The Latest from AI at Sea

Weekly roundup: Schneider Electric, NVIDIA and HPE build AI agents that read alarms and maintenance history on site without the cloud, while Siemens cuts inspection set-up time by 40% at one plant, which is an engine room in everything but name; Mistral and NVIDIA bring AI to engineering design for cars, aircraft and chips as HD Hyundai and Kiewit plan a digital shipyard in the US, with no ship shown yet; Anthropic's draft stock market filing shows $16.3bn of revenue in six months and about $450bn promised to cloud companies, which says something about the price of the AI you rent; AI finds two holes that sat in PHP for about twenty years and governments take websites offline to patch them; South Korea plans a permanent test area for self-driving ships off Geoje before Samsung's temporary permit runs out; China puts fishing grounds on the electronic charts of more than 16,500 vessels; EuroAfrica moves Syroco's AI voyage planning to its whole fleet of four ships and keeps its own savings number private; and we ask what carries over when tools built for factories and offices come aboard

Weekly roundup: CMA CGM's AI phone agents reach more than 4,000 customers and close 46% of cargo release calls on their own, with the supplier part-owned by the carrier and no error rate given; a Boston Whaler now docks itself with six cameras and hands control back when it is not sure, with no price, no test data and no word on who is liable; the EU sends its first AI Act requests after agents from two AI labs got out of their test environments; Korea puts about $1bn and Fincantieri almost €100m into AI for shipyards, with no targets yet; nine shipping tech bosses say what they would never buy again; an independent Dutch lab tests a self-navigation system on 55 DNV traffic scenarios and nobody publishes the scores; ABS says test smart glasses task by task, after a study found people wearing one headset hit their heads about twice as often; and we ask who is in charge of the AI on board when the satellite link drops

Weekly roundup: ABS runs a port approach 100 times on a computer and finds that taking one person off a four-person bridge pushes the officer on watch into overload in 61% of runs instead of 8%, in a simulation and not on a real ship; a Busan container terminal puts 28 AI cameras on 14 quay cranes to warn machine drivers when a worker on foot is in their lane, with no accident figures yet; Lloyd's List Intelligence adds satellite pictures to check whether a ship was really at the position it broadcast, with no customer or price given; Singapore plans to register a ship in minutes instead of three days from 2027 and puts AI, autonomous vessels and drones into a new law in the same week; camera maker ShipIn says Prime Tanker Management recorded 76% fewer safety gear breaches in 60 days, in a release that does not explain how they were counted; China's national standard for naming and grading ship data takes effect, written by the yards and owners that will use it; and we ask the question that sits under all six, which is whether the person next to the machine can still cope

Weekly roundup: ship manager Acheon Akti publishes its own number for camera AI, with bridge near misses down almost 20 times on one pilot ship and no word on who did the counting; a Marcura and Thetius survey finds 63% of maritime staff use AI every day while only 8% say their company governs it, from a sample of 60 people; Flexport lets outside AI agents book freight at the customer's own rate without logging in, and says every exception goes to a human; Sea launches a charter party assistant in the same week chartering chiefs in Singapore ask who trains the juniors when AI does the basic work; Ulysses and GTMaritime both go after the email inbox, a market our own team is also in; PSA and Granite Asia put $50m into port technology while the Netherlands backs 13 projects with more than €20m; and we write down the one list most companies still do not have, which is what an AI agent may do without asking

Weekly roundup: Huawei turns its work at Shandong and Tianjin into port AI products it can sell, claiming 90% of machine-made plans are accepted and under 0.1% intervention on driverless trucks, figures nobody outside the company has checked; four Black Sea tankers broadcast positions from central Lima, and the startup that caught it argues normal data cleaning deletes the evidence of spoofing; CYTUR finds a system with 200 known software flaws scored almost the same risk as one with 20, from a small sample it openly calls unrepresentative; Germany starts a second phase of AI support for vessel traffic controllers, with a demonstrator due in 2028; a training company chief asks the IMO to make AI literacy an STCW competence before the review closes in 2029 to 2030; WiseTech offers a free four-week freight capacity forecast without an accuracy record; and the Nautical Institute rewrites the evidence guide that once said electronic records cannot lie, in the same week a P&I club put a second position source on offer to its members

Weekly roundup: Shipping Technology gives its inland track pilot control of the engine as well as the rudder, on the back of 200 vessels and a 5% fuel saving that a Dutch government agency checked in the field, while the new 10% claim is still the vendor's own; Galileo produces the first civil satellite position protected against spoofing in both its data and its signal, with the service due in 2027 and free worldwide; Yilport reports 21% more crane moves per hour at Gebze after three months of Kaleris optimisation with dispatchers still in charge; India's JNPA awards a $9.7m AI digital twin contract only after a live proof of concept on its own port; TS Cyanergy calls an ABS approval a landmark validation when the certificate itself says non-essential; China Merchants launches a vessel operating system from chips to applications; and Anthropic names its first outside AI inspector and pays for it itself, which is the question shipping has lived with since the first class society sent the owner an invoice

Weekly roundup: three separate bodies certified maritime AI software in six days and every one of them stopped short of the number that matters, with Korean Register reviewing ShipWatch's calculation method but not its fuel-saving claims, Normec Verifavia certifying 90POE across six emissions regimes while leaving data quality with the owner, and Bureau Veritas approving Inmarsat's NexusWave for cyber; CMA CGM's head of ship management describes taking Orca AI watchkeeping from trial to 100 ships at Seaspan and reveals the data changed conversations with pilots, not just captains; Nauticus sells ROV autonomy as a retrofit that keeps the pilot in charge and claims over 20% efficiency; Sea puts BIMCO's standard clauses inside the fixture workflow; Hanwha Ocean pays MaumAI KRW2.1bn for autonomous welding; the Cannes navigation prize goes to assisted position-keeping rather than AI perception; South Korea forms a taskforce to write its autonomy technology into the international rulebook; and the people building the AI models publicly asked their own industry to slow down, in the same week shipping certified three speedometers

Weekly roundup: CMA CGM starts designing an AI-assisted containership with SDARI and Bureau Veritas and puts the real question in writing, which is how much control stays with the crew; ShipIn's chief executive answers a serving master mariner in print and publishes fleet numbers for the first time, then converts the company to a Public Benefit Corporation with a legal duty to report; Marcura launches a procure-to-pay platform where AI reads the invoice and matches it against the order; Fugro builds an AI marine survey hub in Singapore with government money and claims some analysis runs up to ten times faster; Kaleris buys Newport Systems to automate container repair approvals across 25 data links built over twenty years; GTT Marine turns the voyage data recorder from accident evidence into a weekly bridge monitor; Cydome ships a tool for tracking your suppliers' cyber certificates under IACS UR E27; and we count what shipping actually launched at SMM Hamburg and find plenty of products, almost no published results, and one question nobody wants to answer about who is paying for the AI accounts already in use

Weekly roundup: a serving Maersk master mariner argues in print that AI cameras detect behaviour but cannot measure safety culture, invoking Goodhart's law against a category now running on more than 1,300 vessels without published incident-reduction data; ClassNK puts a natural-language assistant called Survey Compass over its own survey rules and becomes the user of AI rather than the auditor of it; Anschütz integrates SEA.AI machine vision into its AUTONOMICS suite and adds AUTONOMICS CGA, which calculates COLREG-compliant avoidance trajectories and hands heading-only manoeuvres straight to the autopilot; DBMatic and Weforsea Shipping merge into MariControl in Urk to sell hardware, connectivity, software and AI as one stack across shipping, offshore, inland and fishing; Østensjø Rederi expands Dipai's fleet platform across its offshore vessels and tugs five months after onboarding the first ship; Hexagon buys Guidance Marine and its roughly €22m of relative-positioning sensor revenue into the NovAtel division; ESVAGT fits NextOcean's WavePredictor to an SOV at Dudgeon so the officer can see what the vessel will do in the next few minutes before making a transfer call; and we argue the eyes got installed this week while nobody agreed what they are for — the signal is whose decision the sensor changes, the officer's or the office's