Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Ask for the baseline behind each of the three figures before modelling any saving: planning hours before the agent, gate transactions per hour before the gates, seconds per box before the tally model. No baseline, no business case, and raise it at your next service review.
- Replace average throughput with the 95th percentile truck turn time in your own KPI set and track it monthly for two quarters, because a gate that emptied its queue will look excellent on average and fail on the tail.
- Measure your own arrival variance at the gate for four weeks. If the standard deviation is above two hours, fix booking discipline before crediting the terminal, since you are the constraint in that case.
- Test a six-hour cut in your documentation cut-off on one string for one month, with a rollback trigger at two missed sailings, to learn whether the terminal's re-planning speed actually reaches your cargo.
- Log the gap between the berth window you were given and the one actually used for thirty days. A median gap above ninety minutes means the plan is still behaving like the old two-hour plan.
Shandong Port Group has launched the Xingyu SA5200, billed as China's first AI chip designed for port operations, alongside 26 priority deployment scenarios. At Qingdao's Qianwan terminals a berth-allocation agent has cut planning from one to two hours down to minutes, smart gates lifted vehicle throughput 60% without new lanes, and Rizhao Port's tally model trimmed per-box time 6%.
The Analysis
Three numbers came out of Shandong Port Group on 6 August 2026, and they are not the same kind of number. A berth-allocation agent at Qingdao's Qianwan terminals cut planning from one to two hours down to minutes. Smart gates lifted vehicle throughput 60 percent without adding lanes. A tally model at Rizhao cut per-box time 6 percent. Alongside them sits a chip, the Xingyu SA5200, described as the first AI chip designed for port operations, and 26 priority scenarios.
A chip is a component. A terminal is a set of queues that touch each other. The question worth asking is which of those three figures changes capacity, and which one only changes how quickly you discover that capacity is fixed.
I have spent enough time around gates and yards to distrust a percentage without a denominator. One of these three does not have one.
Take the terminal apart. Vessels queue for berths. Boxes queue for quay cranes and then for yard slots. Trucks queue at the gate. Each of those is a service rate against an arrival rate, and the terminal's capacity on any given day is set by whichever one binds first, not by the average of the three. Improve a queue that is not binding and you get a better dashboard. Improve one that is binding and you move the constraint to the next queue, where it sits until somebody improves that one too. This is why port efficiency programmes so often deliver a good first year and a flat second one.
Berth allocation is not a physical process, it is a commitment process. The plan is a promise made at once to a carrier, a crane crew, a yard block and a trucking base, and it is made against a window. At one to two hours of planning time you plan once and freeze, because re-planning inside the day costs more than the disruption it would repair. That freeze is the buffer. It is not slack in the bad sense, it is the thing that lets every downstream party commit resources without calling you every twenty minutes.
Compress planning to minutes and the buffer becomes optional. The terminal can absorb a late feeder or a delayed discharge inside the same shift instead of carrying the cost into tomorrow. That is a real gain, and it is a gain in responsiveness rather than in berth capacity. Berth metres did not increase. What increased is the terminal's ability to use the metres it has against a moving reality.
Which is where the cost moves. A plan that can change in minutes will change in minutes, and every party landside books against a window: a trucking company books a driver for a shift, a rail operator books a block train for a week, a feeder books a slot. When the window moves after they have booked, they absorb the variance, and their cost of absorbing it is a function of how long their commitment runs. The shortest commitment absorbs the least. A drayage operator can re-sequence a morning. A rail block cannot. A shorter planning cycle therefore pushes flexibility risk down the chain to whoever has the longest lock-in, and that party is usually not the terminal.
Now the gate. Sixty percent more vehicle throughput without new lanes is a service-rate result, and service rate against a fixed arrival rate does not raise steady-state throughput by sixty percent. It empties the queue. Throughput in steady state is capped by arrivals; what falls is waiting time. That is genuinely valuable, and it also means the sixty percent measures how long the queue used to be, not what the gate will do next year.
A queue is a buffer. Take it away and the variance it was absorbing does not vanish, it relocates: onto the approach road, into the appointment book, or upstream into the yard, which is now the next constraint. A gate with no queue is efficient and brittle. One lane closure, one system outage, one hour of weather, and there is nothing left to absorb the shock. The honest metric here is not throughput. It is the 95th percentile turn time, and that number is not in the release.
Six percent of what, exactly. The release gives a relative improvement on per-box tally time and not the base. If a tally takes ninety seconds, six percent is five and a half seconds. If the terminal handles five million boxes a year, that is about 7,500 labour hours, which is a real number with a budget line attached. Both of those inputs are my assumptions, the announcement supplies only the percentage. This is the arithmetic that decides whether the figure matters, and it is exactly the arithmetic handed to the reader.
What the coverage leaves out is that the three figures are not additive, while the release invites you to add them. One is a latency gain on a decision. One is a throughput gain on a physical process. One is a duration gain on a labour task. Only the second changes capacity, and even that one changes it by removing a queue rather than by adding a lane. The first changes responsiveness, which converts into capacity only if planning latency was the binding constraint. The third changes cost per box, which is a different budget entirely. Stack them into one story about a smarter port and you hold nothing you can plan against.
The other thing nobody says out loud is what a dedicated chip implies about the operating model. A port group designing its own silicon is a port group that has decided inference latency and the data path are worth owning rather than renting. Whether the chip is good is not the interesting question and I have no way to judge it. What matters is what happens to the integration burden once compute is in-house: it stops being the vendor's roadmap and becomes the terminal's own engineering capacity. That is a slower and more durable kind of advantage than a purchase, and it is also a bottleneck that relocates inside the organisation.
The read flips on one unknown: berth occupancy. At seventy percent, planning latency is a rounding error and the minutes saved buy comfort. At ninety-five percent, latency is the difference between a recovery and a cascade, and minutes are the most valuable thing in the building. The release does not give occupancy. Ask for it before you build a business case on the berth agent, because the same announcement means two completely different things at those two readings.
The trap for a shipper is assuming a faster terminal shortens door-to-door. It does not. It means the terminal can absorb more of your lateness, which is a different product, and it is only worth something if you stop producing lateness. Most operations consume new slack within a quarter and land back on the old variance, because the variance was theirs all along. Set the threshold before you negotiate: if your own gate-arrival standard deviation is above two hours, the terminal's new flexibility never reaches your P&L. Fix that first, then talk about the port.
↑ Back to the key points