As AI processors become assemblies of GPUs, memory stacks, and specialized chiplets, the optical link is no longer just a data center networking component. It’s becoming part of advanced packaging. CEA-Leti, Scintil Photonics, and NcodiN are each working on a different stage of that transition, from multi-wavelength laser sources for co-packaged optics to nanolasers designed for photonic interposers.
Historically, silicon photonics in data centers has been associated with the front panel of the rack, said Eléonore Hardy, partnership manager at CEA-Leti. Electronics inside the rack perform the computing, while optical transceivers convert electrical signals into photons that can travel through fiber across the data center or beyond.
Hardy said the interconnects now drawing attention are “closer to the chiplets, closer to the GPU, CPU, and memories.” These links are inside the rack, inside advanced packages, and across interposers. Today, many still use copper. But, she said, they are “starting to switch to being also the domain of the photons.”
That progression runs from front-panel transceivers to optical engines inside the rack, co-packaged optics near high-bandwidth electronics and, eventually, photonic interposers that route data among chiplets inside a 3D package. “You get inside the 3D package,” Hardy told EE Times.
View All That’s where the distinction between Scintil Photonics and NcodiN becomes useful. Scintil is industrializing a heterogeneous integration process for single-chip photonic components, including DWDM laser sources that sit outside the package and supply light to co-packaged optical engines. NcodiN, a younger company, is working on nanolasers that could be placed directly in a photonic layer inside the package.
Two routes inward
Yannick Paillard, chief commercial officer at Scintil, described the company’s LEAF Light product as “an integrated photonic multi-wavelength laser source.” It sits inside an external laser small form-factor pluggable (ELSFP) module, usually plugged into the front panel or switch of a GPU system, where it supplies DWDM light to co-packaged optics.
On the die, Scintil integrates distributed feedback lasers, multiplexing stages that combine different wavelengths into a single DWDM source, four optical output ports, monitoring photodiodes, and an on-die frequency reference. The frequency reference is used to keep the wavelengths aligned to the grid customers require.
According to Paillard, the reason for DWDM is that AI systems need a “slow and wide” approach. Instead of simply increasing baud rate, which raises power consumption, traffic can be split across multiple lower-rate channels. “If you have multiple colors, you need multiple lasers to provide these different colors,” he told EE Times.
For Scintil, that creates the need for an integrated laser source that can be manufactured in volume. Paillard said the company is sampling now, with an evaluation kit for customers to test the laser source and connect it to co-packaged optics. The company expects hundreds of samples in 2026, thousands in 2027, a production ramp in 2028, and full production in 2029. Scintil is targeting the connections that allow many GPUs to work together as one larger computing system. Optical links are already widely used to move data across longer distances within and between data centers. Paillard said the next transition is in shorter GPU-to-GPU links inside large AI systems, where copper is still widely used but may struggle as bandwidth rises.
“What we address is the GPU-to-GPU interconnect, what is called scale-up,” he said. “This scale-up is today in copper.” Paillard expects the shift toward optics to occur around 2028 and 2029.
NcodiN is aiming at a shorter link. Francesco Manegatti, CEO of NcodiN, said the company is not targeting scale-out or scale-up networking. Its target is “scale-in”: communication within the package, including links between GPUs and high-bandwidth memory in a single processor.
“We’re actually in the scale-in,” Manegatti told EE Times. “Within the package.”
NcodiN’s core technology is a nanolaser integrated on silicon. Unlike external laser modules or more conventional integrated DFB lasers, NcodiN’s approach uses much smaller lasers that can be densely integrated for low-power, high-bandwidth-density links inside the package.
The company’s target use case is chiplet-to-chiplet communication. Instead of fiber, the links use waveguides. Manegatti said the goal is to allow customers to package more GPU and memory dies in a single system, helping shift the constraint from data movement back toward compute.
NcodiN said its platform operates with a “wide and fast enough” approach, with many parallel channels transferring data bidirectionally. The company claims a bandwidth density of roughly 40 Tbit/s per millimeter, optical-link power of about 0.15 pJ/bit, and integration capacity of about 5,000 nanolasers per square millimeter.
Manegatti does not expect optics to replace copper everywhere. For very short links, copper may remain sufficient. But as chiplet systems grow toward centimeter-scale and eventually wafer-scale distances, he said optics becomes more compelling.
“We use copper when we can and optics where we must,” he said.
The hard part is making it manufacturable
The common challenge behind both approaches is not simply designing a laser. It’s making that laser fit into a semiconductor manufacturing flow.
Silicon photonics can route light efficiently, but silicon is not well suited to generating the light itself. That’s why companies need III-V materials, such as indium phosphide, for laser sources and optical amplifiers. “Silicon alone is great for routing,” Hardy said. “But III-V material is necessary for a laser source or an optical amplifier.”
The difficulty is bringing those materials into a silicon manufacturing flow. Hardy said III-V materials are often processed on four-inch wafers, while silicon photonics uses 200-mm or 300-mm wafers. Integrating small amounts of III-V material onto larger silicon wafers is critical if photonic devices are to reach the volumes and costs required by data center markets.
CEA-Leti’s answer is heterogeneous integration, particularly direct bonding. The institute places III-V material only where it is needed on a silicon photonics wafer, rather than covering the entire wafer. The bonding is molecular, without paste or glue. The surfaces are prepared so that the materials behave as though they are one piece.
According to Hardy, fabless companies such as Scintil and NcodiN may have promising devices, but they need a path to integrate those devices into foundry-compatible processes. “We are the bridge between those guys and the volume foundries,” she said.
The maturity levels are different. Scintil is a Leti spinoff and has licensed several patents from the institute. Hardy said Leti also develops processes that Scintil can take to volume foundries as it builds its supply chain. Scintil announced in February that it ported its heterogeneous integration process to Tower Semiconductor’s silicon photonics platform and is now running on 200-mm production lines.
NcodiN, by contrast, is a CNRS spinoff working with Leti on industrialization. Manegatti said the collaboration is intended to move NcodiN’s technology from “lab to fab” by reproducing it on 300-mm wafers in a CMOS pilot line. He expects first wafers with integrated lasers by the end of the first quarter of 2027.
Several barriers remain. Hardy said system architects have worked for decades with electrical interconnects and will use copper as long as they can. She also pointed to the lack of standards, which makes integration difficult when many companies propose different approaches. Thermal behavior is another issue as lasers move closer to high-power GPUs, and packaging cost remains a significant challenge.
Paillard said Scintil is still generating reliability data. The company receives wafer batches every month, assembles devices, and puts them through high-temperature operating-life testing to accelerate aging and produce data for customers.
“Do we have reliability data now? No, because we haven’t produced enough volume for this,” Paillard said. “That’s what we are currently doing.”
NcodiN faces the same industrial test at an earlier stage. Manegatti said manufacturability, yield, reliability, and cost efficiency must all be validated before volume production.
The direction of travel is becoming clear. Photonics is moving from the edge of the rack toward the package, the interposer, and the chiplet fabric. But the decisive question isn’t only whether tiny lasers can be placed closer to GPUs. It’s whether they can be manufactured, tested, and packaged there at semiconductor scale.
Read also:
CEA-Leti CEO: AI’s Real Bottleneck Is Architecture AI’s growth is hitting an architectural wall, not just compute—discover why integration trumps raw power in this exclusive interview.
AI Data Centers Push Silicon Photonics Toward 300-mm Scale AI data centers are torching copper’s reign as ST pushes 300-mm silicon photonics for faster, denser optical links.