GXO Logistics (GXO), the world’s largest pure-play contract logistics provider, says its humanoid robots have moved more than 100,000 totes at one Georgia warehouse. This week’s news that two more sites will follow carried no price for a single tote, because the machines are rented rather than bought. Thirteen days earlier, AutoStore began selling a tool that builds the business case for buying its cube-storage systems, and that tool ships next to the vendor’s own math on the downtime and floor space a buyer will save. Put the two releases side by side and a quiet handover shows up. The numbers your payback rests on now come directly from the company selling you the gear, and the one figure that would settle the question, what a unit of work costs, is the figure the tote count skips.

The big picture

A payback calculation needs four inputs. It needs the price of the gear, the work a site really does in a year, the cost of the labor it replaces once attendance and turnover are counted, and the uptime the machines deliver. Sellers have always given buyers the first one. What changed this month is that they now sell the other three as products, built from a set of lookalike projects rather than from your building.

AutoStore states this plainly. CubeExplore uses benchmarks and data from comparable projects to help partners and would-be buyers build a stronger case, from someone else’s floors rather than their own. CubeStudio tests layouts before anything is built, and CubeFleet tracks upkeep and, on AutoStore’s own math, can cut surprise downtime by 35% to 45%. FlexBinsPort is credited with a 10% to 15% gain in space use, and each figure may be fair, though every one of them comes from the side that gets paid.

The same notice shows how far the repackaging goes. Pio Pal pitches piece picking to smaller firms as a lighter commitment under robots as a service (RaaS), reaching up to 540 order lines an hour at a single station. That is a real gain in clarity, because a price per line is easier to check than a sheet of guesses. It also puts those same guesses inside a contract, where the buyer sees the rate and still cannot see how busy the machines must be for that rate to hold.

GXO is the other half. It logs the work done: more than 100,000 totes at Flowery Branch, and more than 65,000 hours of live run time across Agility’s commercial operations. It notes what the machines can do and the factory that will build them, planned for up to 10,000 units a year. It logs a network spanning more than 1,000 sites and 200 million square feet. It carries no dollar figure anywhere, and no rate for an hour of robot work.

That is the clash. One firm publishes how much work a machine did, the other sells a model of how much work a machine will do, and neither publishes what one unit of that work costs. I read this silence as a structural shift rather than a secret, because under a service contract the buyer never sees a capital number at all. A buyer who signs a multi-year deal ends up paying down a plan it never measured, at a rate it cannot take apart.

By the numbers

  • 100,000-plus — Totes moved: the volume GXO claims for its first live site in Georgia, where the machines have worked since June 2024 [1].
  • 65,000-plus — Hours logged: the live run time Agility reports across its commercial operations, given as a total with no fleet count to divide it by [1].
  • 35% to 45% — Cut in downtime, vendor estimate: what AutoStore says its upkeep alerts can achieve, offered as an input to the case rather than a measured result [2].
  • 540 — Order lines an hour: the speed AutoStore quotes for its lighter picking offer at one station, for orders that are already picked [2].

What I’d watch

  • The first measured cost per tote. I am watching for a buyer to publish this, because the number lives in a logistics provider’s customer deal or in a firm’s own time study, and both stay private by default.
  • The audit clause in the next large deal. A buyer who cannot take a rate apart will sooner or later ask for the throughput guess, the site baseline and the list of what was left out. The reply to that request is the true test of the model.
  • The term behind “lighter commitment”. AutoStore uses this phrase for Pio Pal rather than calling it a short deal, and the exit terms are exactly where a mid-market operator now carries the risk that a fixed purchase used to carry.
  • Factory output sold ahead of the proof. Agility sized its plant for up to 10,000 robots a year while its public record is 65,000 hours of work. The question is whether that capacity gets placed through pilots rather than through repeat, measured deals.

The catch

The honest limit here is that sellers often know more than their buyers. A storage supplier with two thousand systems in the field has a stronger view of a typical project than a mid-market operator with no throughput data of its own. Shrugging off that knowledge would be a mistake, and a buyer who insists on measuring everything alone will take longer and land on a worse first guess.

Renting the machine, the route this desk examined two weeks ago, does not remove the model either. It moves the same guesses into a rate, which is easier to read on a bill and no easier to check. The 100,000 totes are real work. A humanoid that has handled that volume in a live building has shown it can run, and that is not nothing.

What the count cannot show is that the site pays. My view is that since no party publishes a payback or a cost per tote, any figure a reader has seen for robot economics came from a model, and every model of this kind has an author. The one number worth building yourself is the cost of the work you already do, because it is the only input no vendor can hand you.

At a glance

  • The Big Shift: In the same fortnight, AutoStore began selling a GenAI tool that builds the business case for its own systems, and GXO expanded a robot fleet whose public proof is a count of totes moved. The case for warehouse automation now comes from the seller.
  • Why It Matters: A payback rests on how busy the machines are, how often they break, how much space they save and how much labor they replace, and all four are now vendor math. An operator that takes them at face value is paying down someone else’s guesses, which is how a fleet can look proven at 100,000 totes and still carry no rate anyone can defend.
  • What I’d Watch: I am watching for the first buyer that asks for the guesses in writing, because that request is what turns a vendor model into a decision a finance desk can stand behind.
  • The site baseline: the operator’s own count of units handled per year, which no set of lookalike projects can supply.
  • The loaded labor rate: wage plus attendance, overtime and turnover, the figure a vendor guess tends to understate.
  • The contract terms: length, exit and repricing clauses, which decide who holds the throughput risk when the rate moves.
  • The Catch: Vendor benchmarks are better than most buyers’ guesses, and RaaS makes a rate visible where a capital number used to hide. Neither fixes the deeper problem, which is that nobody in this market publishes what a unit of warehouse work costs.

Related reading

Sources

[1] GXO Logistics, “GXO and Agility Robotics Expand Partnership to Accelerate Deployment of Humanoids in U.S. Operations,” news release, Oct. 26, 2026. https://gxo.com/news_article/gxo-and-agility-robotics-expand-partnership-to-accelerate-deployment-of-humanoids-in-u-s-operations/ [2] AutoStore, “AutoStore Advances the Path Toward Autonomy with Fall 2026 Product Announcement,” news release, Oct. 13, 2026. https://www.autostoresystem.com/news/autostore-advances-the-path-toward-autonomy-with-fall-2026-product-announcement [3] DC Velocity, “GXO to expand use of humanoid robots from Agility,” Oct. 26, 2026. https://www.dcvelocity.com/material-handling/robotics/gxo-to-expand-use-of-humanoid-robots-from-agility