Eliyan's NuLight targets the copper wall choking AI scale-up

Eliyan, a Santa Clara-based connectivity chipmaker, has introduced NuLight, a platform it describes as the first 2D microVCSEL interconnect solution designed to carry AI compute traffic beyond the reach and bandwidth limits of copper cabling. The launch, unveiled at the OCP Summit 2026 this week, centres on two chiplets: Nexus, a 1.6 terabit-per-second reverse gearbox built on TSMC's 3nm process, and Oscar, an electrical interface chip that bridges high-speed signals to arrays of vertical-cavity surface-emitting lasers (VCSELs) and photodiodes. Both are sampling now.

The announcement crystallises a constraint that is quietly becoming one of the most consequential bottlenecks in AI infrastructure. As hyperscale operators pack ever-larger clusters across multiple racks, copper interconnects struggle to sustain the bandwidth and reach that training and inference workloads demand. Eliyan calls this the "copper wall", a point at which the physics of electrical signalling impose hard limits on AI scale-up architectures.

Optical economics, not optical premiums

The strategic proposition of NuLight is as much commercial as technical. Optical interconnects have historically carried a significant cost premium over copper, making them impractical for the dense, rack-to-rack links that AI clusters rely on. Eliyan's approach uses mature gallium arsenide (GaAs) laser technology, a more widely available and cost-stable supply chain than the indium-phosphide lasers that premium optical modules typically require, and targets an initial reach of approximately 50 metres at what it positions as "copper economics."

The architecture distributes bandwidth across more, slower optical lanes running parallel NRZ signalling rather than fewer, faster PAM4 lanes. The effect, the company says, is lower per-lane signal-processing complexity and better power efficiency than conventional optical interconnects. A patented on-the-fly lane repair mechanism redirects traffic to spare optical lanes if an active lane degrades, without interrupting the link, a feature aimed squarely at the uptime demands of large-scale AI training runs, where a single interconnect failure can stall operations across thousands of accelerators.

Modular 1.6Tbps slices are designed to aggregate to 6.4Tbps and 12.8Tbps configurations, with a stated architectural path toward co-packaged optics (CPO), where future Oscar variants would interface directly with host ASICs via die-to-die connectivity.

The wider convergence stakes

The timing of this launch reflects a broader infrastructure inflection point that sits well beyond a single chiplet product. The hyperscale capital expenditure cycle is entering a phase where compute density is outrunning the interconnect fabric designed to support it. Analysts tracking AI infrastructure spend have noted that networking and interconnect now represent a growing share of total cluster cost, as operators find that adding more GPUs yields diminishing returns without commensurate fabric upgrades.

For the cross-sector investor, the strategic read-across is significant. The shift from copper to optical at the rack and multi-rack level will reshape procurement across the AI supply chain: laser component suppliers, optical module integrators, and the GaAs wafer fabs that underpin microVCSEL production all stand to benefit if platforms like NuLight achieve broad adoption. Conversely, established copper-interconnect suppliers face structural headwinds as AI cluster sizes continue to grow.

There is also a geopolitical dimension. Nexus is designed on TSMC's 3nm node, placing it firmly within the Taiwan-centred advanced-node supply chain that remains subject to ongoing export-control scrutiny and capacity constraints. As sovereign AI buildout accelerates across the US, GCC, and Europe, the availability of advanced interconnect chiplets, and the fabs that produce them, will become an increasingly live policy question for governments seeking to onshore AI infrastructure capability.

"Scaling AI infrastructure requires efficient, reliable connectivity that reaches beyond a single rack," said Ramin Farjadrad, co-founder and CEO of Eliyan. "With our NuLight chiplet platform, customers can scale aggregate bandwidth through reusable building blocks."

Whether NuLight achieves the industry adoption its launch positioning claims will depend on whether it can compete with pluggable optical module ecosystems already embedded in hyperscaler procurement cycles, a competitive question that sampling availability alone does not yet resolve.