The Port Runs on Rented Steel

The night gang pulls the first box off the vessel just after the shift change. In the tower, the dispatcher is watching a bank of screens that all seem to be talking to each other. One is a terminal-operating system. Another is the truck-appointment queue. A third is a Coast Guard webpage the terminal has been told to keep open, in case a new advisory drops about the crane the operator is sitting inside. All three are talking to the same steel arm. Only one of them was built here.

That third window is the story the sovereign-AI conversation in Washington keeps missing. The steel is Chinese. The software on top of the steel — the terminal-operating system, the routing agents FedEx and UPS are rolling into their networks, the AI models coordinating a million-plus Amazon warehouse robots — is being deployed at software speed onto hardware that will take a decade to replace. The gap between those two clocks is the sovereignty gap that matters.

What the paperwork says

Read the paperwork in order. In April 2025 the White House issued Executive Order 14269, "Restoring America's Maritime Dominance", and on 13 February 2026 followed with America's Maritime Action Plan, threading AI, digital twins and additive manufacturing through a plan to rebuild the industrial base under the ships. The Coast Guard's final cyber rule for the Marine Transportation System went into effect on 16 July 2025; the training deadline for terminal personnel passed on 12 January 2026; the full-compliance date is 6 July 2027. On the sovereignty of the machinery itself, MARSEC Advisory 2026-007, currently in force until 21 October 2026, still names the same three vectors it has named for two years running: ZPMC ship-to-shore cranes, Nuctech scanners, and LOGINK.

The stack of documents is real work. It also does the thing documents can do, which is describe a state the operator on the pier does not yet inhabit. Roughly 80% of the ship-to-shore cranes at US ports are still built by Shanghai Zhenhua Heavy Industries, a share that has moved barely at all since the FBI first surfaced concerns at the Port of Baltimore. PACECO and Mitsui E&S restarted final assembly on the US West Coast in 2024 for the first time since 1989, and delivered a first batch of rubber-tired gantry units to the Port of Long Beach through early 2026. Real progress, and a rounding error against an installed base of several hundred large cranes across the country.

That is the hardware layer. The data layer is not much prompter. Legislation blocking federal funding to any port that uses LOGINK, the Chinese-state logistics platform, has been on the books since the fiscal 2024 NDAA. LOGINK now partners with more than twenty ports internationally, none of them in the US. But the goods US carriers move touch its network on the far side of the ocean constantly, and a Pentagon that will not use a foreign port because that port runs LOGINK still buys fuel through the ships and the routing services that do.

The software runs faster than the hardware

Put the software layer next to it. The Port of Long Beach unveiled an AI-powered terminal-optimization system this year, integrating vessel, crane and gate data with a stated ambition of trimming truck wait-times by up to 20% and terminal throughput by up to 15%. UPS has said AI and human support will together handle more than 98% of customer service requests by the end of 2026, and is pushing agentic tooling across planning, routing and execution. Amazon deployed its millionth warehouse robot alongside a foundation model it calls DeepFleet, coordinating fleet movement with a claimed 10% improvement in robot travel-time.

Treat those vendor numbers with the caution they deserve. Ports and carriers, like the labs, publish the flattering measurement. A 20% wait-time reduction at Long Beach and a 10% robot-travel gain at Amazon are the intended results of the systems, not the audited results of them. What is not in dispute is the pace. AI is being pushed into US logistics on a two-to-three-year clock, at latitude the Coast Guard cyber rule gives operators through 2027, while the substrate underneath (cranes, scanners, logistics data plumbing) is on a ten-to-fifteen-year replacement clock at best.

This is the mismatch the sovereignty conversation elides. A US-built terminal-optimization model, running on a Chinese-built crane, telling a Chinese-built scanner what to inspect, and reporting throughput to a US-flagged carrier whose overseas transit lands at a LOGINK-connected partner port on the other end, is a sovereign system in perhaps one of those five places.

The operator's decision

The person the paperwork lands on is not a policy staffer. It is a terminal manager and, one desk over, a shop steward. The International Longshoremen's Association's 2025 contract explicitly bans fully autonomous cranes at East and Gulf Coast ports through 2031 and requires a new hire for every semi-automated one. The ILWU's standing policy on the West Coast opposes AI used to surveil, discipline or displace workers, with contract talks due in 2028.

Set that against the AI programs above. A terminal-operating system that promises to trim throughput is, by definition, a system that reallocates human effort. The union positions do not stop that; they slow it, price it, and give it a face at the negotiating table. It is a legitimate constraint, and it is one the vendor deck for a terminal-AI product usually leaves out.

The most cited counter to the union view is the AJOT piece asking whether the ILA is wrong about automation, which points out that paid ILWU hours at the two fully automated Los Angeles/Long Beach terminals grew 31.5% between 2015 and 2022, more than twice the 13.9% at the non-automated ones. It is a fair number and it deserves an answer. Two of them, in fact. First, the base was smaller and the comparison covers a period of exceptional container growth in the San Pedro Bay; growth on a small base is not the same argument as safety of the workforce elsewhere. Second, the mix of jobs shifted. Total hours can rise while the specific longshore role that automation most directly displaces shrinks, and the residual work moves toward remote-operator seats that are counted the same way in the payroll but paid and staffed on very different terms. The union's fear is not "fewer jobs" as a scalar. It is a change in the composition of the jobs the contract was written to protect. That distinction survives the chart.

What breaks first

Sit in the tower long enough and you learn where a system fails. The sovereign-AI plan for US logistics fails, in the order it will fail, at three points.

It fails first at the labor negotiation, because the ILA contract runs to 2031 and any AI plan that quietly assumes a smaller longshore workforce before then is planning a strike into its schedule. It fails second at the hardware replacement clock, because ZPMC has told US port operators plainly that a punitive crane tariff without a viable domestic alternative would idle terminals, and however self-serving that warning is, the arithmetic is on their side; no US manufacturer is ready to fill the pipeline this decade. It fails third at the data layer, because a rule that keeps US ports off LOGINK does very little about the twenty-plus foreign ports where a US carrier's freight actually lands and gets routed on the other end.

None of that is an argument against the direction of the sovereign-AI plan. It is an argument for reading its calendar honestly. The software half of the sovereignty project is moving at the pace of an enterprise agent rollout. The hardware and data halves are moving at the pace of a shipyard and a treaty. The dispatcher in the tower runs the interface between those clocks every night, and does not have the option of pretending they are the same.

What stays true when this cycle ends

The night I described happens in some form every night at every major US container terminal, and it has been happening in some form since the shipping container replaced break-bulk sixty years ago. The people who dispatch cargo have always worked with equipment they did not build, on software written somewhere else, under rules written later still. What is different now is that the equipment, the software and the rules are being pulled in three directions by three governments at once, and it is the dispatcher, at eleven at night, who has to hold all three while the box comes down. Sovereignty in that room is a slow, unfinished thing. It has been that way for a long time, and it will probably stay that way through this cycle too.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.