Marvell details vision of optically-interconnected data centers spanning across thousands of kilometers — new interconnects sampling later this year would allow CSPs to pool resources based on workload

While hyperscalers rush toward expansion amid the swelling demand for AI data centers, Marvell last week shared its vision for an optical interconnect solution that can theoretically pool resources between discrete data centers across thousands of kilometers.

Optical interconnections are steadily being deployed across the industry, over both short and long-distance connections, and we're going to be seeing much more in the future, according to Matt Murphy, Chief Executive at Marvell, speaking at Computex 2026.

"Imagine future data centers, a globally optically interconnected data infrastructure," Murphy said. "These rigid boundaries we have today, and the systems we have, they begin to disappear. Compute can now be pooled, memory can be pooled, and infrastructure can be composed dynamically at scale."

Constrained by distance

Murphy says that workloads no longer fit within one data center, which is why hyperscale cloud service providers increasingly need to build entire campuses consisting of multiple data centers connected by high-speed links, as clusters are becoming larger than a single data center.

Today, connecting multiple data centers within a single campus is not easy or cheap, but relatively straightforward. However, Marvell envisions that in the future it will need to connect data centers that are located at considerable distances from one another.

This is why Marvell is working on coherent optics and long-haul scale across optical networking technologies, which will connect data centers separated by thousands of kilometers. Marvell already has products which enable such connectivity today, including the Colorz 1600 1.6 Tb/s coherent optical solution based on a 2nm DSP, which targets inter-data-center connectivity and will sample later this year.

In addition, Marvell says it will offer the Ara 1.6 Tb/s family of interconnect solutions for data centers (with 3nm DSPs) as well as the Teralynx T100 102.4 Tb/s Ethernet switch, which supports 512 ports running at 200 Gb/s or 64 ports running at 1.6 Tb/s.

Murphy argues that today's architectures are constrained by distance because of copper interconnects: CPUs sit near memory because latency matters, GPUs sit near memory because bandwidth matters. As a result, workloads must be partitioned according to those physical limits. The head of Marvell claims that once optical interconnects penetrate scale-up interconnects, scale-up domains will not be limited by copper cable lengths, and those constraints will begin to disappear.

Nowadays, scale-up AI solutions, such as Nvidia's NVL72, are connected using copper wires, but scale-out connections tend to use optical interconnects. Once the number of AI accelerators within scale-up systems increases, they will also have to move to optical links, according to Marvell. This means that virtually all data center-grade interconnections will become optical, which might inspire hardware developers to reconsider the architecture of data centers.

Pooling resources

Murphy presented a rather interesting vision: firstly, optics will expand scale-up domains from 72 or 144 accelerators to 1,000 or more. But after that, optical connectivity will enter servers themselves. This will enable developers to disaggregate CPUs, accelerators (Marvell calls them XPUs), and memory into separate pools as distance will no longer matter, enabling better configurability and utilization.

"It is a data center without distance, where compute, memory, networking, and photonics operate as one unified system, where millions of resources across the data center can work together as if they were one machine," the head of Marvell said.

Keeping in mind that hyperscalers deploy hardware worth billions of dollars, even a 10% higher utilization will save a lot of money, and companies like Nvidia are clearly paying attention.

"In today's systems, the ratio of CPU and XPU or GPU is fixed, so these ratios have to be defined at the time the system is built and deployed, but no two workloads require exactly the same ratio," Murphy stressed. "Imagine a completely disaggregated architecture, XPUs in one system, memory in another, generic CPUs in another."

Today, companies buy something like an NVL72 system and get a fixed ratio of CPUs, GPUs, and memory, which may be efficient for certain workloads and inefficient for others. In the future, operators will be able to assemble a virtual machine from shared pools of systems, allowing for customization and flexibility, based on the type of workload. If a workload needs more memory than compute, operators often have to buy additional GPUs just to get the extra HBM, but they may just get memory in the future if Marvell's vision comes to pass.

"Once we decompose the system into separate pools of compute, memory, and they are all optically interconnected, we can then compose dedicated systems on the fly, which are then optimized wherever the workload is," Murphy said. "For the first time, architects can begin designing AI systems around the needs of the model, not around the limits of the interconnect."

One detail

While Marvell has the know-how to interconnect data centers across thousands of kilometers and technologies that enable pooled data centers, these visions do not necessarily intersect. Data centers located thousands of kilometers away cannot share resources — a 1,000 km round-trip takes light 10ms — which makes such long-distance resource sharing inefficient from a latency point of view.

However, Marvell's technologies enable hyperscale CSPs to synchronize AI campuses, access distributed storage, replicate data, and perform other operations that do not depend on latency. Meanwhile, the synchronization of AI campuses on different continents in a matter of hours could be a killer app for hyperscalers.