Total Cost of Ownership for a Warehouse Humanoid: The 14 Cost Lines Vendors Omit

A Digit-class humanoid leased through Robots-as-a-Service costs roughly $30 per operating hour. The US warehouse worker it would displace costs $22 per hour fully loaded. That gap looks like a closed argument against humanoids — until you include the 14 cost lines that vendors leave off the sticker.

Total cost of ownership for a warehouse humanoid is not a single number. It is a function of duty cycle, downtime, integration scope, and how aggressively the platform ages. The published per-hour figures from Agility, Apptronik, and Figure are useful starting points and bad ending points. Below is what actually rolls into the spreadsheet, with numbers traceable to public sources.

Capex versus RaaS — two different problems

Two ownership models dominate humanoid warehouse work in 2026. Outright purchase puts a Unitree H1 at $90,000, an H1-2 at $99,900, a G1 at $16,000, and a Boston Dynamics Atlas Electric at an unpublished figure that industry reporting (The Robot Report, March 2025) places between $150,000 and $200,000. Figure does not sell hardware. Tesla has not opened the Optimus order book. Apptronik moved Apollo to a leased model after early customer pushback on capex.

Robots-as-a-Service has become the default for industrial customers. Agility Digit ships under multi-year RaaS contracts to GXO Logistics, with Damion Shelton quoting a target operating range near $30 per hour at TechCrunch Disrupt 2024. Apptronik has not disclosed Apollo rates publicly; Mercedes-Benz characterized its 2024 trial as "pilot pricing" without numbers. Figure's BMW deployment runs under a confidential agreement.

The split matters because capex and opex hit different parts of the procurement budget. A $200,000 Atlas Electric purchase enters as a fixed-asset depreciation schedule across three to five years. A $30/hour Digit lease enters as operating expense and clears procurement faster but compounds over the platform lifetime. At 2,000 productive hours per year — single shift, no overtime — RaaS at $30/hour totals $60,000 per robot per year. After three years the cumulative spend lands at $180,000, in the same range as buying an Atlas outright but without ownership at the end.

The 14 cost lines that vendors omit

The sticker price covers the unit. The operating budget covers the deployment. The gap is large and consistent across customers we know about.

Integration engineering. Connecting a humanoid to existing warehouse management systems, AMR fleets, and conveyor controls runs $50,000 to $150,000 per site in published Apptronik and Agility deployment notes. This is one-time but always understated in pilot proposals. Operator training. GXO trained four cross-functional staff per Digit deployment site, per their June 2024 announcement. At a fully loaded operator cost of $35/hour and 80 hours of certified training each, that adds $11,200 of front-loaded labor before the robot does any productive work. Charging infrastructure. H1 ships with a 864 Wh battery and a single-unit charger. Multi-robot operations need bay chargers, distribution wiring, and in many warehouses an upgraded service panel. Public Digit deployment notes from Amazon's BFI1 facility describe new 80A circuits added specifically for charge bays. Downtime and spare units. A humanoid that cannot work earns nothing. Agility's published mean-time-between-failure data does not exist; The Robot Report's 2024 Digit field profile estimated MTBF in the 200-400 hour range for first-generation units, with maintenance cycles consuming 5-10% of calendar time. The industry rule of thumb is one spare per six deployed units, which moves the effective per-hour cost up by roughly 17%. Insurance. US warehouses operating any robot above 50 kg with collaborative work envelopes typically carry a separate liability policy. Brokers we have spoken to quote $3,000 to $8,000 per robot per year for fleets under 20 units, with the rate dropping at scale. Certifications. ISO 13482 for personal-care robots, IEC 62443 for control system cybersecurity, and local fire-safety inspections all carry recurring audit costs. Agility has stated publicly that ISO 13482 work for Digit is in progress as of 2026 but not complete; once complete, the certification cost passes through to customers. Software updates and platform fees. Several humanoid platforms now ship with paid software subscriptions on top of hardware lease. Apptronik's Apollo platform tier was disclosed at $1,200/month per unit in a leaked Mercedes pilot document covered by Bloomberg in February 2025; Apptronik has not confirmed or denied the figure. Network and compute backhaul. Humanoids that run cloud-side ML inference need reliable, low-latency network egress. Real warehouse network conditions rarely match the lab. Customer integration notes from Amazon's BFI1 program describe a private 5G overlay specifically to keep Digit telemetry within the 50 ms round-trip the controller expects. Decommissioning. A humanoid at end-of-lease has to leave the building. Some RaaS contracts include logistics; some do not. Trade-in residual credits, where they exist, run 30-40% of acquisition price after three years per BiorobotAI's internal residual tracker. Unitree H1 specifically is tracking at 35% residual by year three.

The other lines — facility modifications, payroll for a robotics technician on staff at large sites, third-party testing for safety-critical work, and the executive time spent justifying the program upward — vary too much for industry-wide numbers but are real.

Energy is small. Runtime is the real cost.

Energy per shift is not the bottleneck people expect it to be. A Unitree H1 running its 864 Wh battery to depletion in 105 minutes draws about 0.49 kW continuously. At US industrial electricity rates of $0.08/kWh, that comes to under $0.04 per operating hour. Even at European rates near $0.25/kWh it stays under $0.13/hour.

Runtime, not energy, is where the budget breaks. H1's 105-minute envelope means a single robot covers a fraction of a shift before swap-out. Without hot-swap (H1 has none) the operation either accepts gaps or buys a second unit per swap cycle. Agility Digit has hot-swap and a 240-minute base runtime, which is why Digit shows up in real warehouse deployments and H1 does not. Figure has not published runtime numbers for Figure 02. The Boston Dynamics Atlas Electric runtime is undisclosed but Marc Raibert's 2024 talk at CMU described battery life as "the constraint we are pushing hardest on."

For TCO purposes, treat runtime as an effective utilization multiplier. A platform with 240 minutes of work per battery and 15 minutes of swap covers an 8-hour shift cleanly with two batteries. A platform with 105 minutes and no hot-swap covers roughly half a shift before the operation has to schedule around it. The hourly rate is the same on paper but the productive hour count per robot per day differs by a factor of two.

Depreciation patterns differ across manufacturers

Capex buyers face a depreciation question that RaaS customers do not. The honest answer in 2026 is that humanoid residual values are still being discovered by the secondary market.

BiorobotAI's internal residual tracking, which compares published list price against verifiable secondary-market transactions and end-of-lease buyout offers, shows three patterns. Unitree H1 holds 70% of value at year one, 50% at year two, and 35% at year three — driven down by the H1-2 release and the expected H2 announcement window. Boston Dynamics Atlas Electric holds value better because of constrained supply, sitting near 80% at year two on the rare resale, but the small transaction sample makes that number soft. Agility Digit has effectively no secondary market because Agility retains hardware title under RaaS; the question does not apply.

The pattern that matters for procurement: humanoid generations are turning roughly every 18 months. A three-year ownership horizon on capex hardware almost guarantees the platform will be one or two generations behind by exit. Buyers who treat humanoids as five-year capital equipment in the spreadsheet will be wrong on residual by 30-40%.

Realistic payback windows by scenario

Three real-world configurations, with TCO numbers built from the components above.

Scenario A: Single-shift tote transfer, RaaS. GXO-style deployment. One Digit on a Robots-as-a-Service contract at $30/hour, 8-hour shifts, 250 working days per year, plus $80,000 in one-time integration and training. Annual run rate: $60,000 RaaS + $26,667 amortized integration (3-year basis) = $86,667. Displaces approximately 1.2 full-time warehouse workers at $52,000 fully loaded each = $62,400. Net cost: $24,267/year above the human baseline. Payback against labor savings: never, under these assumptions. The deployment justifies itself on injury reduction, retention, and shift coverage rather than labor cost. Scenario B: Two-shift operation, RaaS, mature platform. Same Digit-class unit at $25/hour (assumed 15% RaaS discount at scale), 16 hours of productive work per day across two shifts using a swap cycle, 250 days per year. Annual: $100,000 RaaS + $26,667 amortized integration = $126,667. Displaces 2.4 FTEs = $124,800. Net: break-even by year two, marginally positive by year three. This is roughly where the industry is starting to see real ROI calculations work. Scenario C: Owned hardware, three-shift research lab. Unitree H1 purchase at $90,000, plus $40,000 in setup and SDK development, plus $9,000/year operating cost (electricity, parts, occasional repair). Lab use does not displace labor; the ROI question is research throughput per dollar. At a three-year horizon the cumulative cost is roughly $157,000 with a $31,500 residual recovery — net $125,500 over three years for a research platform. Whether that pencils against alternative hardware (Cassie, Atlas Electric where available, simulation-only stacks) depends entirely on what the lab is trying to publish.

What the published numbers actually tell you

Two things stand out from the public data. First, single-shift RaaS does not pay back on labor alone — it pays back on availability, injury rates, and the ability to staff shifts where workers refuse to take the job. Second, the platforms that win industrial deployments in 2025 and 2026 are the ones with hot-swap batteries, ISO 13482 progress, and RaaS pricing. Hardware specs (DOF count, height, speed records) are not driving procurement at the warehouse level.

The humanoid TCO conversation in 2026 looks more like the conversation around forklifts in 1985 than the conversation around laptops in 2010. Buyers are not optimizing total cost. They are buying around labor shortages, injury liability, and shift coverage. Vendors who sell that story honestly, with the 14 cost lines on the table, are closing deals. Vendors who only show the per-hour sticker are not.