Led by Coherent, the Data Center Is Turning to the Light
Hyperscale data centers buckle under the heat, weight, and waste of their hundreds and thousands of miles of copper connections linking together thousands of graphical processing units (GPUs). Coming to the rescue is light in the form of optical connections, replacing copper in the nervous system of the AI data center.
As workloads scale, every additional GPU, accelerator, and memory module must move data faster, farther, and at the lower power and greater fidelity only optics can provide.
The company that already defines this transition is Coherent. Its leadership in optical transceivers — now shipping at 1.6 terabits per second — gives it a commanding position in the practical rollout of photonics across the modern data center. Coherent’s technology moves light across racks and rows where performance, power, and reliability converge at hyperscale.
The Physics Problem
Copper has long done this job, but the physics are starting to protest. Each new generation of AI chips pushes more data through the same metal traces, generating heat, distortion, and wasted energy. The signal boosters that keep copper alive add their own overhead, until the cost of moving bits begins to rival the cost of computing them.
Light sidesteps all of that. Photons travel cleanly through fiber, immune to electromagnetic noise, needing no re-timers or shielding. They carry more data, run cooler, and consume far less energy.
As shown in Table 1, the difference between electrical and optical interconnects is both physical and economic. Optical links deliver longer reach, greater bandwidth per lane, lower power per bit, and virtually no signal interference. These advantages compound as systems scale. Where copper struggles to sustain signal integrity beyond a meter or two, photonics can span entire racks with consistent performance. Power drops by multiples, latency shrinks, and thermal load plummets — translating directly into lower total cost of ownership.
At hyperscale, those gaps turn into megawatts. Light isn’t just faster; it’s inevitable — the next medium for an intelligence that’s outgrown metal.
Coherent’s commanding lead in AI optics
Coherent’s fiscal-2025 results (ended June 20, 2025) show what inevitability looks like when it hits the income statement. Data center and communications revenue climbed 51% for the year, with AI datacenter sales up 61%.
Yet the company began shipping 1.6-terabit transceivers only in the fourth quarter, marking a new performance tier. Demand for modules below that level remains strong, and Coherent is already working on 3.2-terabit designs.
These are active transceivers — the small modules that sit at the edge of racks, converting electrical signals into light. They’re not fiber cables; they’re the engines that drive the link itself.
Coherent 1.6T Optical Transceiver Module
Source: Coherent
Coherent has achieved scale production and qualification with nearly every major hyperscaler. Optical’s next step is to replace metal on a more micro scale aimed at chiplet or board-level integration. Startups including Ayar Labs, Lightmatter, and Celestial AI are developing experimental technologies to do this. As these companies are all pre-IPO and not yet publicly traded, we will report more on them soon in our sister publication, Gilder’s Private Reserve, which shows readers how to invest in start-ups at an early stage.
For now, however, Coherent is the one. It is already a multibillion-dollar public company serving global hyperscalers.
System-level optics has the economics
Inside a hyperscale rack, every watt counts. As shown in Table 2, moving from copper to optical links at the rack level sharply reduces total interconnect power, heat output, and lifetime cost. The example models a rack with thousands of 1.6 Tb/s links. By eliminating re-timers and minimizing signal loss, optics reduce power draw by roughly two-thirds and free up energy for additional GPUs within the same electrical envelope. These efficiencies translate into higher throughput per rack and lower cooling requirements, directly improving operating margins for hyperscalers.
For hyperscalers, that difference translates directly into more GPUs per rack — more compute for the same electricity.
Coherent anchors the optical stack, now and to come
The startups mentioned here have been getting lots of publicity. But their work does not displace Coherent’s; it depends on it. If successful, each would add functionality closer to the chip — faster I/O, optical interposers, and photonic fabrics that will eventually draw from the same light sources and integration pipelines Coherent already supplies. Together, they would form an ecosystem where Coherent’s reliability and capacity make the startups’ ideas commercially viable.
Photonics in the data center is not a zero-sum race between veterans and startups. It is a layered transformation, where each company plays a distinct and necessary role. Coherent commands the outer layer — the one already generating revenue and scale. Its transceivers move light across racks, linking GPUs, switches, and servers with proven, manufacturable technology. That foundation is what allows the next generation of optical innovation to take shape.
Investors sometimes imagine that disruptive startups immediately unseat incumbents. In photonics, maturity is not a handicap — it’s a prerequisite. Coherent has the wafer fabs, module lines, and hyperscaler relationships to deliver real volume. The startups could extend the addressable market, but Coherent’s technology remains central to every architecture the company aims to build upon.
Coherent should not be seen as a transitional supplier but as the platform company around which photonic computing grows. As AI clusters demand more light at every layer — from rack to package — Coherent’s share of that energy pathway increases. The startups illustrate what’s possible; Coherent ensures it happens.
As light enters the data center — step by step, layer by layer — Coherent is the most experienced and scaled name in the field. It carries the light forward.