Singapore forum.
"It'll be in factories first, I don't think it'll be in homes for a while," Raibert said in a profile, adding that the machines will not all resemble the human body — some may have four legs, one big arm, or wheels.
The implication for investors is uncomfortable: the humanoid form factor that dominates headlines may not be the form factor that dominates revenue. A "mixture-of-robots" fleet — quadrupeds for inspection, arms for manipulation, wheeled platforms for transport — is more likely than an army of bipeds.
Why the Industry Is Consolidating Now
The second layer of the story is ownership. Physical AI is entering its consolidation phase, and the deal flow shows which players are positioning for a long, capital-intensive buildout rather than a near-term consumer breakout.
In September 2026, SoftBank Group agreed to acquire the Robotics and AI Institute (RAI) from Hyundai Motor Group. Financial terms were not disclosed, and the transaction is under review by the Committee on Foreign Investment in the United States. RAI is a research engine, not a robot company: it works on dexterous manipulation, learning-based control, physical interaction models, and navigation, with a research horizon measured in years rather than product quarters. Hyundai spun the institute out in 2022 with an initial investment of more than $400 million, and that initial funding has now ended. SoftBank is buying an option on the underlying science, not a product roadmap.
The same month, Hyundai said it was pursuing the acquisition of SoftBank's remaining 9.65% stake in Boston Dynamics after SoftBank exercised a put option. Taken together, the two moves redraw the map: Hyundai is concentrating ownership of the robot maker itself, while SoftBank reaches upstream into the research layer. Both are betting that value in physical AI will compound through patents, data, and integration know-how rather than through a consumer launch. It is worth noting that SoftBank owned Boston Dynamics from 2017 to 2021, selling a controlling stake to Hyundai in a 2021 deal that valued the company at $1.1 billion; the circle is closing with SoftBank re-entering the ecosystem from a different angle.
Boston Dynamics is also changing leadership. On Oct. 7, 2026, the company appointed Rohit Prasad as chief executive officer, effective immediately. Prasad spent 12 years at Amazon, most recently as senior vice president and head scientist for Alexa and artificial general intelligence, where he helped scale a consumer AI service used in hundreds of millions of households. The appointment is a signal that the next phase of robotics is less about mechanics and more about language models, task planning, and the ability to generalize from instruction. The company said the move "accelerates Boston Dynamics' Physical AI strategy by combining world-class robotics with advanced AI capabilities to commercialize intelligent machines at scale."
The industrial beachhead is already being built. In September 2026, Boston Dynamics opened its Robotics Metaplant Application Center at Hyundai Motor Group Metaplant America in Georgia — a testbed for training Atlas humanoids directly inside automotive manufacturing operations. Robert Playter, the former chief executive, said at the Global HR Forum in October 2026 that he expects humanoid robots to begin sorting automotive parts within two years, freeing workers to focus on problem-solving and repairs. That is a concrete, near-term use case with a named customer and a defined task. It is exactly the kind of deployment Raibert's "factories first" framing predicts.
The Second-Order Question: Who Owns the Stack?
The first-order read of this news is straightforward: robots are coming to factories, and the hype is ahead of the hardware. The second-order question is harder, and it is where the money will actually be made: which layer of the stack captures the margin?
There are four layers. The hardware layer — actuators, gearboxes, batteries, sensors — is a margin trap unless a player achieves scale, and Unitree's $16,000 price point suggests Chinese manufacturers are already pushing the chassis toward commoditization. The platform layer — robot operating systems, simulation, and developer tools — is where Qualcomm is moving, having agreed to acquire PickNik Robotics and keep the MoveIt open-source manipulation framework, and where Google's Intrinsic released open-source capabilities in September 2026. The model layer — robot foundation models and task planners — is where Prasad's background matters, and where Nvidia has expressed readiness to invest in Korea's robotics and physical-AI segments. The integration layer — knowing which robot, in which cell, doing which task, for which customer — is where Boston Dynamics, Figure, and Agility are fighting.
The counter-intuitive point is that the company that ships the robot may not own the profit pool. If the model layer generalizes, hardware becomes a low-margin commodity and integration becomes a services business with modest multiples. If the model layer fragments, the platform owners with developer ecosystems capture the rent. Raibert's emphasis on the customer understanding the technology points to integration as the near-term bottleneck.
"One of the things people underestimate is the need for the customer to understand the robotics technology," Raibert said at a recent industry event. "When we buy cars, we're all very experienced users of cars. But robots are not like cars yet."
His conclusion: "You need an educated user base, and a culture where they know how the robot is going to be used."
That framing carries a direct valuation implication. A company that ships robots but does not own the workflow knowledge — the task library, the failure modes, the customer-specific integration playbooks — is selling boxes. A company that owns the workflow owns the recurring revenue. This is the difference between a hardware multiple and a software multiple, and it is the difference the market has not yet priced cleanly across the sector. Agility Robotics, for instance, reported $1.8 million in revenue with Digit deployed at nine customer sites ahead of its proposed SPAC merger — a toehold, not a scale business, but a disclosure of the kind the sector rarely provides.
The Counter-Thesis: Why the Hype Might Be Rational
The strongest case against the cautious read is that the market is not pricing robots as they are today; it is pricing the cost curve. Exponential technologies look absurdly overvalued at the start and absurdly cheap at the end. The transistor, the smartphone, and the cloud all followed that path. If humanoid robots follow it, today's $39 billion valuations will look like the entry fee, not the top.
The evidence for the optimists is real. Tesla is vertically integrated in a way no pure-play robotics company can match — it makes its own actuators, batteries, and chips, and it has its own factories as first customers. Unitree is already shipping at $16,000, an order of magnitude below Western competitors, which suggests the cost curve can bend faster than incumbents expect. And the model layer is improving on a steep trajectory: every quarter, robots demonstrate manipulation skills that were impossible the quarter before. If generalization arrives, the installed base does not need to be large for the option value to be enormous.
The optimists also have a point about first-mover advantage in data. Every robot in a factory generates task-specific data that improves the model, which makes the next robot more capable, which attracts more customers — a flywheel that rewards the company that ships first, even at thin margins. Figure operated Figure 03 robots for eleven months at the BMW Spartanburg plant, completing more than 30,000 vehicle production cycles, before announcing a Leipzig deployment for summer 2026. That is the kind of real-world data asset that cannot be replicated in simulation.
The rebuttal is that flywheels only spin if the unit economics work. A robot that costs $200,000 and requires constant human supervision is not a flywheel; it is a consulting engagement with legs. The data flywheel requires machines that are deployed, utilized, and retained. Unitree's 53% post-IPO decline is the market's first serious vote on whether the utilization story holds. Until utilization rates and gross margins are disclosed at scale, the flywheel remains a hypothesis, not a mechanism.
The falsifying signal for the cautious thesis is specific: if a major platform — Tesla, Figure, or Unitree — reports sustained gross margins above 30% on robot hardware at volumes above 1,000 units per quarter, combined with demonstrated unsupervised operation exceeding 90% of a shift, the "factories first, profits later" view is wrong. That combination would prove the cost curve and the autonomy curve have crossed simultaneously. Nothing in the public record shows that today.
Outlook: Three Horizons, Three Trades
Short term (6–12 months): sentiment and liquidity. The sector will remain volatile. Unitree's post-IPO drawdown set the tone: public markets are willing to fund the story, but not at any price. The CFIUS review of the SoftBank-RAI transaction and the expected closing of SoftBank's roughly $5.4 billion acquisition of ABB's robotics business are the near-term catalysts. Expect headline-driven moves, not fundamentals-driven ones.
Medium term (1–3 years): fundamentals and deployments. This is where the thesis gets tested. The metrics that matter: robots in sustained commercial operation, utilization rates, gross margins, and customer retention. Boston Dynamics' RMAC in Georgia and Figure's BMW deployment are the watchlist. If Atlas is sorting automotive parts within two years, as Playter expects, the industrial thesis is confirmed. If deployments stall in pilot purgatory, valuations compress further.
Long term (5+ years): structural. The structural call is clear: physical AI is a regime shift, not a cycle. The convergence of foundation models with affordable actuators and sensors is permanent, and it will not revert on its own. The question is not whether robots matter in 2035 — they will. The question is which companies capture the value, and the answer depends on stack ownership, not demo quality. The winners will own the model layer, the platform layer, or deep vertical integration — not the ones with the most viral videos.
Base case: humanoid robots reach tens of thousands of units in industrial settings by 2030, with the market valued in the tens of billions — broadly consistent with Goldman Sachs Research's $38 billion 2035 projection. Upside case: generalization arrives earlier than expected, cost curves bend to $20,000 per unit at scale, and the market re-rates toward Morgan Stanley's $5 trillion 2050 view. Downside case: autonomy plateaus, utilization disappoints, and the sector consolidates into a handful of survivors trading at hardware multiples.
The physical-AI trade is not a bet on whether robots will work. It is a bet on who gets paid while the rest of the world finds out.
Explore more exclusive insights at nextfin.ai.
