Part IV: Forward Deployed Robotics

Figure's 02 supported production of more than 30,000 BMW X3s at Spartanburg over about ten months (90,000 components moved, 1.2 million steps, roughly 1,250 operating hours on ten-hour shifts), and Figure 03 has moved on to logistics sequencing there, picking from unsorted containers into just-in-sequence trolleys. BMW discloses no robot count and no commercial terms. Agility's Digit runs the first formal robots-as-a-service humanoid contract at GXO's Flowery Branch warehouse (pilot, then contract, then a 100,000-tote milestone), with 65,000-plus real-world hours across customers including Schaeffler's Cheraw plant, live since early 2025, and Toyota's Canadian manufacturing operation. Agility has signed a definitive agreement to go public through Churchill Capital XI at about $2.5 billion; it had not closed as of this writing. Apptronik's Apollo works intralogistics at Mercedes' Berlin-Marienfelde digital factory and in Kecskemét, Hungary, with Jabil contracted toward commercial quantities in 2027. 1X's NEO runs the model in its most literal form inside homes: autonomous on roughly 90% of door tasks and 80% of fetching, escalating everything harder to remote "Expert Mode" operators, at $20,000 or $499 a month. And in China, UBTech's Walker series has booked roughly ¥800 million in orders, rolled its thousandth unit off the Liuzhou line, and deployed into BYD, Geely, FAW-Volkswagen, Foxconn, SF Express, and Airbus, targeting five thousand units a year of capacity in 2026.

These deployments share one operating model: put robots into customer sites with a human ready to take over by remote control, keep an engineer close, charge for outcomes, and run depots where fleets are received, QA'd, rehearsed against replicas of the customer's task, and staged. The teleop fallback keeps the robot billable before autonomy fully arrives. Each deployment also produces training data from real edge cases and operating proof for the next customer. The engineer travels with the robot, and what breaks on site becomes next quarter's roadmap. Tooling built to survive deployments becomes the product.

$5,975Formic's flat month, zero capexMarket maturity can be measured by two missing disclosures. No humanoid operator anywhere publishes a customer-facing price; Agility's cited $10–12 an hour is its own operating cost from trade press, and the closest public anchor is Formic's conventional-automation service at a flat $5,975 a month, zero capex, uptime guaranteed. Robots-in-field counts per site are also almost never published. Every deployment is still a negotiated first.

Every deployment above is run by the robot's own maker or hosted by a logistics giant. Nobody yet operates fleets as a neutral party across manufacturers.

$1.5Mrevenue per employee at PalantirTwo public P&Ls bound what that compounding is worth. The floor is the staffing industry itself: Randstad, Adecco, and Manpower averaged roughly 20% gross and 3% operating margin over the past decade, and the current numbers are worse; ManpowerGroup ran 0.6% operating in early 2026. Randstad's €23.1 billion of revenue across 38,000 corporate staff works out to about €600,000 per employee. A century of scale, and thin all the way up, because a staffing firm accumulates no proprietary record from the work it places. The ceiling is Palantir: forward-deployed engineers whose field work exists to generate the product, at roughly $1.5 million of revenue per employee across 4,429 people. The gap between €600k and $1.5M per head prices the value of turning deployments into an accumulating asset. Palantir spent nearly two decades unprofitable getting there, so the early years of any agency still look like the floor.

For a robot operator, the accumulating asset has a specific name: comparative performance data across platforms, in identical tasks, under one scorecard. That asset is only collectible by a neutral party. Customers already want this; access is scarce enough that companies rent robots through intermediaries and shared floors, though today's brokers (Formic and peers) cover conventional automation, and no intermediary yet brokers across humanoid OEMs. A manufacturer babysitting its own pilots cannot be the accountable operator across a mixed fleet, and the OEMs' own economics keep them out of the role: Apptronik burned on the order of a billion dollars just to reach manufacturing and handed production to Jabil; running national fleet operations on top is a second business the balance sheet does not have. The auto industry settled the same division of labor: carmakers ceded fleet management to Holman (whose fleet business dates to 1948), Element, and Ayvens, who run millions of vehicles the OEMs build. A true single-platform giant like Tesla can self-operate its captive fleet, and that share of the market is closed. The mixed-fleet remainder is the addressable market. GXO already runs three humanoid brands concurrently in its own warehouses, proving one party can operate across rivals. GXO does it inside its own buildings; no operator yet offers the service across customers.

Amazon's Delivery Service Partner program supplies the origination precedent: thinly capitalized local operators made financeable by an investment-grade demand guarantee and a scorecard.

$28US robot-teleoperation roles, an hourThe wage data prices it. US robot-teleoperation roles average about $28 an hour on the job boards. A Philippine industrial-robot operator earns about $4 an hour, against a 2026 minimum wage near $1.60. That spread is the arbitrage in one line, and at least one operator runs it as the stated strategy: Eastworld Labs operates thirty-plus humanoids from Kuala Lumpur on low-latency teleoperation into richer economies: robot business-process outsourcing, by design. 1X's Expert Mode operators, taking over the 30–40% of home tasks NEO can't finish, are the entry rung of the same ladder in its most consumer-visible form.

The number that would price the whole layer is not public anywhere: robots per operator, and the slope at which it improves. NEO's pricing implies a ceiling of well under an hour of human help a day; Agility's cost figures imply supervision ratios; managed-teleoperation providers sell around 60–65% operator utilization as a design target. But no operator has published its ratio, and the entire economics of teleop-backed deployment rests on that undisclosed number. As autonomy rises, the operator's share of each robot-hour falls and one person supervises more machines, so the labor footprint grows with the fleet. At what rate, nobody outside the operators knows.

62%ILA raise over six yearsThe 2024 longshore strike supplies a documented political precedent, and it did not end in a ban. In October 2024, 47,000 longshoremen struck 36 East and Gulf Coast ports; the ILA's opening demand was a total ban on automated cranes, gates, and container-moving trucks. The settlement, ratified in February 2025 with 99% approval, traded a 62% raise over six years and guaranteed job protections for limited automation. Automation proceeded with worker protections attached. Low-wage remote operation of machines inside rich-country workplaces reads as wage arbitrage when that is all it is, and the operators who survive scrutiny will need two things: a real ladder from teleoperator toward technician, and a published scorecard that makes the deal legible. No operator publishes a scorecard today. The first one to do so will become the industry's reference point in the hearings that follow.

$46Mrobots sold, opening festivalChina built the first version and runs it as state theater. The Beijing Robot Mall in Yizhuang opened August 8, 2025, timed to the World Robot Conference: four floors and 4,000 square meters, fifty-plus robot models from forty-plus manufacturers, ten anchor tenants including the Beijing Humanoid Robot Innovation Center, UBTech, and Galaxea. The shopping festival around its opening logged nineteen thousand robots and related products sold for over ¥330 million (about $46 million), with a quarter of the order value above a million yuan per order. A "2.0" upgrade followed within months. And the mall is one node in a national grid: a 5,000-square-meter humanoid "training school" opening in Shanghai's Zhangjiang district, a "6S store" demonstrated in Shenzhen alongside a ¥10 billion municipal AI-and-robotics fund with training, data, and model vouchers, and roughly forty government-funded training centers addressing the data shortage directly. The policy backdrop is measured: the formal deployment target is a modest ten thousand humanoids by end-2026, while MIIT's own officials estimate actual 2026 output above one hundred thousand. The state is solving the same deployment-and-data bottleneck the neutral operator solves privately, with public money, at national scale.

No Western analogue is in progress. No city, state, or company has built the neutral version: a capture floor running on posted consent instead of surveillance ambiguity, rival fleets rehearsing under one roof, and a public venue where fleet owners can watch their assets work.

$6.6BCHIPS grant to TSMC Arizona fabsSemiconductor projects bracket what a city might pay to host it, along with the risk. TSMC's Arizona fabs drew a $6.6 billion CHIPS grant plus up to $5 billion in federal loans against a $65 billion commitment. Foxconn's Wisconsin project drew roughly $3 billion in pledged subsidies for a promised $10 billion plant; the state spent about a billion dollars on infrastructure and the factory never came. Cities will pay heavily for an industry's front door, and they get burned when the demand guarantee is hollow. A robot building therefore needs working fleets and signed deployment contracts before public incentives. The first credible project should attract municipal bids.

7M TEUTriton runs, largest container lessorThe shipping container is the founding case, told from the operating layer. Malcom McLean's Ideal X sailed on April 23, 1956, with 58 truck-body containers; the Tantlinger corner casting and twist-lock became the design; McLean released Sea-Land's patents royalty-free so every country's boxes would share his geometry; ISO standardized the box in 1968, twelve years after the first sailing. Then the box commoditized (China builds nearly all of them) and the rent moved to the layer that operated and financed the standardized unit: Triton, the largest lessor, runs over seven million TEU, and the top five lessors control two-thirds of the leased fleet. Give away the format; own the network built on it.

Payment-card security shows a private rulebook becoming law-without-legislation. Visa launched CISP as one brand's internal security program in October 1999; PCI-DSS 1.0, modeled on it, shipped in December 2004; the PCI Security Standards Council followed in 2006, governing all five card networks. No statute anywhere requires PCI compliance; the networks enforce it by gating access, and everyone complies. PCI had a natural chokepoint: five networks controlling access to the payment system. Robotics has no equivalent gate today; financing can create one. EDI shows the buyer-mandate route: a coordinating committee's 1975 spec became ANSI X12 in 1979, and then Walmart and Kmart mandated it until three hundred thousand companies ran it. The CUDA-versus-ONNX split supplies the cautionary pair: a proprietary internal spec can become the standard and the lock-in, which is why an open interchange format emerged against it. A telemetry standard that labs, OEMs, insurers, and lenders must all trust needs an open format to win adoption from parties who refuse lock-in.

The robotics versions of these are already moving at the plumbing layer. MCAP, the open robot-log container format launched in 2022, is the pattern in progress: it began as Foxglove's tooling and is now the default in ROS 2 and NVIDIA's Isaac. VDA 5050, driven by the German auto industry, standardizes AMR-to-fleet-controller communication across brands, with OTTO, MiR, and third-party fleet managers certifying against it; Open-RMF and MassRobotics run the adjacent open-interop efforts; Formant and InOrbit sell the fleet-ops layer above. On the graded layer, NVIDIA's Halos holds the machine-safety certification seat with six cert bodies and no dataset scope, and SGS holds the one data-quality certification ever issued.

Four commercial standards are missing: no robot-hour of work is defined in ISO/TC 299, ASTM F45, or IEEE RAS; no standard fleet-servicing SLA exists; no environment-certification scheme for robots exists; no operator publishes its scorecard. Formic's uptime-guaranteed contract and GXO's RaaS agreement are the closest working templates, and they are private documents. An operator that answers those four questions in signed documents is doing standards work whether it calls it that or not. Because deployed formats tend to win, those answers could route everyone else's data through its pipes. The same operating budget that builds the service also establishes the commercial standard.