Marvell pitches optical links spanning thousands of km

Key Takeaways

• Marvell's Colorz 1600 coherent optical solution, built on 2nm DSP, samples later this year with 1.6 Tb/s bandwidth
• The company envisions pooling CPUs, GPUs, and memory across data centers thousands of kilometers apart
• Marvell argues copper interconnects are hitting a wall as AI clusters outgrow single data centers

Marvell wants to connect data centers thousands of kilometers apart using coherent optical links, effectively pooling compute and memory resources across vast distances. CEO Matt Murphy laid out the vision at Computex 2026, announcing that the company's Colorz 1600 coherent optical solution will begin sampling later this year.

The pitch is ambitious. Murphy argues that AI workloads have outgrown single data centers, forcing hyperscalers to build entire campuses of interconnected facilities. Today, those connections rely heavily on copper, which imposes hard limits on how far GPUs can sit from memory and how large clusters can scale. Marvell's bet: replace copper with optical links everywhere, and those limits start to dissolve.

"We are hitting a copper wall," Murphy said. "We are moving to a connectivity-first architecture, transitioning from electrical to optical interconnects to enable larger, more efficient AI fabrics."

What is Marvell actually shipping?

The Colorz 1600 is the headliner. It delivers 1.6 Tb/s of coherent optical bandwidth using a digital signal processor built on 2nm process technology. Marvell positions it for inter-data-center connectivity, the long-haul links between facilities that can be hundreds or thousands of kilometers apart.

Beyond that, Marvell announced the Ara 1.6 Tb/s family for shorter-range data center interconnects, using 3nm DSPs. And the Teralynx T100, a 102.4 Tb/s Ethernet switch, rounds out the portfolio. That switch supports 512 ports at 200 Gb/s or 64 ports at 1.6 Tb/s.

The timing matters. Samples arriving later this year means production volumes in 2027 at the earliest. Hyperscalers planning their next-generation AI infrastructure will be evaluating this technology against competing offerings from Broadcom, Cisco, and others.

Why does distance matter for AI workloads?

Current AI systems like Nvidia's NVL72 use copper cables for scale-up connections, the high-speed links between GPUs within a single cluster. Copper works fine when everything sits close together, but it limits how big a cluster can get before signal degradation becomes a problem.

Murphy's argument is straightforward: as AI models grow, so do the clusters needed to train them. A single data center can't house enough accelerators for the largest workloads. You need multiple facilities working as one.

The "data center without distance" phrase captures the ambition, but it also invites skepticism. Light travels fast, but physics still applies. A round trip across 1,000 km adds roughly 10 milliseconds of latency, an eternity for synchronous AI training workloads. Engineers in Hacker News discussions have raised this point, questioning whether Marvell's vision applies to training or primarily to inference and other latency-tolerant tasks.

The disaggregation play

Marvell's longer-term pitch goes beyond connecting data centers. Murphy described a future where optical links enter servers themselves, allowing CPUs, GPUs, and memory to be disaggregated into separate pools.

Today, when a company buys an NVL72 system, they get a fixed ratio of CPUs, GPUs, and high-bandwidth memory. That ratio might be perfect for one workload and wasteful for another. If a task needs more memory than compute, the operator still has to buy additional GPUs just to access the HBM attached to them.

In a disaggregated architecture, operators would compose virtual machines from shared resource pools, assembling exactly the ratio of compute, memory, and networking each workload requires. Murphy claims even a 10% improvement in utilization would save billions, given the scale of hyperscaler deployments.

This is a compelling vision for CFOs watching AI infrastructure costs spiral. It's also a vision that requires fundamental changes to how data centers are built and operated. The industry has tried disaggregated computing before, with Intel Rack Scale Design and similar efforts. Those projects stalled on software complexity and latency sensitivity. Optical interconnects address the latency piece, but the software challenge remains.

Scale-up to scale-out transition

Murphy outlined a progression. First, optics will expand scale-up domains from today's 72 or 144 accelerators to 1,000 or more. That's a meaningful expansion. Nvidia's current flagship connects 72 GPUs over copper. Moving to optical would allow significantly larger coherent clusters.

After that, optical connectivity would push into servers themselves. At that point, the line between scale-up and scale-out blurs. Everything becomes a fabric.

The power efficiency angle also matters. Electrical-to-optical conversion burns energy. But copper cables at high data rates require significant power for signal integrity, and that power scales with distance. For long-haul connections, optics already win on efficiency. The question is where the crossover point sits for shorter links.

Market timing and competition

Marvell isn't alone in this space. Broadcom, Cisco, and startups like Lightmatter are all pushing optical interconnect technology for AI infrastructure. Nvidia itself is rumored to be working on co-packaged optics for future systems.

The timing of Marvell's announcement, at Computex alongside major GPU and accelerator reveals, positions the company as an infrastructure enabler rather than a compute competitor. That's a comfortable spot. Whoever wins the AI chip race still needs connectivity, and Marvell wants to supply it.

The coherent optics market is projected to grow substantially through the decade, driven by AI workloads and cloud expansion. Marvell's 2nm DSP in the Colorz 1600 represents a meaningful process lead if it delivers on specifications.

Frequently Asked Questions

When will Marvell's Colorz 1600 be available?

Marvell says the Colorz 1600 coherent optical solution will begin sampling later this year, with production volumes likely following in 2027.

How fast are Marvell's new optical interconnects?

The Colorz 1600 and Ara families both deliver 1.6 Tb/s bandwidth. The Teralynx T100 switch offers 102.4 Tb/s aggregate throughput.

Can optical links really span thousands of kilometers for AI training?

Technically yes, but latency becomes a factor. A 1,000 km round trip adds roughly 10ms of latency, which may limit use cases to inference and latency-tolerant workloads rather than synchronous training.

What problem does disaggregated computing solve?

Today's AI systems bundle fixed ratios of CPUs, GPUs, and memory. Disaggregation would let operators compose custom configurations for each workload, improving utilization and reducing waste.

Who competes with Marvell in optical interconnects?

Broadcom, Cisco, and startups like Lightmatter are developing competing optical interconnect solutions. Nvidia is also rumored to be working on co-packaged optics.

Source: Latest from Tom's Hardware