Eliyan raises $145M at $1B valuation for chip interconnects for AI networking

Eliyan, a maker of chiplet interconnect technologies for AI, raised $145 million at a $1 billion valuation, led by Seligman Ventures with participation from Cisco Investments and Lumentum. The funding will expand the company into electro-optical interconnect technology for AI infrastructure, addressing bandwidth, power efficiency, and reliability demands in next-generation AI systems.

Become a member of GB MAX to gain exclusive access to the industry and to the most influential global B2B leadership community in the business of gaming, entertainment, and tech. Join now https://go.gamesbeat.com/gb-max/ and also get a VIP ticket to GamesBeat Next Nov 2-3, SF . Eliyan https://www.eliyan.com , a maker of connectivity technologies enabling next-generation AI, said it raised $145 million at a $1 billion valuation. Patrick Soheili, cofounder of Eliyan, said in an interview with me that the funding positions the company to expand into the rapidly growing electro-optical interconnect technology market for AI infrastructure. I last covered them in 2024 https://venturebeat.com/business/eliyan-raises-60m-for-chiplet-interconnects-that-speed-up-ai-chips when they raised $60 million. And it reinforces growing industry confidence in Eliyan’s differentiated approach and the company’s progress toward commercializing a diverse range of electrical and optical AI infrastructure architectures to meet unprecedented performance scaling, bandwidth density, high link reliability, and power efficiency requirements. Seligman Ventures led the unicorn round, with participation from additional new and existing investors, including two new strategic investors: Cisco Investments and Lumentum. The latest addition to Eliyan’s strategic investor list is further testament to Eliyan’s technology fit at major hyperscalers, chip developers, memory suppliers, and optical module providers. As part of the financing, veteran technology investor and operator Umesh Padval, managing partner of Seligman Ventures, will join Eliyan’s board. Die-to-die, chip-to-chip, and rack-to-rack Eliyan makes chiplets, or small chips that can be embedded within larger monolithic silicon chips that are connected with fast networking technologies. These chiplet interconnect technologies are designed to address the increasingly complex connectivity demands of next-generation AI systems spanning advanced package-level integration, multi-package scale-up architectures, memory expansion, and emerging high-bandwidth infrastructure fabrics. Soheili likened the processors, memory and other components on a chip to major features on a semiconductor chip, while Eliyan’s SERDES networking chiplets are like the roads that keep electrical traffic flowing within the chip. The same kind of chips can be used to connect server trays to each other or connect racks to other server racks in a data center. They could also possibly be used to connect data centers to each other, Soheili said. “We’re effectively a connectivity company. We can bundle ourselves up with optical engines and do rack-to-rack connections with optical technology, or we can build AECs, which is basically an electric-based module, so you can go across many different connectivity applications and potentially find us,” Soheili said. “If you don’t have good, high-efficient roads to connect them all together, then it’s slower. Sometimes you’re connecting buildings. Sometimes you’re going across town,” he said. “We can build all those roads, if you will.” But he said the main purpose is scaling-in making servers more efficient and scaling-up internet servers making servers more capable , rather than focusing on networking for scaling out to more servers or larger data centers. As AI infrastructure scales from single accelerators in the form of chips to rack-scale and cluster-scale systems the size of cabinets or rows of cabinents, interconnect becomes more important. It helps relieve the bottleneck stress in areas like bandwidth, memory capacity, and power efficiency — which are increasingly becoming primary constraints on system performance. Technology Platform: Electro-Optical Interconnect Architecture Eliyan’s technologies are designed to eliminate these bottlenecks by enabling efficient communication across compute, memory, and electro-optical domains. The company’s silicon-proven NuLink PHYs and NuGear chiplet families provide scalable connectivity solutions optimized to address the growing challenges of I/O, memory, and power “walls” in AI scale-up and scale-in networks. The company’s expansion into optical interconnect will complement its existing strengths in electrical interconnect as it supports the industry transition toward increasingly heterogeneous, co-packaged, and bandwidth-intensive compute fabrics. Together, these capabilities position Eliyan’s platform to span the full continuum of AI infrastructure connectivity from intra-package electrical interconnects to system-level optical fabrics supporting the industry’s transition toward larger, more distributed, and more power-efficient AI compute systems. “Our expertise is building interconnects, SERDES serializer, deserializer chips , if you will. We’ve been doing that as a team, effectively for the probably 20 or 25 years, including at the startup Aquantia, which went public and later was purchased by Marvell. Eliyan spun out from Marvell with the original team. “So we’ve been developing and working on SERDES for a long time. We’ve done a bunch of chips and chiplet type of things back in the Aquantia days. We continued that work at Marvell, and at Eliyan, we’re basically focused on addressing the memory, power and I/O walls,” Soheili said. “For the memory wall, we’ve got products effectively that sit on the the base die of a custom HBM, which now has gone logic.” He added, “It’s very high speed, very low power density, so it doesn’t melt the DRAM and sits on top of it, and enough bandwidth to drive effectively to HBM4Es in a daisy chain or back-to-back way, and we’ve got some unique patented architectures that effectively take the input and routes it onto the custom HBM and gives you a second port on the other side.” The tech allows a lot more high-bandwidth memory in a system. He said Eliyan can build 30 chiplets that fit inside of a graphics processing unit GPU or central processing unit CPU . All of that is concentrated on creating enough networking, bandwidth and power efficiency. Eliyan’s tech is useful with chip-to-chip interconnections, die-to-die or even rack-to-rack. So two CPUs or GPUs sitting next to each other can communicate so fast on the data network that they are essentially functioning as one CPU or one GPU. The financing “We raised the money effectively to bolster the product line,” Soheili said. He said the result will be more product announcements coming. The company refashions its designs so that its chips can work with any ecosystem or be manufactured on a variety of manufacturing processes. I asked what was driving the valuation of the company forward. Soheili said the number and quality of strategic investors has helped, as many of the biggest chip manufacturers are now investors. There are lots of highways for chip data to flow. They can flow from chip to chip, or from module to module, or connect via wires from server to another server, or from one rack to another rack. At the end points for these cables, there are often SERDES chips, either with electrical connections or optical connections. Soheili said the company focuses on scale-in and scale-up, but not as much in scale-out, where Eliyan has more development work to do in the future. “We think the market appetite and the timing for scale up and scale in is much better and much more immediate,” he said. The financing will be used to expand Eliyan’s development efforts of its next-generation interconnect technologies and products, expand ecosystem partnerships, and scale manufacturing. It follows significant commercial momentum across leading hyperscale, AI accelerator, memory, and infrastructure customers evaluating or deploying Eliyan technologies. “AI infrastructure is undergoing one of the largest architectural transitions in computing history,” said Ramin Farjadrad, CEO of Eliyan, in a statement. “As compute, memory, packaging, and networking become increasingly interconnected, traditional approaches to system connectivity are reaching their limits. Eliyan was founded to address these challenges at the architectural level. This financing, together with the support of leading strategic and financial investors, positions us to accelerate product development, customer adoption, and ecosystem expansion as we help enable the next generation of AI infrastructure.” “We are thrilled to partner with repeat founders Ramin, Patrick, and Syrus as they build a leader in high-speed interconnect technology for AI infrastructure,” said Umesh Padval, managing partner at Seligman Ventures, in a statement. “As AI systems continue to scale, efficient electrical and optical interconnects that enable future scale-up and scale-out networks between compute and memory are becoming the performance-defining technologies of modern AI data centers.” Padval added, “Eliyan has developed a differentiated architecture to address this challenge, combining deep semiconductor expertise and a scalable approach to the evolving needs of next-generation AI systems. The team’s exceptional technical depth, portfolio of more than 100 patents, marquee customer base, and validation from 12 leading strategic investors made our decision to lead this financing compelling. I look forward to bringing my board, semiconductor, and operating experience to help the company scale alongside the founding team.” Santa Clara, California-based Eliyan has raised close to $250 million to date and it has more than 100 employees. Competitors include Nvidia with its NVLink technology, AMD with Pensando, and Broadcom. But sometimes Eliyan is also a supplier to these companies too. The best results are based on power, performance and area, Soheili said. It depends on what customers require, based on the applications they use the most. One size does not fit all. “If their internal team didn’t have the bandwidth or specific skill to do things, they end up being a customer,” Soheili said.