Most of the useful AI news this week did not come from shipping. It came from a paper mill in Chile, an electronics plant in Germany, a design office in Paris, the accounts of an AI company getting ready to sell shares, and a piece of web software that runs most of the internet. That is normal. Shipping rarely invents its own tools. It buys what factories and offices already use, a few years later. So the useful question is not "is this maritime?" It is "what happens to this tool when it goes to sea?" A factory has an IT team on site, steady power and a fast line. A ship has twenty people, no IT department and a link that drops. Read the stories below with that difference in mind. We finish with five questions to ask before any of these tools comes aboard.

The week's most important developments in shipping & oceans — distilled into a 5-minute read.

🔗 Quick Links

• 🏭 Factory AI agents read alarms and maintenance history on site, without the cloud — Automation News

  Schneider Electric, NVIDIA and HPE are building them. Siemens reports 40% less inspection set-up time at one plant. Swap "factory" for "engine room" and the idea fits a ship

• 📐 Mistral and NVIDIA bring AI to engineering design, and HD Hyundai plans a digital yard in the US — Mistral

  Built for cars, aircraft and chips. No ship, no customer and no results shown yet. CMA CGM is already a Mistral customer on the office side

• 💰 Anthropic's draft filing: $16.3bn revenue in six months, about $450bn promised to cloud companies — VKTR

  The AI you rent sits on very large fixed bets and a shortage of computing power. That is a reason to read your AI contracts for price and capacity terms

• 🩹 AI finds two holes that sat in PHP for about twenty years — TechCrunch

  PHP runs about 70% of websites. Government sites in several countries went offline to patch. No attacks reported. Ships run old software longer than anyone

• 🇰🇷 South Korea plans a permanent test area for self-driving ships off Geoje — Splash247

  Samsung's temporary permit ends in December. Approval expected by year end. No size, budget or results published

• 🇨🇳 China puts fishing grounds on the electronic charts of 16,500+ vessels — Splash247

  About 4,500 merchant and ro-ro ships and 12,000 fishing boats now see the same risk areas. Not AI. No collision figures yet

• ⛽ EuroAfrica moves Syroco's AI voyage planning to its whole fleet of four ships — Splash247

  The operator says it saves fuel but does not say how much. The 10% figure is the supplier's own average

• ⚖️ AiatSea: tools built for factories and offices are coming aboard. What carries over? — AI at Sea

  A factory has an IT team, steady power and a fast line. A ship has none of the three. Five questions to ask before you bring any of this week's tools to sea

🚀 Big Moves This Week

1. Factory AI Agents That Read Alarms and Maintenance History, on Site and Without the Cloud

Two announcements this week show where industrial AI is going. Read them with your engine room in mind.

The first. On 22 October, Schneider Electric, NVIDIA and HPE said they are building AI agents for process plants. An agent here means a program that watches the plant, decides something is wrong and suggests what to do. One type of agent combines equipment settings, plant layout, alarms and maintenance history to spot problems early and judge how serious they are. A second type looks for unusual traffic on the control network. A third looks for ways to use less energy. The important detail: the agents run on site, on HPE computers in the plant, using NVIDIA's openly available AI models. They do not depend on a link to a distant data centre. The companies say people keep control of critical decisions. The one named user is CMPC, a Chilean paper and packaging maker that has used Schneider's control system monitoring for three years.

The second. NVIDIA opened early access to a set of tools for building factory agents, and Siemens tested them at its electronics plant in Erlangen, Germany. Siemens reports that setting up its camera inspection took 40% less time. Other figures, such as a 20% cost saving, are estimates and targets, not results.

What is missing. No price, no launch date for the Schneider agents, and no savings figure. The announcement says the agents can "recommend or initiate" corrective action, but it does not explain who decides which of the two happens.

Why a superintendent should care. Look at the ingredients: alarms, equipment settings, maintenance history. Every modern ship has all three, in the alarm system and the maintenance system. And the design choice that matters most for a ship has been made here for a different reason: the AI runs where the machines are, not in the cloud.

What to ask your suppliers. When someone offers AI for your engine room, ask three things. Does it run on board when the link is down? Can it only advise, or can it act, and who sets that line? And which of your alarm and maintenance data does it need, in what format? A supplier that has thought about ships will answer the first question before you finish asking it. [Automation News · NVIDIA · Siemens]

2. AI for Engineering Design Arrives From Outside Shipping. Nobody Has Shown a Ship Yet.

On 21 October, the French AI company Mistral and NVIDIA announced work on AI for engineering design. The goal is to speed up the slowest part of engineering: simulation, the computer tests of strength, heat and flow that run before anything is built. They also plan public tests that measure how well AI agents can use design software, check their own work and keep the computing cost down. The examples are aircraft, cars and computer chips.

The background. In May, Mistral launched a product for industrial engineering with Airbus and BMW as partners, and its biggest shareholder, the chip machine maker ASML, said a joint tool ran 120 times faster at similar accuracy. That figure was shown at Mistral's own event and nobody outside has checked it. Mistral also has a new large model, Large 4, in public preview, and says it will publish the model itself for anyone to run by the end of October. Mistral's link to shipping is on the office side: CMA CGM signed a five-year, €100m deal with Mistral in 2025, mainly for customer service.

A shipyard in the same week. HD Hyundai and the US builder Kiewit plan a new joint shipyard on the US Gulf Coast, using HD Hyundai's digital twins, automated production and robot welding. A digital twin is a computer copy of the yard or ship that is kept up to date with real data. Building should start in 2027, with full operation in 2029 if all goes to plan, and up to 3,000 jobs. No investment figure, no capacity and no targets are given.

What is missing. Neither story shows a ship. Mistral names no customer for this week's announcement and gives no results.

Why it still matters to owners. A car model is built a million times. A ship design is built four or eight times. Design AI learns from data, and shipping has far less of it per design. So the first useful jobs are likely to be the boring ones: checking drawings against class rules, comparing hull options faster, finding clashes between pipes and cables before steel is cut. Last week Korea and Fincantieri put money into AI in their yards. The design office is the next step.

What to ask. At your next newbuilding meeting, ask the designer and the yard whether AI is used in the design, for which jobs, and who checks its results before class sees them. The honest answer today is probably "a little, carefully." That is fine. "Not at all" and "everywhere" are both answers to look into. [Mistral · VentureBeat · Digital Ship · HD Hyundai · AiatSea, 18 October]

3. The AI You Rent Sits on Very Large Bets. Anthropic's Filing Shows How Large.

Anthropic, the company behind the Claude AI models, is preparing to sell shares on the stock market. Its draft filing, seen by the Wall Street Journal and reported this week, gives the first detailed look inside one of the companies whose AI many office workers now use every day.

The numbers, as reported. Revenue of about $16.3bn in the first half of 2026, against about $1.1bn a year earlier. A net loss of $80.7bn, of which about $72bn is an accounting charge linked to the rising value of its own financing, not cash spent. Spending on computing power of nearly $9bn in six months. And about $450bn promised to Google, Amazon, Microsoft and others for cloud computing over the next seven to ten years, most of it payable whether the capacity is used or not. With a chip lease and deals still being negotiated, reports put the total much higher.

Two details matter more than the big numbers. First, the company says its cloud partners were not willing to provide all the capacity it wanted, so it is leasing data centres, some from operators with a short track record. Second, a large share of revenue comes from a few big customers who are not on long contracts.

Our caveats. This is a draft. The figures come through press reports, not a final public filing, and different outlets give different totals for the cloud commitments. Revenue in the first half was higher than the computing bill, so this does not show that AI is sold below cost.

Why a ship manager should care. For ten weeks we have asked who pays for the AI accounts already used in shipping offices. This filing shows what sits behind those accounts: very large fixed costs, a shortage of computing power, and a business still changing fast. Prices, usage limits and plans can all move. Last week we reported that Accenture capped its own AI spending at $29 per person per month after some staff used much more.

What to do. Know which AI services your staff use and what you pay for each. In any AI contract, look for three things: how the price can change, what happens when the supplier is short of capacity, and how you take your data out. And avoid building a key process on one supplier only.

Disclosure: this newsletter is produced with the help of Claude, an AI model built by Anthropic, the company in this story. We have tried to hold it to the same standard as every other company in this issue. [VKTR · AI Weekly · AiatSea, 18 October]

4. AI Finds Two Holes That Sat in Web Software for About Twenty Years

PHP is the software behind roughly 70% of websites, including most sites built with WordPress. On 21 October, its developers released fixes for two critical security holes that had been in the code for about two decades. The US security agencies CISA and NSA told everyone to update at once, or to take a site offline until they could.

Who found them. AI did, with people guiding it. A security firm called Calif found one hole using an OpenAI model, as part of an OpenAI programme that searches widely used free software for flaws. Wiz, a security company owned by Google, found holes using its own AI agent and AI models from Google and OpenAI. Both worked with the PHP team before going public. OpenAI says the holes were found before any attacker found them.

What happened next. Government services in the Netherlands, Finland and France, among others, went offline for a while as a precaution while they patched. No attacks using the holes had been reported. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Anything on version 8.1 or older no longer gets fixes and needs an upgrade.

Why this matters beyond one bug. Holes like these have always existed. What changed is speed. AI can now search old code for weak spots much faster than people can. Wiz also warned that AI makes it easier to turn a known hole into a working attack. The same speed is available to defenders and attackers.

Why it matters at sea. Ships keep software longer than almost anyone. Many devices on board have a small web page for their settings: satellite terminals, routers, cameras, tank and engine monitoring, printers. Some of those pages run on old web software, and few are ever updated. On shore, crew portals, supplier portals and company websites often run on PHP.

What to do this week. Ask your IT team or provider whether any company website or portal runs PHP, and which version. Ask your main shipboard suppliers whether their products use PHP or similar web software, and how they deliver fixes to ships. And make sure nothing on board can be reached directly from the internet without a reason that someone has written down. [TechCrunch · The National · Science Times]

5. South Korea Wants a Permanent Test Area for Self-Driving Ships

South Gyeongsang province in South Korea has asked the Ministry of Oceans and Fisheries to create a permanent test area for autonomous ships, in the waters linking Hansan, Gajo and Chilcheon islands off Geoje. Approval is expected by the end of 2026.

Why now. Samsung Heavy Industries already tests there under a temporary permission granted in 2024. That permission ends at the end of December. Its test boat is small: 24 tonnes, 12 metres, with automated engine and steering control, route following, collision avoidance and an AI system that combines cameras and other sensors to understand what is around it. It cost about KRW2bn, roughly $1.4m, to develop. Separately, Korea is spending KRW30bn on an AI data platform for future autonomous ships.

Who was asked. The coastguard, fishermen's representatives and local authorities were consulted, and officials say the zone will not get in the way of fishing, leisure or commercial traffic.

What is missing. No size for the area, no budget, no rules for speed or weather, and no test results from the years Samsung has already sailed there.

Why it matters. Temporary permits are how experiments start. Permanent test areas are how an industry plans. A yard can only promise an autonomy product to owners if it knows where and how it can keep testing. Korea has also set up a national team to shape the international rules for autonomous ships, so the test area and the rulebook are moving together.

What to watch, and ask. A test area produces data. Last week an outside lab tested a navigation AI on 55 scenarios and nobody published the scores. If Korea opens a permanent zone, owners and class should ask for the results from it: how many hours, how many encounters with other boats, and how often a person had to take over. [Splash247 · AiatSea, 18 October]

6. China Puts the Fishing Grounds on 16,500 Vessels' Charts

Anyone who has kept a night watch off the Chinese coast knows the problem: dozens of fishing boats, nets you cannot see, and boats that move in ways a merchant officer does not expect. China's Ministry of Transport says it now has a fix on more than 16,500 vessels.

What it is. A special electronic chart that adds fishing information to the normal navigation chart. It shows coastal shipping routes, fishing grounds and the areas where merchant ships and fishing boats are most likely to collide. It was developed by the ministry's East China Sea Navigation Support Centre. The data includes more than 20,000 fishing-related items, 38 high-risk warning areas and 359 shipping routes. Updates come through China's official chart service.

Who has it. More than 4,500 merchant and passenger ro-ro ships and more than 12,000 fishing boats, in provinces including Shanghai, Liaoning, Shandong and Guangdong. It started on two fishing boats in December 2025.

Be clear about one thing. This is not AI. It is digital navigation used on a very large scale. We include it because it solves a real bridge problem, and because it is the kind of data any future AI collision avoidance will need. A system that does not know where the fishing grounds are cannot plan around them.

What is missing. No collision or near miss figures, before or after. No word on cost, or on whether ships must use it. China says it will propose parts of it to international organisations, so it could reach charts used by everyone.

Why the design is smart. It gives both sides the same picture. The merchant officer sees the fishing grounds, and the fishing skipper sees the shipping routes. Most collision tools only help one side.

For managers with ships trading to China: ask your chart provider whether this information is available outside Chinese systems, and brief officers on the high-risk areas before the voyage. And watch whether China publishes collision figures, because that is the number that would justify the rest of the world copying it. [Splash247]

7. A Customer Moves AI Voyage Planning to Its Whole Fleet, and Keeps the Number Private

EuroAfrica, a liner operator registered in Cyprus, has moved Syroco Live to its whole fleet. The fleet is four multipurpose ships, and the shipping director says all four now use it every day.

What it does. Syroco, a company from Marseille, builds a digital twin of each ship, a computer model of how that particular hull and engine behave. It combines the model with live data from the ship and the weather forecast, and recommends the route and engine power for the voyage. The crew and the office see the same picture.

How they got here. A trial started in mid-2025. A pilot voyage at the end of 2025 ran from Gdynia in Poland to Poland's research station in Antarctica, through hard weather. After the trial, the operator decided to use it on every ship.

The honest part. EuroAfrica says it saw fuel savings and good crew acceptance. It does not say how much. The only number in the story is Syroco's own: an average saving of 10% per voyage, with no word on how it was measured or against what.

Why we still run it. A customer that tries something, measures it on its own ships and then buys it for the whole fleet is a stronger signal than any brochure. Four ships is small, but every ship is in. That is a real decision, made with real money.

But the number matters. Five weeks ago we reported on an inland autopilot whose 5% fuel saving was checked in the field by a Dutch government agency. That is the standard. A saving the customer measured but will not publish is better than a supplier's average, and not as good as a number someone else checked.

For managers looking at voyage planning tools: ask the supplier for a reference customer who will tell you their own number, and how they measured it. Then run your own trial with a fixed method agreed before the first voyage. If the saving is real, it will survive being measured. [Splash247 · AiatSea, 20 September]

8. AiatSea: Tools Built for Factories and Offices Are Coming Aboard. What Carries Over?

Look at where this week's tools were born. Agents built for a paper mill. Inspection AI tested in an electronics plant. Design AI built for cars and aircraft. Office AI run by a company that just showed its books. Web software that runs on most of the world's servers. None of it was made for ships. Most of it will end up on ships anyway, because that is how shipping has always bought technology.

Some things carry over well. The factory agents read alarms, settings and maintenance history, the same things an engine room produces every day. They run on site, which is what a ship needs. The design AI could speed up the boring checks in a design office. The office AI already writes and reads the email that runs a fleet.

Some things do not carry over at all. A factory has an IT team on site. A ship usually has none, and the chief engineer becomes the IT department at 0200. A factory has a steady, fast connection. A ship has a link that is slow, expensive, or gone. A factory updates software on a Tuesday afternoon. A ship may wait until the next port, or the next dry dock. And a factory makes the same product a million times, so its AI has plenty of examples to learn from. A ship design is built a handful of times.

The PHP story shows why this matters. Old software is safe only as long as nobody looks hard at it. AI now looks hard, fast. Ships keep software for ten or twenty years. The tools we add to ships this decade will need fixes we cannot yet imagine, delivered to places a technician rarely reaches.

And the Anthropic story shows the other side. The office AI your staff use sits on very large fixed costs and a shortage of computing power. It may change price, limits or features faster than your contract changes.

Five questions before any of this week's tools comes aboard. Does it work with no link for three days? Who installs the fixes on board, how, and how quickly? What happens to the price and the service if the supplier's own costs change? Was it trained and tested on ships, or on factories and cars? And who measured the result, the supplier, the customer, or someone with nothing to sell? A tool that passes all five is worth a trial. A tool that fails the first two should stay ashore. [NVIDIA · Siemens · TechCrunch · VKTR · AiatSea, 18 October]

📊 Why It Matters — Strategic Impact Table

🔭 On Our Radar

• 🗳️ Who Is Actually Paying for Shipping's AI Accounts? — carried for an eleventh week: this week Anthropic's draft filing showed the large fixed costs and capacity limits behind the AI services many offices rent, we monitor whether any ship manager publishes a policy on which AI tools staff may use and who pays, and our reader survey is still open.

• 🏭 Does a Marine Supplier Put an Offline AI Agent in the Engine Room? — factory agents now read alarms and maintenance history on site, we monitor whether any engine maker or monitoring supplier offers an agent that runs on board without the link, and whether it may act or only advise.

• 📐 Does Anyone Show a Ship Designed With the New Engineering AI? — Mistral says it will publish its new large model by the end of October, we monitor whether any ship designer, yard or class society says publicly how it uses design AI, and how the results are checked.

• 💰 What Does the Final Filing Say About Price and Capacity? — the figures so far come from a draft seen by the press, we monitor the public filing for price, capacity and customer terms, and whether AI suppliers change plans or usage limits for business customers.

• 🩹 Which Maritime Suppliers Confirm Their PHP Status? — governments took sites offline to patch, we monitor whether shipboard and shore software suppliers say publicly whether their products were affected and how fixes reach ships.

• 🇰🇷 Does Korea's Test Area Publish Results? — approval is expected by the end of the year, we monitor the decision, the rules for the area, and whether Samsung or the authorities publish hours, encounters and human takeovers.

• 🇨🇳 Do Fishing Collisions Fall Off China's Coast? — 16,500 vessels now share fishing data on their charts, we monitor whether China publishes collision or near miss figures, and whether it brings the idea to international chart standards.

• ⛽ Will EuroAfrica Publish Its Own Fuel Number? — the operator measured savings and bought for the whole fleet, we monitor whether it or Syroco publishes the measured result and the method behind it.

📅 Critical Maritime AI Research Areas for Managers

1. 🏭 Offline AI Agents for Machinery: factory agents combine alarms, settings and maintenance history on site. Research should test whether the same approach works with real shipboard alarm and maintenance data, with no link for days, and measure false warnings per watch.
2. 📐 Design AI With Little Data: ships are built in small series, so design AI has few examples to learn from. Research should test which design checks AI can do reliably for ships, such as rule checks or clash detection, and how class should review AI-assisted designs.
3. 💰 The Real Cost of AI Services in Ship Management: AI suppliers carry large fixed costs and capacity limits. Research should track what ship managers actually pay per person for AI tools, how prices and limits change over a year, and the cost of switching supplier.
4. 🩹 Patching Software on Ships: AI now finds old software holes faster. Research should measure how long security fixes take to reach shipboard systems, which devices are never updated, and what a realistic patch process looks like with limited bandwidth and no IT staff.
5. 🇰🇷 Shared Results From Autonomy Test Areas: permanent test areas will produce years of data. Research should propose a simple common report for test areas: hours, encounters, takeovers and failures, that authorities and yards could all publish.
6. 🇨🇳 Fishing Data in Collision Avoidance: fishing boats are a leading collision risk in Asian waters. Research should test whether shared fishing ground data on charts reduces close encounters, and how it can be fed into AI collision avoidance systems.
7. ⛽ Customer-Measured Fuel Savings: buyers decide on measured savings that are rarely published. Research should set a simple, shared method for measuring voyage planning savings that a customer can apply and publish without a supplier's help.

📈 Top Investment Opportunities

1. 🏭 Industrial AI Agents That Run Without the Cloud — Schneider Electric, NVIDIA and HPE are building agents that read alarms and maintenance history on site — the opportunity is in versions for ships and offshore units, where the link is weak and nobody on board is an IT specialist, with clear limits on what the agent may do on its own — Schneider · NVIDIA · HPE
2. 📐 AI for Ship Design and Yard Engineering — design AI is arriving from aircraft, cars and chips while yards invest in digital twins and robot welding — the opportunity is in tools trained for small-series ship design, rule checking and clash detection, and in data services that connect design to production — Mistral + NVIDIA
3. 🩹 Security Fixes Delivered to Ships — AI now finds old software holes at internet scale, and ships keep software for decades — the opportunity is in services that list every device on board, check its software version and deliver fixes over limited bandwidth, with records class and insurers can read — PHP fixes · TechCrunch
4. 💰 AI Cost Control for Companies — the companies selling AI carry very large fixed costs and capacity limits — the opportunity is in tools that track AI use and spend per person, route work between several AI suppliers, and keep a company's data portable — Anthropic filing · VKTR
5. 🇰🇷 Autonomy Testing and Shared Navigation Data — Korea plans a permanent test area and China has put fishing data on 16,500 vessels — the opportunity is in test services, scenario libraries and navigation data products that both autonomous systems and human officers can use — Korea test area

📅 Top Monthly Picks

1. 📞 CMA CGM Puts AI on the Phone, and Almost Half the Calls Never Reach a Person — a top carrier's AI agents close 46% of cargo release calls; this week Anthropic's draft filing showed what sits behind the AI services companies now rent, and why price and capacity terms matter — Digital Ship
2. 🏭 Shipyards Spend Real Money on AI. Owners Should Ask for the Data. — Korea and Fincantieri put about $1bn and €100m into yard AI; this week design AI and a new digital shipyard in the US took the same idea one step earlier, to the drawing board — Splash247
3. 🛒 Nine Tech Bosses Say What They Would Never Buy Again — growing cloud bills and lock-in topped their regrets; this week's AI filing shows why those two questions now apply to AI services too — Splash247
4. 🛶 The Inland Autopilot Gets the Engine, and One of Its Numbers Was Checked by Someone Else — a 5% fuel saving checked in the field by a Dutch government agency; this week a fleet bought AI voyage planning without publishing its own number, which shows how rare that kind of check still is — Smart Maritime Network
5. 🧮 Two Hundred Software Flaws or Twenty, the Risk Came Out Almost the Same — a study found ship systems with very different numbers of known flaws scored almost the same risk; this week AI found two old holes in web software used everywhere, which makes knowing what runs on board more urgent — Digital Ship

Have feedback or a story tip? Email us or connect on LinkedIn.