Light Over Copper: Lightmatter and Qualcomm Set Data Transfer Speed Record for AI Clusters

As AI models grow in scale, a question that seems purely engineering-related at first glance becomes increasingly urgent: how quickly data moves between thousands of chips inside a data center. Copper cables, which have handled this task for decades, are starting to bottleneck the entire process – literally. Consequently, the industry is looking more intently toward optics.

Limitations of Copper Interconnects in AI Clusters

Why the «Bottleneck» Isn't Just a Metaphor

Modern AI clusters aren't just one powerful computer; they consist of hundreds of thousands of specialized chips that must constantly exchange data. The problem is that the capabilities of copper interconnects are physically limited: over long distances and at high speeds, they heat up, lose signal integrity, and consume massive amounts of energy. This is known as the «interconnect bottleneck» – a data transfer squeeze point that prevents the system from scaling further.

Fiber optics, which transmit data via light pulses, have long been a known alternative. However, integrating optics directly into server hardware at the chip level has proven to be a technical challenge. This is precisely the hurdle that Lightmatter is clearing.

Lightmatter and Qualcomm Achieve 1.6 Tbps Throughput per Fiber

A Record Worth Explaining 💡

On March 11, 2026, Lightmatter announced a joint milestone with Qualcomm Technologies: they successfully achieved a throughput of 1.6 terabits per second over a single optical fiber. For comparison, existing solutions in the same class provide roughly 8 times less speed. This isn't an incremental gain – it's a quantum leap.

Put simply: where eight fibers were previously required, one now suffices. This means fewer cables in the data center, streamlined infrastructure, and lower costs – all while transferring a significantly higher volume of data in the same amount of time.

At the heart of this development is a technology that sends 16 simultaneous light streams across different wavelengths through a single fiber. Each stream carries data independently without interfering with its neighbors. This allows for a manifold increase in total data capacity without needing to replace the cable itself.

Silicon Photonics Technology and Strategic Partnerships

Who Is Behind This and Why It Matters

Lightmatter is a company specializing in photonic computing – technologies where light is used not only for data transmission but potentially for processing as well. Their Passage platform is a silicon photonics engine that integrates into existing hardware, serving as an optical interface between chips.

In this partnership, Qualcomm Technologies is responsible for the high-speed SerDes components – specialized circuits that convert data streams for optical transmission. Their involvement isn't just about supplying parts; it's a strategic collaboration aimed at creating a unified solution ready for mass production.

An important detail: this isn't just a lab prototype. The company describes the technology as «silicon-verified» and ready for high-volume manufacturing – meaning it's a solution that has already been tested on real hardware and can scale to meet the needs of major cloud providers.

Impact of Optical Interconnects on Data Center Infrastructure

What This Changes for Data Centers

The goal outlined by Lightmatter is 100 terabits per second per chip package. The current result of 1.6 Tbps per fiber is a major step on that path. When you multiply this figure by the number of fibers that can be housed in a single connector, the numbers become staggering.

For operators of large-scale AI infrastructures, this is vital for several reasons. First, reducing the number of physical cables simplifies management and lowers the chance of failure. Second, optics consume significantly less power than copper at comparable speeds. For data centers where energy is measured in megawatts, this represents a substantial line item in the budget.

Analyst Vlad Kozlov from LightCounting notes that the industry has reached a point where incremental improvements can no longer keep pace with the growth of AI clusters – which is why breakthroughs in bandwidth density like this are becoming key to the entire ecosystem.

Future Deployment and Commercial Availability of Photonic Solutions

What's Next

The first samples are already available for testing by key customers. Lightmatter will showcase the development at the OFC 2026 Optical Fiber Communication Conference in Los Angeles, which runs from March 15 to 19.

One question remains: how quickly will such solutions reach mass deployment? The journey from «production-ready» to «installed in thousands of server racks» typically takes more than a year. However, the fact that the technology has already passed silicon verification suggests that the gap between the lab and practical application has narrowed significantly.