China's Cole Winke signs embodied-AI warehouse robotics deal at WAIC

The Analysis

On July 18 at WAIC 2026, Cole Winke, the supply-chain arm of China Technology, signed a strategic cooperation with Daka Robotics to industrialise embodied AI in its warehouses. The numbers attached to the release are the interesting part: close to 200 warehouses, roughly one million square metres, five million orders on a peak day, 99.95 percent inventory accuracy. Real operating scenarios are being opened for robot-as-a-service deployment, which means the machines arrive as a metered service rather than as capital equipment you own.

Model it before you admire it. A warehouse is a queueing network: receiving, putaway, picking, packing, staging, dispatch, plus one small station nobody photographs where the exceptions go. Adding capacity at a single station raises system throughput only if that station was the constraint. So the question is not whether the robots work. It is which station in a five-thousand-square-metre building was actually binding.

The density in these numbers is the first thing I would put on a whiteboard, because it decides what kind of problem this is.

One million square metres across 200 sites averages 5,000 square metres a site, which is a mid-size building, not a mega-DC. Five million peak orders across the same 200 sites averages 25,000 a site on the peak day. That is five orders per square metre per day at peak, and I am assuming an even split, which no network has ever had. The consequence is that at this density the scarce resource is floor, not motion. A picker's travel time is a function of how much of the building is walkable, and staging lanes, charging bays, safety fencing and robot corridors all come out of that same floor.

Read the 99.95 percent again, this time as a subtraction. The remaining 0.05 percent of five million is 2,500 exceptions on a peak day across the network, roughly twelve or thirteen per site. A site that has been running for a decade absorbs twelve exceptions a day with two people and a trolley. Which tells you something the release does not: this deal cannot be sold on accuracy. There is almost no headroom left. If precision were the pitch, the business case would be a rounding error. The case has to rest on labour hours per line and on cubic utilisation, and those are the two numbers the announcement withholds.

Follow the labour argument into its mechanism. Robots raise lines per labour hour at the pick face. That is a service rate, and a service rate only helps if the queue in front of it is the long one. If inbound receiving and putaway are unchanged, the higher pick rate pulls work forward into the staging buffer, and the buffer is floor. Floor is fixed at 5,000 square metres. As the buffer fills, pick faces get blocked, travel paths lengthen, and the effective pick rate slides back toward where it started. That is a balancing loop, and it explains the automation pilot that looks magnificent in month one and flat in month six. The gain is real. The system eats part of it. You keep it only by widening the buffer or shrinking the variance feeding it.

Which brings up the parameter that actually moves the system: the coefficient of variation of inbound and of order mix, not the mean. Two sites with identical daily volume need different buffers if one takes full pallets on a fixed schedule and the other takes mixed cartons whenever a supplier feels like shipping. The 99.95 percent is an average across a network, and averages hide tails. The tail concentrates in the SKU classes that look alike on a shelf and in the orders that mix them, which is precisely where a vision model has the least margin and an experienced picker has the most. Embodied AI does not remove that tail. It moves the same tail, at the same rate, onto fewer people, who now supervise machines instead of picking.

Split the network by site type and the rollout order writes itself. Stable, high-velocity, single-temperature sites with predictable order profiles go first, because variance is low and a robot cell can be packed in densely. Sites with seasonal assortment, high SKU churn, or mixed pallet-and-each picking wait, possibly for years, because each of those features lifts the exception rate and depresses utilisation of a metered robot. For an importer the consequence is asymmetric. If your cargo lands in an automated node, your own receiving discipline becomes the binding constraint on your side of the fence, and a missed appointment costs more than a late truck used to. If your cargo lands anywhere else, nothing changes this year except possibly a rate review to fund somebody else's capex.

Direction matters as well. Full-pallet inbound into these buildings is the most compatible flow, because a pallet is a unit a robot can reason about. Loose cartons, LCL deconsolidation, anything arriving without a stable unit load, are the least compatible, and those are exactly the flows where a human exception desk earns its keep. Exporters shipping out of such a node meet the mirror image: cut-off times tighten, because outbound is now sequenced against a robotic pick plan that tolerates a late truck far less gracefully than a shift supervisor did.

Announcement to pilot is a quarter or two. Pilot to a line running in production is another two to four quarters, because the hard part is not the arm, it is teaching the warehouse management system to hand work to a fleet and to take it back when the fleet stops. What an importer actually experiences is a change in appointment windows and cut-off times at the next contract cycle, plus a revised accessorial schedule. Peak is where it becomes visible, because peak is when the buffer is already gone.

Here is the part no conference stage covers. Every automated warehouse that underperforms, underperforms at the handoff, not at the robot. A tote reaches the exception station without a location code the human can act on. A pallet sits eight centimetres out of tolerance and the cell stops. An order needs a substitution decision nobody wrote a rule for. Each of those is a transfer of state between a machine and a person, and every transfer carries a cost in minutes and a failure rate in percent. Count them. Handoffs per thousand order lines is the most predictive number I know for whether an automation programme holds its gain, and it is missing from every press release because it is unglamorous and because it is a design choice, not a purchase.

Now the arithmetic. All of it is my assumption and none of it comes from the release. Take one 5,000 square metre site at 20,000 orders on an average day, 1.4 lines per order, so 28,000 lines. Assume a manual picker sustains 80 lines an hour across a 7.5-hour effective shift, which is 600 lines, so about 47 picker-shifts a day. Let the robot cells cover 30 percent of those lines in year one, 8,400 lines, roughly 14 shifts. If cells plus charging and fencing take 12 percent of the pick area, you have traded 600 square metres of floor for 14 shifts.

Whether that is a good trade depends entirely on your wage and your rent, and it flips if the floor you freed is what lets you hold the peak buffer. Then the tail: at a constant 99.95 percent, exceptions scale with volume. Grow lines by 25 percent and you add 625 exceptions on a peak day network-wide, which is about 125 labour hours at twelve minutes each, about 16 shifts nobody budgeted, because the automation case only ever counted the pickers.

The counterfactual is not no robots. It is sequence. Re-slot so the fastest movers sit nearest the pack bench, smooth inbound to a two-hour appointment window, re-cut the wave logic, and you recover a meaningful share of the same throughput with no capital and no integration risk. That work is boring, it is measurable, and it is a prerequisite, because automating a badly slotted building locks the bad slotting into steel. What reverses my conclusion is low variance plus high labour cost plus a lease long enough to amortise. In that case buy the robots and skip the poetry. In every other case the queue in your building is already telling you which station to fix, and it is usually not the one being sold.

↑ Back to the key points

— By Ravi Chandran