JD Logistics launches China's first 'autonomous vehicle plus metro' city delivery model

The Analysis

Whenever a logistics operator puts a new city delivery model in front of me, I reach for the same mental picture: a queue. Arrivals on the left, a service station in the middle, departures on the right, and a waiting line that nobody photographs. What JD Logistics described in Shenzhen is a city delivery model that puts its own unmanned ground vehicles on the two ends, collection at one end and the final hand to the customer at the other, and borrows spare off-peak capacity on the metro network for the long middle leg, with no intermediate sortation hop in between. The company put three things on the table: transport cost down by roughly 60 percent, capacity utilisation up by about 10 percent, and same-day movement and delivery across the city. My reaction is not excitement. My reaction is: which station did you delete from the queue, and what did that do to the tail?

The three claims are not the same kind of claim, and that is the part people skip. A cost ratio is only as good as the denominator it is measured against. A utilisation number tells you about somebody's asset, and I want to know whose. A same-day promise is a statement about the middle of a distribution dressed up as a statement about the whole distribution, and the tail of that distribution is where your customer complaints live. If you are an importer or exporter moving goods into or out of a dense Chinese city, you are not buying a robot and you are not buying a metro ticket. You are buying a change in the shape of a delay distribution, and you should price it exactly that way.

I have watched versions of this before: rail and metro capacity repurposed for freight in the quiet hours, automation grafted onto routes that were never designed for it. The pilots that survived shared one property. Somebody, by name, owned the fallback when the scheduled run did not happen. The ones that quietly died had a beautiful throughput chart and no answer to the question of what happens when you miss the train. Before anyone in your company forwards this to a board slide, model it, put numbers on it, and only then talk about optimisation.

Start with the picture. A parcel arrives at a city gateway, gets collected, rides to a sortation point, gets sorted, rides again to a delivery base, gets handed to a courier, and reaches a door. Six touches before the customer sees anything. Every touch is a service station with its own mean service time and its own variance, and every handover between two stations is a place where a parcel can sit, get mis-scanned, get damaged, or disappear for four hours while nobody owns it. If you have ever wondered why a parcel that was ten kilometres away took two days to arrive, the answer is almost never the driving. The answer is the waiting.

In queueing terms the flow has an arrival rate and each station has a service rate, and the ratio between them is utilisation. Waiting time does not grow in a straight line with utilisation. It bends. Push a station from 70 percent utilised to 90 percent utilised and the queue behind it does not grow by a tidy fraction, it can multiply, because the backlog only clears when arrivals happen to pause. This is why the question of where the bottleneck is always comes first, and why the honest answer is usually the station nobody has instrumented.

So what did the Shenzhen model actually do to the queue? On the description published, it removed the intermediate sortation hop. That is not a small change. Deleting a station is the strongest lever available in a flow, stronger than speeding one up. It removes the station's service time, removes the queue in front of the station, and removes the variance the station contributed. If a sortation hop used to occupy, on average, ninety minutes of a parcel's life including the wait to be processed, then deleting it does not save ninety minutes of driving. It saves ninety minutes plus whatever queue was sitting in front of it.

There is a catch that belongs in the same breath. Stations rarely disappear for free. A metro line is not a truck that leaves when it is full; it runs on a timetable designed for passengers, not parcels. The moment you move your line-haul onto a scheduled service you have converted a continuous server into a batch server. Parcels arriving one minute after a departure do not travel, they wait for the next departure. You have traded a variable service time for a variable waiting time, and depending on the headway that trade can be excellent or quietly terrible.

Let me take the cost claim apart, because this is where most of the mispricing happens. Transport cost down by roughly 60 percent. The word transport is doing an enormous amount of work in that sentence. Is the 60 percent measured on the line-haul leg alone, on the whole in-city movement including collection and last-mile, or on the operator's internal cost of running the flow as a business? Those three denominators can differ by a factor of three, and if you assume the wrong one you will build a budget line that does not survive contact with an invoice.

Here is the same problem with the utilisation claim, and it is subtler. Capacity utilisation up by about 10 percent. Whose capacity? If the number refers to the metro's spare off-peak capacity being filled, then it is utilisation on somebody else's balance sheet. It is good for the transit operator and it makes the service politically easier to keep running, but it does not move your unit economics at all. If the number refers to load factor on the unmanned ground vehicles, or to the share of a vehicle's available daily hours that is actually carrying revenue parcels, then it is your number and it belongs in your model. Two very different meanings behind one percentage. Ask which one before you repeat it.

The third claim is the one I care most about and the one with the least information behind it: same-day citywide movement and delivery. Same-day is a promise about a typical parcel. It says nothing about the slowest five percent, and in my experience the slowest five percent generate more than half of the customer contacts. If I am buying this service I am not buying same-day. I am buying a distribution, and I want two numbers from it: the median door-to-door hours and the ninety-fifth percentile door-to-door hours, both measured from the moment the parcel is handed over rather than from the moment an order was placed.

Now the arithmetic, with the assumptions written where you can argue with them. Assume your current all-in city delivery cost per parcel is 5.00 yuan, split as follows: collection 0.50, first-mile movement 0.70, sortation 0.80, line-haul 1.30, last-mile 1.70. Assume further that the 60 percent applies only to the line-haul leg, because that is the leg the metro replaces. Line-haul becomes 0.52, you save 0.78 per parcel, and the all-in cost falls from 5.00 to 4.22. That is a 15.6 percent saving on your delivery cost. Not a 60 percent saving.

Now assume a friendlier reading: removing the intermediate sortation hop takes the sortation cost out as well. You save 0.78 on line-haul and 0.80 on sortation, so 1.58 per parcel, and the all-in cost falls to 3.42. That is a 31.6 percent saving. Still not 60 percent, and this is already the generous reading. To reach a genuine 60 percent reduction on a 5.00 base you would have to take 3.00 yuan out of the flow, which would mean the collection and last-mile legs change materially too. That is a different claim and it should be stated as one.

Put a volume against it. Assume 15,000 parcels a day through this city and 300 operating days a year, so 4.5 million parcels. The narrow reading, 0.78 yuan saved per parcel, is about 3.5 million yuan a year. The generous reading, 1.58 yuan per parcel, is about 7.1 million yuan a year. The headline reading, 3.00 yuan per parcel, would be 13.5 million yuan a year. Three numbers out of one press item, with a four-fold spread between the cautious and the enthusiastic version. This is exactly why the order has to be quantify first, then optimise.

Now subtract what the item does not mention. Assume you need one integration engineer for three months to connect your order system to the new handover events, at a fully loaded 45,000 yuan a month, so 135,000 yuan. Assume that during the first eight weeks your exception rate runs half a percentage point above normal, that you are handling 15,000 parcels a day, and that each exception costs 12 yuan to resolve in manual re-runs and customer service. That is 75 extra exceptions a day, 4,200 over eight weeks, roughly 50,000 yuan. Both numbers are small next to the annual savings, and that is the honest conclusion: the ramp cost is not what kills a pilot like this. A broken fallback is.

Which brings me to the question I ask first in every one of these conversations. When the model fails, who covers it? Not what is the contingency plan, which is a phrase that means nothing. Who, in writing, pays for the road re-run when a metro window is missed? Who pays when a customer promise is broken because an unmanned vehicle could not complete a handover? In your business that cost lands on your marketplace service level, your chargeback rate and your seller rating. In the operator's business it lands on a service credit that may or may not exist. Get the answer in the contract, in numbers, before the first parcel moves.

If you are going to test this, test it like an experiment rather than a rollout. Pick one district for the new flow and one district that stays on the current road-plus-sortation flow. Match them on parcel mix, on average delivery distance, on daily volume band, and on the time-of-day profile of order arrivals, because a district whose orders arrive at nine in the morning behaves very differently from one whose orders arrive at nine at night. Run both arms for at least six weeks and do not read the result before then except for safety. Write the metric definitions down in week zero, before anyone has an incentive to redefine them.

The metrics I would collect, all of them per parcel or per thousand parcels so that volume swings do not confuse you: all-in cost per delivered parcel; median and ninety-fifth percentile door-to-door hours from handover to signature; first-attempt delivery rate; exception rate per thousand parcels; damage rate per thousand; metro handover miss rate, meaning parcels that reached the handover point and did not make the run they were planned for; human takeover rate per hundred kilometres on the unmanned vehicles; and cancellation rate, because a service that looks cheap until customers walk away is not cheap.

Then write the stop lines into the same document as the business case, not into a risk register nobody reads. Mine would read like this. Freeze any expansion if the ninety-fifth percentile door-to-door time in the test district runs more than 20 percent above the control district for two consecutive weeks. Freeze if the exception rate exceeds the control district by more than half a percentage point on a rolling two-week window. Freeze if the metro handover miss rate exceeds 2 percent. Freeze if the human takeover rate on the unmanned vehicles exceeds 3 per hundred kilometres. And freeze if the all-in cost per delivered parcel is not at least 10 percent below the control district by the end of week six. Those thresholds are mine, not the operator's, and the point of writing them down is that you stop on evidence rather than on mood.

Back to the queue, because the scheduled-server point deserves numbers. Assume the usable off-peak runs on the metro line you would ride are 45 minutes apart during your window. Assume also that parcels reach the handover point uniformly across that window. The average added wait is then about half the headway, roughly 22 minutes, and the unlucky parcel waits the full 45. Compare that with a road line-haul that departs when a vehicle is full: on a dense lane it might leave every 15 minutes, on a thin lane every two hours. Whether the metro version is faster depends entirely on the headway and on how close to a departure your parcels naturally arrive. Five minutes of arithmetic tells you more than any press release.

There are two permission sets between the model and your cargo, and both are yours to verify rather than the operator's to promise. One is road access for the unmanned vehicles: which routes they may use, which districts they may enter, the insurance position, the remote supervision arrangement, and what happens in heavy rain or after dark. The other is carriage on the metro: which commodity classes may travel, what size and weight limits apply, what screening is required, and who is liable between handover and collection. I am not going to quote a regulation in a column like this, because the specifics differ by city and they change. What I will say is that somebody on your side should hold a current, dated, written permission covering your actual districts and your actual commodity list. If your mix includes batteries, liquids, aerosols or anything pressurised, check that list before anything else, because that category is usually where these schemes stop.

The inventory angle is the part that genuinely interests me, and it is the part where people get fooled. If same-day citywide movement is real and, more importantly, consistent, then the replenishment lead time into your city stocking point shortens, and safety stock falls with the square root of lead time. Assume daily demand of 1,000 units with a standard deviation of 400 and a 95 percent service target, so a z of 1.65. At a replenishment lead time of two days with a lead time standard deviation of half a day, safety stock works out near 1,250 units. Move the mean lead time to half a day at the same variability and it lands near 620 units. That is a real release of working capital, and it is the strongest argument in the model.

Now the trap. Run the same calculation with the lead time variability you would actually get from a scheduled service. Keep the half-day mean, but set the standard deviation at 1.2 days, because missing a run does not cost you minutes, it costs you a whole headway or a whole day. Safety stock goes back up, to roughly 2,000 units. You end up holding more stock on a faster service than on a slower one. That is not a paradox, it is simply what variance does, and it is the single most common way a faster mode fails to pay for itself. A shorter mean with a fatter tail is a bad trade, every time.

Contract definitions will decide whether you capture any of this. Ask what same-day means in the document: same day by which clock, measured from which event, and with what cut-off for order acceptance. Ask how long parcels may sit at a handover point before dwell charges start and who pays them. Ask the billing basis, because per parcel, per kilogram and per volumetric kilogram rank your SKUs differently, and a rate card that looks cheap per parcel can be expensive on a bulky mix. Ask about oversize and overweight surcharges, minimum volume commitments, and who insures the parcel during the metro segment. And ask for event-level scan data rather than a daily summary, because without timestamps at each handover you will never compute the percentiles that matter.

On alternatives, keep the list short and priced. You can stay on road line-haul and improve consolidation, which is boring and often worth 10 to 15 percent. You can move inventory closer to demand and accept more nodes holding more safety stock in total, which costs working capital but removes the dependency on any single mode. You can change the promise you make to customers rather than the network that serves it, which is free and frequently overlooked. Whichever you pick, price the way back: how long to return to the old flow, what it costs, and whether the contract locks you in for a term that outlasts your confidence. Reversibility is a real economic attribute of a pilot.

The wider claim in the item is that this is a replicable template for dense urban markets. Replicable is doing work in that sentence too. For the model to reproduce in another city you need, at minimum, parcel density high enough to fill a scheduled run, a metro network with genuine spare off-peak capacity on the corridors you use, a local rulebook that lets unmanned vehicles operate on the roads you need, and a commodity mix the metro will accept. Score your own cities on those four before you copy anything. Two out of four is not a template, it is a pilot.

I do not think this is marketing noise. Removing a station from a queue is the real thing, and a real change in the shape of a queue is worth real money. What I have in front of me is three claims from the operator: one hard percentage with an unstated denominator, one ambiguous percentage, and one promise with no distribution behind it. Our job this quarter is to convert each of them into a measurement with an owner, a baseline and a line that stops us. Quantify first, then optimise. That order is not caution, it is just how you avoid paying for a chart.

↑ Back to the key points

— By Ravi Chandran