Three constellations sit above your vessel right now. None of them is the answer. All of them, in the right combination, are.
Most boats reach for one provider. They sign with Starlink because it is the loudest name and the speeds are real. Then they discover, the hard way, what a single uplink actually buys: a great pipe most of the time, and an embarrassed crew the rest of the time. The Pentagon found this out with their drone fleet in late 2025. Yacht owners find it out the first time they cross the South Atlantic and the regional cell drops.
The right way to think about satellite connectivity for AI at sea is the same way you think about anything else on a vessel that matters. Build redundancy, separate concerns, and assume the link will fail, because the link will fail.
Here is the working comparison.
What "satellite for AI" actually means
Three different jobs, three different constellations.
The first job is bulk pipe. You are syncing a model update, pulling down a 30GB RAG corpus refresh, uploading three days of voice transcripts and sensor logs to shore-side. Throughput matters. Latency does not. You want LEO, you want gigabit-class, and you do not care if the round trip is 50 milliseconds or 500.
The second job is interactive. Crew is using a cloud-hosted tool, a guest is on a video call, the captain is running a weather routing app that pings shore. Latency matters now. Round-trip time over GEO satellite is 600+ milliseconds and feels broken. LEO at 30 to 70 ms feels like home.
The third job is heartbeat. The vessel is alive, the AIS is talking, the GMDSS distress channel is armed, somebody onshore can verify that the ship has not gone dark. Throughput is irrelevant. What matters is that this channel works when nothing else does, including in a storm, including over the poles, including when SpaceX has a bad routing day in São Paulo.
Different jobs, different physics, different orbits. Which is why you do not pick one constellation. You stack three.
Starlink Maritime: the bulk pipe
This is the one everyone defaults to and everyone should have. Starlink Maritime, marketed as Mobile Priority, runs on the same LEO constellation as residential Starlink with prioritized routing and a flat-panel terminal hardened for marine use.
The hardware is around $2,500 for the High Performance kit. Service tiers, after the March 2025 pricing reset, run from a $250/month entry plan with 50GB of priority data up through Global Priority plans where you are paying north of $5,000/month for multi-terabyte allotments. Throughput is 40 to 220 Mbps down, 8 to 25 Mbps up, with bursts past 400 Mbps when the cell is healthy and you have paid for it.
Latency is 20 to 40 milliseconds in the well-served regions. In the middle of the South Atlantic or off the coast of Madagascar, you will see 100 ms or more, because the packet is hopping through ground stations that are nowhere near you. That number matters if you are routing inference to a cloud GPU, which is part of why you should not be routing inference to a cloud GPU.
What Starlink is genuinely good for:
- Pulling down model weights and updates between voyages
- Syncing your local RAG corpus from shore-side
- Uploading bulk telemetry, sensor archives, and inference logs after the fact
- Crew video calls and entertainment, when the cell is healthy
What it is not:
- A safety system. Starlink has had multiple multi-hour regional outages in the last 18 months. We covered one in our note on the Pentagon drone outage.
- A polar solution. Above ~75° latitude, coverage is thin to none.
- A reason to put your AI in the cloud. We have written about why Starlink's gigabit ambitions do not fix vessel AI, and the math has not changed.
Starlink is your bulk pipe. Treat it that way.
Eutelsat OneWeb: the LEO redundant
OneWeb, now part of Eutelsat after the 2023 merger, runs a 654-satellite LEO constellation that completed global coverage rollout in 2024. In January 2026, Eutelsat awarded Airbus a contract for 340 additional satellites to expand capacity, which tells you something about where the funding is going.
OneWeb is sold differently. There is no consumer signup. You buy through a service partner (Marine Data Solutions, Concord, AST Networks, others), with terminals from Intellian (the OW10HM and OW11FM are the marine units) and data plans that range from 50GB up to 15TB per month. The OW11FM terminal sustains 100 Mbps down and 20 Mbps up. The OW10HM is closer to 30/6.
Latency is 30 to 70 milliseconds, comparable to Starlink. Throughput per terminal is lower. The constellation has more usable capacity per square kilometer in certain regions because the terrain is not as congested.
Why OneWeb in 2026:
- It is the one credible LEO alternative to Starlink for maritime use. That is the whole pitch.
- Eutelsat is a regulated European operator with different geopolitical exposure. If you have ever wondered what happens if SpaceX decides not to serve you, OneWeb is the answer.
- The hardware vendors have been building marine terminals for years, not retrofitting a dish on a roof.
What OneWeb is genuinely good for:
- A warm standby to Starlink. Run both, and your network policy fails over when one degrades.
- Primary connectivity in regions where Starlink's regulatory or routing posture is weak.
- Operators who cannot take the single-provider risk of running on SpaceX's roadmap.
What it is not:
- Cheap. Enterprise-only contracts mean you negotiate, and the floor is higher than Starlink's consumer plans.
- A throughput-leader on paper. If raw download speed is your only metric, Starlink wins. The point of OneWeb is that raw download speed is not your only metric.
Iridium Certus: the L-band heartbeat
Iridium Certus is the odd one out in this comparison because it is not competing on bandwidth. It is competing on the fact that it works.
Certus runs on the Iridium NEXT constellation, 66 cross-linked LEO satellites in a unique mesh architecture that gives true 100% global coverage. Including the poles. Including the regions where Starlink draws blanks. The hardware is L-band, which means the antenna is small, the link is weather-resilient, and the terminal pulls less than 50W when transmitting.
Throughput is 88 Kbps to 704 Kbps depending on service class. The next generation, Certus 1.4, brings that to about 1.4 Mbps. Latency is roughly one second, end to end.
Those numbers look bad next to a 220 Mbps LEO link until you remember what the job is. You do not run video over Iridium. You run:
- AIS position pings and vessel telemetry
- GMDSS distress alerts (Iridium has been a recognized GMDSS provider since 2020)
- Engine and bilge status messages
- A heartbeat channel that proves the boat is still on its feet, even if the LEO links are dark
For an AI deployment, Iridium is a control channel. You cannot serve a 70B parameter model over it. You can absolutely use it to confirm the vessel is alive, push a one-line incident alert to shore, and trigger a service call.
Pricing varies by reseller. A typical Certus 700 maritime plan runs $300 to $600/month for the data tiers a yacht would actually use, plus $4,000 to $8,000 for the terminal. There is a 3-month minimum commitment and a $50 activation fee on most resellers.
What Iridium is good for:
- Safety. GMDSS-grade reliability is not optional on commercial vessels and increasingly expected on serious yachts.
- The "is the vessel up" channel that runs when both LEO providers are degraded.
- Anywhere your route takes you above 75° latitude, where Starlink and OneWeb thin out.
- A control plane for your on-vessel AI: deployment alerts, model update notifications, "the local LLM is responding" health pings.
What it is not:
- A bulk pipe. Do not even try.
- Cheap on a per-megabyte basis. It is expensive bandwidth that exists because the alternative is no bandwidth at all.
The architecture: do not pick one
If you have been counting, the right answer for a serious vessel is all three. Each handles a different layer of the stack.
The architecture looks something like this:
- Primary path: Starlink Maritime. Most of your bandwidth, most of the time. Crew traffic, guest entertainment, bulk uploads, weather data, model syncs.
- Standby path: Eutelsat OneWeb. Hot standby, ideally with automatic failover at the network layer. When Starlink degrades or has a regional outage, your traffic shifts and most of your services keep working.
- Heartbeat path: Iridium Certus. Always on. Carries safety, status, and the small set of messages that have to get through regardless. Carries your AI control channel.
This is the same logic as redundant power on a vessel. You do not run the whole boat off one inverter. You do not run the entire connectivity stack off one constellation either. The broader argument for layered satellite design lives in our pillar on satellite connectivity for AI workloads at sea; this post is the per-constellation cut.
The deeper reason it matters is the same reason we keep coming back to local-first AI. Your inference, your knowledge base, your agents need to keep working when the link is degraded or gone. We have written about running 70B-parameter models offline on a vessel and about why cloud AI fails at sea. The connectivity stack above is what surrounds a sovereign, on-vessel AI deployment. The boat does its own thinking. The satellites carry the bits that have to leave.
This is what the knowledge ark looks like in practice. The library is on board. The connectivity layer is for syncing the library, broadcasting that the library is healthy, and reaching for help when something goes wrong. None of those jobs is well served by treating one provider as your only option.
Cost realism for a yacht-sized deployment
For a single 60m+ yacht, the all-three architecture pencils out roughly like this in 2026:
- Starlink Maritime: $25,000 to $40,000 per year, terminal amortized at ~$2,500
- Eutelsat OneWeb (mid-tier enterprise plan via Intellian terminal): $20,000 to $35,000 per year, terminal $15,000 to $25,000 once
- Iridium Certus 700: $5,000 to $8,000 per year, terminal $5,000 to $8,000 once
You are at $50,000 to $80,000 per year for connectivity, plus $25,000 to $35,000 in one-time hardware. For a vessel where the crew payroll is north of $1M and the annual operating budget is $5M+, this is a rounding error against the resilience it buys.
For a fleet operator running 50-100 vessels, the economics shift. Volume discounts on OneWeb get real, you can negotiate Iridium tonnage rates, and the per-vessel Starlink cost drops as you move into Global Priority enterprise pricing. The architecture is the same. The total bill scales sub-linearly.
The number that matters more than the bill is the one you avoid. A single connectivity-driven AI outage during a charter, a press tour, or a board cruise costs more than three years of OneWeb standby. A safety incident traceable to the vessel being out of contact costs more than every Iridium subscription you will ever pay.
What ShipboardAI actually deploys
We design for this stack. Our on-vessel AI runs locally regardless of link state. The connectivity layer is a tool for keeping the local deployment healthy and reaching shore when something needs to change, not a load-bearing assumption that anything in the cloud will be reachable when the bridge needs it.
If you are scoping a refit or a new build and want a working architecture for connectivity, on-vessel compute, and the AI workloads that sit on top of all of it, reach out. We have done this on actual ships, not in a deck slide.