LA Overseas Warehouse: 11,016 Bins in 600 sqm, Storage Up 400%, Labour Cost -75%

The Analysis

Eleven thousand and sixteen bins. Six hundred square metres. Storage capacity up four hundred per cent. Labour cost per unit down seventy-five per cent. Pick speed three times what it was. Those five numbers are the entire story coming out of a Los Angeles overseas warehouse, and not one of them is a fact. Each one is a ratio that got dressed up as a fact somewhere between the operator's press release and the headline you read. That is not a knock on the warehouse. It is a warning about what you are allowed to do with the numbers, because the numbers are about to be used to sell you something.

I read a warehouse the way I read everything else in this trade: as a queue with a roof on it. Cartons arrive, they wait, somebody or something works on them, they wait again, they leave. Storage is the waiting room. Picking is the server. The dock door is the exit. When somebody tells me the waiting room got five times bigger and the server got three times faster, the question I ask is not how impressive that sounds. It is which of the two was actually the reason things were slow. If the answer is neither, you have bought a very elegant machine that changes nothing.

So I start where I always start: where is the bottleneck, and which single variable moves the whole system when you touch it. Here I think I know the answer, and it is not the gantry robot. It is density. Roughly eighteen bins to a square metre changes the price of holding stock near the US West Coast, and once the price of holding stock moves, every inventory parameter sitting downstream of it is suddenly the wrong number. That is the part worth an evening of your time, not the robot.

The facts, kept exactly as they were reported. A Los Angeles overseas warehouse put 11,016 bins into 600 square metres using a gantry robot and a three-dimensional shuttle system. Storage capacity rose 400 per cent. Labour cost per unit fell 75 per cent. Pick speed tripled. That is the whole of what was published. Everything after this paragraph is my model of it, and I will say so each time I am modelling rather than quoting.

Why now, and who is pushing. A warehouse is a machine for converting rent and labour into shipped units, and on the US West Coast both inputs have been moving against the operator for years. Industrial rent near the port does not come down in any year that matters, and California warehouse labour has been repriced by wage floors, by turnover and by a tight market. When both inputs rise, the only way to hold cost per unit flat is to change the denominator: get more units through the same box. Storage density and pick automation both attack the denominator, which is why they show up together and why they show up now, in the run-up to a peak season where the cost of being unable to ship is higher than the cost of the hardware.

Start with the arithmetic nobody printed. Eleven thousand and sixteen bins into six hundred square metres is 18.36 bins per square metre. Round it to eighteen. Now run it backwards. If capacity is up 400 per cent, the new level is five times the old one, which means the old site held about 2,203 bins in the same 600 square metres, or 3.67 bins per square metre. If the phrase instead means the new level is four times the old one, the old figure is 2,754 bins, or 4.59 per square metre. Either reading puts the old density between three and a half and four and a half bins per square metre, which is entirely plausible for conventional racking picked on foot. So the headline is internally consistent, and that is worth saying out loud, because a lot of automation claims are not.

But eighteen bins per square metre is not a property of the building. It is a property of the bin. The report does not say how big a bin is, and that omission is not a footnote, it is half the story. If a bin holds one shoebox, your 11,016 bins hold 11,016 shoeboxes. If a bin holds forty phone cases in an inner pack, the same 11,016 bins hold 440,640 phone cases. Same robot, same building, forty times the answer. Anybody quoting 18.36 at you as though it were a constant of nature is selling something. The variable that actually decides your capacity is units per bin, and nobody published it.

Now the part I enjoy. Think of it as a queue. Orders arrive as order lines at some rate, call it lambda lines a day. The picking system serves them at a rate, call it mu. Nothing about a warehouse gets interesting until you look at utilisation, rho, which is lambda over mu. Pick speed tripling turns mu into three mu, so at unchanged volume rho falls to a third of what it was. Waiting time in a queue behaves roughly like rho divided by one minus rho, and that function is the reason a warehouse feels calm one week and like a disaster the next. At rho of 0.83 the ratio is about five. At rho of 0.28, which is a third of that, the ratio is about 0.39. The queue does not get three times shorter. It gets roughly thirteen times shorter.

That is the real product on sale, and it is not speed. It is headroom. Run the same arithmetic backwards: if your old busy day was rho of 0.83 at a thousand lines a day, the new system does not reach that same level of congestion until roughly three thousand lines a day. You have bought the ability to absorb a threefold demand spike before the building feels as bad as it used to feel on an ordinary Tuesday. Headroom is the thing you actually monetise. You monetise it by carrying more stock, by getting through peak without a hiring scramble, and by quoting delivery promises you can keep. Speed is just how headroom looks on a spec sheet.

Little's law is the other half, and it is where the good news starts to bend. Work in process equals throughput multiplied by time in the system. If what you can store goes up five times and what you can move goes up three times, the time an average unit spends sitting in the building goes up by five over three, which is about 1.67. If the old site carried thirty days of cover, the new one can physically carry about fifty. That is not a benefit. That is a decision you now have to make, and if you make it by accident you will have converted a space constraint into a working capital constraint and called it an upgrade.

Who gains, and by how much. The clearest winner is the shipper with a stable, deep, fast-moving catalogue and enough volume to keep the machine busy. If you ship a few hundred units a day across a few hundred well-behaved SKUs, density pays you twice: your rent per unit of stock falls to roughly a fifth of what it was, and your queue all but disappears, which is the part your customer actually feels. That shipper should be re-running the safety stock calculation this week rather than next quarter.

The mid-size cross-border seller renting space inside somebody else's overseas warehouse is in a completely different position, and this is where I would slow right down. If your contract is priced per square metre or per pallet position per month, the density gain belongs to the operator, not to you. They have just cut their cost of delivering you a pallet position by roughly four fifths, and unless the rate card changes, they keep every cent. Ask what the rate card is denominated in. If the answer is floor area, you are funding their capital expenditure. The pricing unit is a system design decision wearing a commercial costume, and it decides who captures the robot.

Category fit is the second filter. Bins reward cubic, uniform, boxed goods: consumer electronics accessories, small auto parts, beauty, hardware, anything that arrives in a carton that does not fight back. Bins punish irregular, soft, hanging or oversized goods. Garment on hanger, folded textile that will not hold a shape, anything longer than a bin, anything that arrives in a polybag that slumps: all of that goes to floor space or to a mezzanine, and the 18.36 figure does not apply to a single unit of it. The honest version of the headline is eighteen bins per square metre for goods that fit in a bin, and the share of your catalogue that fits in a bin is the first number you should compute.

The operators themselves are carrying a different risk. A gantry robot and a shuttle system are fixed cost with a long life, and fixed cost does not care that January is quiet. Their economics depend on keeping the box full through the part of the year when cross-border volume falls off a cliff. That is their problem, and it becomes your problem the moment they start pricing to protect utilisation and discounting hard in the soft months. If you are signing a twelve-month deal, ask what the rate is at half volume. The answer tells you how much of the saving you will actually see and how nervous they are about their own forecast.

When does any of this reach your P and L. Installation is the fast part; the slow part is data. A shuttle system is only as good as the dimensions, weights and inner-pack quantities in your item master, because the machine is assigning physical addresses to things it has never seen. If twenty per cent of your SKUs have no recorded dimensions, twenty per cent of your catalogue cannot be binned, and you will find this out in week three rather than week one. Then comes slotting, which is a learning problem: the system does not know which of your SKUs are fast until it has watched you for a while. I would plan for eight to twelve weeks between go-live and the point where the published pick rate is something you would bet a customer promise on. Those weeks are my assumption, not a published figure, and they belong in your project plan as a ramp, not as a switch.

On the inventory side the transmission is slower and far more useful. Changing a reorder point changes nothing until a container lands. Assume a thirty-day ocean transit from South China to Los Angeles, which is your number to check and not mine, and assume your stock turns about once a month. A reorder point you change today first shows up in landed inventory about thirty days out, in your service level about sixty days out, and fully in your working capital by about ninety. Which means, said plainly, that nothing you decide about hardware in late October changes this peak season at all. This peak is already baked. What you can still change for this peak is the parameter, not the building: how many days of cover you carry on the A items, and where that cover physically sits.

And here is the thing that is not in the headline, the one that will actually bite. Storage capacity went up by a factor of five and throughput went up by a factor of three, and nobody divided those two numbers. That ratio, 1.67, is the real output of this project, and it says the warehouse's ability to hold has grown faster than its ability to empty. Fill it to the new capacity and average dwell time rises by two thirds. You will have taken your slowest-moving inventory and given it a home in some of the most expensive industrial square metres in American logistics, sitting in a bin that a fast mover could have used.

The industry will report this as a 400 per cent win and then wonder why inventory turns never improved. Turns never improved because turns per bin is the variable, and turns per bin is not on the press release. Storage utilisation is the wrong KPI for a dense automated site, because storage utilisation is trivially easy to maximise: fill it up. The right KPI is inventory turns per bin per year, and the behaviour that follows from it is keeping the slow tail out of the expensive box.

Four more things the report does not say, each worth one question. Units per bin, which I have already flagged: ask for the internal bin dimensions and ask what that means for your top twenty SKUs by volume, answered in units and not in bins.

Then putaway rate. Every automation story quotes pick rate and almost none quotes putaway rate. A shuttle that picks three times faster but puts away at the old speed has moved the bottleneck from the pick face to the receiving dock, and your containers will queue outside instead of your orders queueing inside. Ask for putaway lines per hour alongside pick lines per hour, and hold both against your inbound container profile in your worst peak month.

Then availability. Eleven thousand and sixteen bins served by one gantry robot is a single point of failure, and the failure mode is total rather than partial. When a forklift in a conventional warehouse breaks you lose maybe ten per cent of capacity for a shift. When the gantry stops, every bin in the building is unreachable or reachable only at walking speed, and the queue that had essentially vanished comes back in an instant, worse than before because you have since promised your customers three-day delivery. Ask for the uptime commitment, ask what the degraded mode is, ask how long a critical spare takes to arrive, and ask what happens in a power event. In California that last one is not a theoretical question.

Then the shape of the cost curve. Labour cost per unit down 75 per cent is a statement about one point on a curve, and it is almost certainly the best point. Fixed cost does not shrink with volume, so at half the volume your unit labour cost is meaningfully worse than the headline, and at a third of the volume it is worse again. Ask for cost per unit at one hundred, fifty and thirty per cent of design volume and put all three in the contract. A rate card with a single number on it is a rate card designed for the seller's best month.

Now the disagreement, because there is one and it is not small. Everything above assumes storage was your binding constraint. If it was not, the entire gain evaporates. Take three plausible alternatives. If your constraint is the outbound parcel carrier's cut-off, because the trailer leaves at three in the afternoon and nothing after that ships today, then tripling pick speed moves your queue from the pick face to the car park and your customer notices nothing at all. If your constraint is inbound, because you have four dock doors and three containers arriving at once, no amount of internal speed helps.

And if your constraint is cash, because you cannot finance more inventory however cheaply it can be stored, then cheaper storage is a discount on something you were never going to buy. In all three cases the honest recommendation is to spend nothing on hardware and go and measure. Measure first, then tune. That is not a slogan, it is the only order of operations that has ever worked on a system like this.

The other way the conclusion flips is demand. The whole reading is built on the volume showing up. If it does not, the fixed cost is spread over fewer units, your cost per unit climbs back toward where it started, and your contract is still running. Cross-border demand is not famous for its stability. The shippers who get hurt by automation are rarely the ones who bought it. They are the ones who bought it against a volume forecast that belonged to somebody else's peak.

So the boundary conditions, stated plainly. This is a good deal if your goods are cubic and uniform, if your volume is high and reasonably steady, if your constraint really is storage or pick labour, if your item master is clean, and if the contract is denominated in something that lets you capture the density. It is a bad deal if your catalogue is long-tailed and slow, if your SKU count grows faster than your units, if the real constraint sits at the dock or at the carrier cut-off, if your data is a mess, or if the pricing unit hands the gain to somebody else. Move any single one of those and the sign flips.

Let me do the arithmetic properly, with the assumptions nailed down where they belong. Rent per bin. Call your monthly rent per square metre R, whatever your contract says. Before, at 3.67 bins per square metre, one bin-month cost R divided by 3.67, which is about 0.27 R. After, at 18.36 bins per square metre, one bin-month costs R divided by 18.36, which is about 0.055 R. That is a fifth of the price for the same physical thing, and it is the hardest number in this whole story, because it does not depend on the robot working well. It depends on geometry. If the four-times reading is the right one, the old figure is 0.22 R and the saving is fourfold rather than fivefold. Either way, the rent you already pay now buys four to five times the storage, and that is real money whatever happens to the machinery.

Labour per unit. I do not have your wage bill, so use an index. Call your old labour cost per unit one hundred. After a 75 per cent reduction it is twenty-five. Across a hundred thousand units a year you remove seventy-five index points on every one of them, which is the equivalent of seventy-five thousand units' worth of labour cost. Put your own currency in if you want a number: at an assumed one dollar and twenty cents of labour per unit, which is my placeholder and not a figure from the report, a hundred thousand units saves ninety thousand dollars a year and three hundred thousand units saves two hundred and seventy thousand. The point of the index is that the saving scales with volume, which is exactly why the question about cost at half volume matters more than the headline does.

Queue headroom, which nobody prices. Assume you run a thousand order lines a day at baseline and peak runs two and a half times baseline, a normal seasonal shape for a cross-border seller. Assume the old system served twelve hundred lines a day, so baseline utilisation was 0.83 and a peak demand of twenty-five hundred lines a day simply did not fit: the queue grew without limit, you hired temps and you shipped late. Tripling the service rate to thirty-six hundred lines a day puts baseline at 0.28 and peak at 0.69. Peak is now inside capacity. You have not bought three times the speed; you have bought the ability to run your peak without the scramble, and that is worth whatever you currently pay in overtime, temp labour, expedited parcel and cancelled orders. Price those four and you have the real business case.

Days of cover, which is where the density actually pays. Assume you used to hold thirty days and the new site physically allows fifty. Assume your annual carrying cost is twenty-five per cent of inventory value, covering capital, obsolescence, shrinkage and insurance, and assume the stock in question is worth five hundred thousand dollars at any moment. Twenty extra days costs five hundred thousand times twenty over three hundred and sixty-five times twenty-five per cent, which is roughly six thousand eight hundred and fifty dollars a year. Now price what that cover saves.

Assume an emergency air shipment of five hundred kilos at a premium of four dollars a kilo over sea freight, my assumption and yours to replace: two thousand dollars a run. Ten of those a year is twenty thousand dollars. The extra cover pays for itself roughly three times over, before counting a single lost sale from a stockout. Run it on your own carrying cost and your own emergency freight rate and you will have a defensible answer to how much safety stock belongs in a bin in Los Angeles.

So what do you actually do. Begin by measuring before you buy anything. Take one ordinary week and count the queue at each stage of the operation you have today: containers waiting for a door, cartons waiting to be put away, orders waiting to be picked, picked orders waiting to be packed, packed parcels waiting for the carrier. Whichever pile is largest and longest-lived is your bottleneck, and it is almost never the one in the sales deck. If the largest pile is waiting for the carrier, stop reading automation brochures and go and renegotiate a cut-off time.

If storage or pick labour really is the biggest pile, go to the operator with a specific list. Ask for internal bin dimensions and the units-per-bin figure for your top twenty SKUs. Ask for putaway lines per hour as well as pick lines per hour. Ask for the uptime commitment, the degraded mode and the critical spares lead time. Ask for cost per unit at one hundred, fifty and thirty per cent of design volume. And ask what the rate card is denominated in. Five questions, and the quality of the answers tells you more than the site tour does. A vendor who cannot answer the third one has not thought about your peak.

Change the unit you are billed in. If you are paying per square metre or per pallet position, you are paying for a constraint that no longer binds, and the operator has just been handed a four-fifths reduction in their cost of serving you. Push for cost per unit shipped, or for a rate per bin that reflects the new density, with a floor and a ceiling written in. If they will not move off floor area, that is information too, and it tells you who the automation was really built for.

Fix the item master before anything is installed. Dimensions, weight, inner-pack quantity and units per bin for one hundred per cent of active SKUs, with a hard rule that anything missing a dimension does not get received. This is unglamorous and it is the whole difference between a system that works and a system that jams. Set the target for inventory accuracy at ninety-nine and a half per cent or better and track inventory adjustments per thousand lines as a weekly number. In a manual warehouse a mis-scan is a nuisance somebody can walk over and check. In a bin-addressed warehouse the machine believes the bin, and one bad record becomes a pick from an empty location in front of a customer order.

Decide the inventory question in numbers rather than by default. If you are going to hold more stock, say how many days, on which SKUs, and with a review date attached. My bias is A items only, an extra fifteen to twenty days on the top slice of the catalogue, and a hard review at the end of the first quarter after go-live, because the entire point of a model is that you get to correct it. Let the space fill because the space exists and you will finish the year with two hundred days of cover on a long tail in a very full and very unprofitable building.

Set the thresholds now, while nobody is panicking. Agree the alert when pick lines per day push utilisation past about seventy per cent, and agree what happens at that point. Agree a review trigger when turns per bin drop below a floor you set in the contract. Agree who pays for expedited outbound when the system is down, and put a number of hours on it rather than a sentence about best efforts. Thresholds written in a calm month are the only ones that survive a busy one.

Alternatives, because automation is not the only way to buy density. Move outward. Assume industrial rent in an inland submarket runs at a third of the Los Angeles figure, which is your number to verify rather than mine; then leasing three times the floor area buys three times the storage for the same rent, with no capital expenditure, no single point of failure and no ramp curve. That gets you threefold where the shuttle gets you fourfold to fivefold, and it gets you a flexibility the shuttle cannot. Re-slot instead of rebuilding: in most overseas warehouses I have looked at, a fifth of the pick travel is caused by slotting decisions nobody has revisited in a year, and fixing that costs nothing. Or split it: fast movers in the dense automated site, slow movers in cheap bulk space an hour away. Two echelons, one catalogue, and the slow tail stops competing with your A items for the expensive bins.

And the pits, in the order people fall into them. Signing a floor-area contract after the operator has densified, which funds their machine and leaves you on the old price. Assuming the tripled pick rate applies to your order profile, when multi-line and single-line orders behave very differently on a shuttle. Promising peak-season delivery on day-one throughput and discovering the ramp in week six. Treating eighteen bins per square metre as a law of physics rather than a function of bin size and carton size. And the big one: buying storage capacity when what you needed was throughput, or buying throughput when what you needed was cash.

Everything here resolves to one sentence, and it is not about robots. The building did not get smarter. The price of a bin-month fell by roughly four fifths and the queue at the pick face got roughly thirteen times shorter, and both of those are changes you can drop into a model tonight and price against your own numbers. Do that before you sign anything. Measure first, then tune. Everything is a queue, and the only question that has ever paid is where the queue actually is.

↑ Back to the key points

— By Ravi Chandran