Starlink is back in the news this week with fresh talk of gigabit-class speeds coming to the fleet in 2026. The Yachting Monthly April issue mentions it. Clarus Networks' performance kit marketing leans on it. A half-dozen maritime IT consultants are already pitching yacht owners on the upgrade path.
I want to take a minute to talk about why, for AI on a vessel, the gigabit number is almost entirely beside the point.
What actually breaks when you lose the link
I spent a decade flying helicopters off warships before I got into AI infrastructure. In that world, "the link dropped" is not a support ticket. It is an operational problem that happens regularly, at the worst possible time, and the difference between a professional operator and an amateur is whether the capability on the other end of the link still works.
The same truth applies at sea in peacetime. The question is not "how much bandwidth can Starlink deliver when the sky is clear." The question is: what happens at 0230 local, in a squall, when the LEO constellation's view is degraded, the rain fade is eating your signal, and a guest in cabin 4 asks the concierge AI why the TV remote is not working?
If your AI is in the cloud, it fails. If your AI is on the vessel, it works. Gigabit on the slide deck does not change the answer to that question by one bit.
The three failure modes nobody puts on a spec sheet
I have been watching this space long enough to see the same three things break again and again. None of them are throughput problems.
1. Weather. Ku and Ka band satellite links (which Starlink Maritime uses)take a hit from rain, snow, and heavy cloud cover. Rain fade at Ka band can cost you 10+ dB of signal in a serious storm. The vessel is still there. The crew is still there. The guests are still there. The cloud AI is not. Gigabit peak throughput has nothing to do with whether your system works when the weather turns.
2. Obstruction. Starlink's LEO birds need a mostly clear view of the sky. Between high-superstructure yachts, the angle of heel on a sailing vessel, large nearby vessels in a crowded anchorage, and the occasional flybridge that sits directly between the antenna and half the sky, there are a lot of legitimate situations where a Starlink link goes from "working" to "no service" in a few seconds. The gigabit pitch assumes the antenna has sky. The reality is that sky is not always available.
3. Latency variance. Even when Starlink is "up," maritime links bounce. Jitter is measured in hundreds of milliseconds, not the tens of ms you would see on a terrestrial connection. An LLM streaming token by token over a 400ms jittery link feels broken even if the average throughput number looks fine. Gigabit does not fix jitter. It compounds the disappointment when the number looks good and the experience feels bad.
What I would actually tell a yacht owner
If a charter manager or a fleet captain asked me "should I wait for gigabit Starlink before building out AI on my vessel?" my answer would be the same answer I have given for the last two years.
No. Build the vessel assuming the link is unreliable, because it is.
Put an AI stack on the vessel that works completely offline. Run your guest concierge, your crew operations assistant, your predictive maintenance models, and your voice interfaces against local hardware and local models. When the link is up, use it for things that benefit from the cloud, model updates, heavy one-time compute, large file sync. When the link is down, nothing that your guests or crew actually touch should notice.
That is what sovereign AI at sea actually means. It is not a fancy phrase. It is the design principle that separates a vessel that impresses people in a harbor with Starlink from a vessel that still impresses people 200 miles offshore in a squall.
The quiet secret nobody in the maritime IT press wants to write
If you ask me, the gigabit marketing is the tell. Maritime connectivity vendors want owners to believe that the next throughput upgrade will finally make cloud-first architectures viable at sea. It will not. The constraint has never been average throughput. The constraint is reliability, weather sensitivity, and the fact that every single LEO constellation ever built depends on conditions that are not always there.
Build for the world as it is, not the world the marketing wants.
And when the squall hits and the link drops and the guest in cabin 4 still gets an answer to their question, that is the day you find out whether you bought sovereign AI or expensive slideware.
Designing an AI architecture for a vessel that has to work when the weather does not cooperate? Let's talk. We build for the squall, not the sales demo.