Six million transceivers move signals.
They burn a power plant doing it.
The number that stopped the industry cold.
At GTC 2025, NVIDIA's CEO put a figure on stage that the data center industry has not stopped thinking about since.
A million-GPU cluster needs roughly six million optical transceivers to move data between accelerators. Those transceivers consume around 180 megawatts. They do no computation. They run no model. They train no weights. They move signals from one place to another and dissipate the rest as heat.
A hundred and eighty megawatts. To move signals.
This is the largest concentrated waste in modern AI infrastructure — and almost no one in AI is working on it. The optical industry that ships these parts today is optimized for telecom, not for the energy reality of a gigawatt data center.
We are.
What we make.
A family of optical transceivers built for the AI data center — at the speeds the fabric now runs at, and at a fraction of the power conventional transceivers consume to do the same job.
800G. 1.6T. 3.2T. OSFP and OSFP-XD form factors. Drop-in compatible with the switches and accelerators already deployed in hyperscale fabrics today. Reach from inside the rack to forty kilometers between facilities.
The difference is what's inside. Or rather — what isn't.
What we removed.
A conventional transceiver running at these speeds contains a digital signal processor. The DSP cleans up signal distortion electronically — and consumes roughly two-thirds of the device's total power doing it. At fleet scale, the DSP is the reason transceivers became a megawatt-class line item on the data center power budget.
We removed it.
Our transceivers are linear pluggable optics — the signal path is analog end to end, with no DSP recovery stage. This is not a software optimization. It is a physical change in the device, made possible by an integrated light source built directly into the photonic wafer rather than bonded on as a discrete component.
The result is a transceiver that delivers the same line rate at a fraction of the power, with a simpler bill of materials, fewer failure modes, and a manufacturing flow that yields at a higher rate than the industry's conventional process.
What changes.
For a hyperscale operator, a transceiver is not a component. It is a multiplier. Replace one transceiver and you save a few watts. Replace six million and you free a power plant.
Detailed performance specifications and qualification data are shared with qualified partners under NDA.
For a fleet of six million transceivers in a gigawatt-scale facility, the difference between a conventional DSP transceiver and a LiteEdge transceiver is on the order of a hundred and thirty megawatts — power that returns to compute, to cooling headroom, or to the operator's bottom line.
Why the industry has not done this already.
The optical transceiver industry is roughly four decades old. Its assumptions are inherited from telecom — long-haul, low-volume, where every device justifies a digital signal processor and a discrete laser bonded onto the package by hand.
AI broke those assumptions in eighteen months. The volume is now a hundred million units a year, growing toward several hundred million by decade's end. The power envelope is now the binding constraint on whether a data center can operate at all. The integration challenge is no longer telecom-grade — it is silicon-grade.
The companies shipping today's transceivers are good at what they do. What they do is the previous decade's product, manufactured the previous decade's way.
We started over.
The manufacturing story.
A conventional transceiver is built by bonding a discrete laser onto a photonic chip, wire-bonding the connections, aligning fibers by hand, and testing each unit through an aging cycle to weed out the ones that didn't survive assembly.
Our process integrates the laser into the photonic wafer at fabrication. The bonding step disappears. The alignment step disappears. The yield-loss step that follows them shrinks dramatically. What is left is closer to a CMOS process than to a hand-assembled optical module — which means the cost curve, the volume curve, and the quality curve all bend in the directions a hyperscale buyer cares about.
Where it fits.
Inside the rack and across the row.
The bulk of an AI fabric's transceiver count is intra-cluster, GPU-to-GPU and rack-to-rack. This is where the megawatt arithmetic is most punishing — and where our power advantage compounds the hardest.
Across the data center campus.
Up to forty kilometers between facilities, at full line rate, on the same architectural family. No separate product line for short-reach versus long-reach.
Future-proofed for the next speed transition.
The 1.6T market is ramping now. The 3.2T market arrives behind it. Our roadmap covers both on the same integrated platform — one architecture, three generations of speed, one upgrade path for the operator.
The economics.
A megawatt of saved transceiver power is a megawatt that does not need to be generated, transmitted, cooled, or paid for. At today's industrial electricity prices, in a gigawatt-scale facility, the annual operating savings from replacing a conventional transceiver fleet with ours is measured in tens of millions of dollars — every year, for the life of the fabric.
The hardware cost is lower than the conventional alternative. The power cost is dramatically lower. The reliability is higher. The lead time is shorter.
The conversation a hyperscale procurement team has about transceivers is usually about price per port. Ours is about the only line item in the data center that grows with every accelerator added — and what happens when you stop letting it grow.