Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Pick the single lane with the highest repeatable volume and log two weeks of real run data before you talk to any vendor: if the same origin-destination pair does not move at least three full loads a day, week in and week out, there is nothing for autonomy to optimise.
- Measure the two ends of the lane before the middle: count minutes lost at fuel or charge, at load securement, at the far dock, and at the exception queue, and name one owner for each, because that is where your cost actually sits.
- Reclassify the driver role on that lane in writing within 30 days: list the tasks that stay human, assign them, and set a target of moving at least 60% of a driver seat's hours off the truck before you commit to any deployment date.
- Model the lane as a fixed cost, not a variable one: build a one-page sheet that shows the annual capital, maintenance and software charge against your committed volume, and pull the trigger only if the lane runs above 80% fill.
- Set a review gate at the next annual service terms: benchmark your quoted cut-off, dwell and exception-response numbers against the baseline you logged in week one, and renegotiate the service level before any price step-up lands.
PepsiCo signed a multi-year agreement with autonomous trucking company Gatik to deploy driverless freight vehicles across its North American supply chain, marking the largest commercial autonomous freight deployment to date. Gatik — already operating fully driverless trucks for Fortune 50 retailers in Texas — will now expand to PepsiCo's distribution network, covering middle-mile routes between distribution centers. The partnership validates autonomous trucking's transition from pilot programs to scaled commercial operations, with implications for driver utilization, fuel efficiency through optimized driving patterns, and 24/7 freight movement capability. The deal reinforces the view that autonomous trucking's first commercial frontier is structured middle-mile routes rather than long-haul highway driving.
The Analysis
PepsiCo signed a multi-year agreement with Gatik to run driverless freight trucks across its North American supply chain, covering the middle-mile legs between distribution centres. Gatik has already been running fully driverless trucks for Fortune 50 retailers in Texas, and this expansion is being called the largest commercial autonomous freight deployment to date. I do not read this as a technology story. I read it as a queue story.
Picture the supply chain as a line of queues. Trucks queue at docks, trailers queue for loads, loads queue for paperwork. Middle-mile is the one queue where the same job repeats from the same point A to the same point B, at volume, every single day. That is why the first commercial frontier turned out to be structured middle-mile routes and not long-haul highway driving, and that is the part worth taking apart.
So where is the bottleneck? Not where the headline points. The news is the driving. The constraint sits at the two ends of the lane, and that is where I would look before I believe any throughput promise.
Start with what is actually claimed, because the numbers here are deliberately thin. One agreement, multi-year, between a large shipper and an autonomous trucking firm. One route class: middle-mile, running between distribution centres. One existing operating record: fully driverless runs for Fortune 50 retailers in Texas. Three claimed consequences: driver utilisation shifts, fuel efficiency improves through optimised driving patterns, and freight can move around the clock. Now notice what is not claimed. No route count, no fleet size, no cost per mile, no delivery window. When a deal is billed as the largest to date but ships without unit economics, read the size claim as a statement about ambition, not about savings. The size is a press fact. The economics are still unproven.
The forces that made this land date back further than the announcement. Gatik did not invent self-driving this year; it spent years on one narrow, repeatable lane type until the exception rate fell low enough that one remote operator could supervise a fleet rather than a single truck. Sensor and compute costs dropped until the commercial case stopped looking like a research budget and started looking like a capex line. Driver wages and the driver shortage pushed the payback window closed at exactly the moment the technology became boring. Boring is the signal. The moment a technology stops being a headline and becomes a line item, procurement picks it up.
Who feels it first. Operators with dense, stable middle-mile networks: big consumer goods, big grocery, big retail, the ones moving full truckloads between their own or contracted DCs on a fixed cadence. For them the lane is already a closed loop with a known route geometry, a known dock and a known average speed. Nothing about the route changes, which is precisely what makes it automatable. Importers and exporters feel it second and only indirectly, through the service level they are quoted: a promise on cut-off, on dwell, on how fast an exception gets resolved. Who barely notices: anyone shipping irregular, low-volume, multi-stop or appointment-sensitive freight. If your load needs a human to negotiate with a receiver, the driver is the least of your problems.
Timing runs on quarters, not weeks. A shipper converts one lane, validates the exception rate, then adds the next; that is an internal engineering calendar you will never see. If you are a customer of a shipper deploying this, you do not meet the trucks. You meet the service promise, and service promises move on the pricing calendar, which means the next annual review or the peak-season capacity conversation. Call it four to eight quarters before any of this shows up in the terms you are quoted, and that is the optimistic read.
Here is the sentence the headline skips, and there are two of them. The first: autonomy converts a variable cost into a fixed cost. A driver is paid for the hours worked. An autonomous tractor is capital plus a maintenance and software contract, and it costs money whether or not the trailer is full. That is wonderful on a dense lane and brutal on a thin one, because you now carry capacity you may not fill. The lane you can automate is the lane you can keep loaded. Volume, not software, is the gating input. The second thing the headline skips: the route geometry quietly becomes an asset. When the same origin-destination pair runs thousands of times, its map, its edge cases and its dock choreography get tuned until the exception rate approaches zero. Whoever owns that tuned loop owns a cost advantage a competitor cannot copy just by buying the same trucks. The real moat in middle-mile autonomy is data about your own route, not the vehicle.
Put a small number on it, with the assumptions said out loud. Assume one middle-mile lane, a 300-mile round trip between two DCs, and steady demand of three round trips a day. Under driver hours-of-service rules, one driver seat realistically covers about one round trip per shift, so call it roughly 250 round trips a year per seat. Three trips a day across 250 working days is 750 round trips, so the lane needs three seats today. A 24/7 autonomous tractor, assuming 85% availability and the same average speed, covers about 2.5 round trips a day, roughly 620 a year, so two tractors clear the lane with headroom. If a fully loaded driver seat costs about 95,000 dollars a year, removing one seat saves roughly 95,000 dollars a year on that single lane.
Now add fuel: assume optimised driving patterns return 5% on fuel, and a truck doing 750 round trips of 300 miles at 6.5 miles per gallon and 3.80 dollars a gallon burns about 131,000 dollars of diesel a year, so 5% is roughly 6,500 dollars. The total is not the point. The shape is. The driver-seat term dwarfs the fuel term by an order of magnitude, and both scale with how full the lane runs. Get the lane utilisation right before you touch the driving.
Reverse one assumption and the conclusion flips. Make the same lane irregular, spot freight, drop-and-hook at an uncontrolled yard, multi-stop with fixed appointment windows, and the autonomous case collapses, because the exception rate sets the cost, not the driving. The condition to watch is not the technology stack. It is the variance of your route. A lane boring enough to schedule a month ahead can run without a driver. A lane that changes every week cannot, and no amount of sensor fidelity fixes a route nobody can predict.
↑ Back to the key points