AWS plans 420 Tbps Sta'O'Nuk subsea cable between Japan and Washington State for 2029
Fibre pairs, landing stations and layered encryption explain why cloud providers now build their own ocean crossings.
Amazon Web Services is building a new subsea cable between Japan and Washington State designed to carry 420 terabits per second, according to a report published on 2 September by Data Center Knowledge, which describes the project as a new US AI route across the Pacific. The system, named Sta'O'Nuk, is scheduled to enter service in 2029 and is intended to support AI workloads, including distributed training of large language models, as well as financial transactions, edge computing and video streaming. The stated purpose is both to serve AI workloads and to expand trans-Pacific network capacity generally.
The cable will comprise 20 fibre pairs and land at Ocean Shores, Washington, at a new landing station being developed by Toptana Technologies, a company owned by the Quinault Indian Nation. Assured Communications is named as a strategic partner and operations provider. From the coast, terrestrial routes are planned to Seattle and to Hillsboro, Oregon, following the Interstate 5 corridor.
The Ocean Shores station is designed to accommodate up to four subsea cable systems and can expand to 16. The report places the project in the context of AWS's wider network, which the company says spans more than 20 million kilometres of fibre. Traffic on that network is protected by encryption at several layers: quantum-safe optical encryption at layer 1, MACsec at layer 2, and TLS, SSL and QUIC at layer 4.
Jimmy Yu, a vice president at analyst firm Dell'Oro Group, told the outlet that hyperscale cloud providers were already among the largest consumers of subsea capacity before the current wave of AI investment.
Why it matters
Cloud providers once leased capacity on cables built by telecom consortia; increasingly they build their own, choosing the landing points, the terrestrial backhaul and the security stack end to end. A 420 Tbps trans-Pacific link is scaled for moving training data and model checkpoints between regions, not for ordinary web traffic. That signals how AI is reshaping network planning: compute is placed where power is available, and the network has to stitch those sites into one system. The landing station's design for up to 16 cables suggests this route is expected to keep growing well beyond 2029.
Layer 4: TLS, SSL and QUIC
Protects each application session end to end, across any number of routers.
Layer 2: MACsec
Encrypts frames between adjacent network devices on a link.
Layer 1: quantum-safe optical encryption
Encrypts the light signal itself, so a tap on the fibre yields nothing readable.
What you can learn from this
Cable capacity is fibre pairs multiplied by wavelengths. A subsea system carries several pairs of optical fibres, two fibres forming a pair because each carries traffic in one direction, and each pair carries many wavelengths of light using dense wavelength-division multiplexing. Dividing 420 Tbps by 20 pairs gives about 21 Tbps per pair, which is achieved by packing dozens of channels into a single fibre. Capacity upgrades often come from new terminal equipment at each end rather than a new cable, which is why fibre count matters for the life of the system.
The landing station is where the sea meets the grid. A cable landing station houses the equipment that terminates the submarine fibre, powers the undersea repeaters and hands traffic to terrestrial networks. Its location determines the onward path to data centres, and a station built for multiple systems becomes a hub. The backhaul routes to Seattle and Hillsboro matter as much as the cable itself, because a fast ocean crossing feeding a congested land route gains nothing.
Hyperscalers build cables to control the whole path. Leased capacity comes with someone else's maintenance windows, routing decisions and commercial terms. Owning the system lets a provider choose diverse landing points to avoid single points of failure, engineer latency for its own regions, and apply its own encryption from the optical layer up. It also converts a recurring lease into a capital asset, which suits companies with large balance sheets and predictable long-term demand. The trade-off is that the owner also carries the repair risk when a cable is damaged.
Encryption at multiple layers is defence in depth for links you do not fully control. Layer 1 optical encryption protects the raw light signal, so a tap on the fibre yields nothing. MACsec at layer 2 protects frames between adjacent network devices. TLS and QUIC at layer 4 protect each application session end to end. Each layer covers a case the others miss: optical encryption does not survive a router, while TLS does not hide traffic patterns on the wire.
Distributed training is a bandwidth problem before it is a compute problem. Training a large model across sites requires exchanging gradients or synchronising checkpoints frequently, and if the network cannot keep up, expensive accelerators sit idle. That is why a cable announcement lists model training as a driver. Physics also sets a floor: light in fibre crosses the Pacific in tens of milliseconds, and no cable can go faster than that. Engineers therefore design training jobs to overlap communication with computation wherever possible.
We teach this
How to use this in practice
Measure your own path across the Pacific. From your machine, run
mtr --reportortracerouteto a public endpoint in an AWS region on the other side of the ocean, then compare the round-trip time with the theoretical minimum, which is roughly the great-circle distance divided by about 200,000 km/s and doubled. Public regional EC2 or S3 hostnames work as targets, and running the test three times at different hours will show that congestion on the land segments varies more than the ocean crossing does. Done looks like a table of measured versus theoretical latency and the number of hops, with a note on where the largest jump occurs.Map the encryption layers onto the OSI model. Draw the seven layers, mark where optical encryption, MACsec and TLS or QUIC each sit, and for each write one threat it stops and one it does not. Add a fourth row for the case of a compromised router inside the path, and note which layers still protect the data. Done looks like a diagram you could use to explain why a service still needs TLS on a link that is already encrypted at layer 1.
Find the cables that serve your country. Open a public submarine-cable map, list the systems that land nearest you, their landing stations, and their ready-for-service years, noting which are owned by consortia and which by a single cloud provider. Done looks like a short table plus one sentence on what would happen to your connectivity if the busiest landing station went offline.
Cost a large data transfer both ways. Calculate how long moving one petabyte would take at 10 Gbps and at 100 Gbps, then look up your cloud provider's published data-transfer pricing and estimate the egress charge for the same volume. Then add a third row for shipping physical drives, which cloud providers offer as a service, and see at what volume it wins. Done looks like a three-row comparison that shows whether time or money is the binding constraint for a transfer that size.
Sources
- AWS Wires a New US AI Route Across the Pacific — Data Center Knowledge
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report