AI Economics for Operators
Eliyan's $145M Round Bets the Bottleneck Is Between Chips
Eliyan raised $145M at a $1B valuation to push electro-optical links from promising silicon into production AI systems.
Eliyan raised $145 million at a $1 billion valuation to move its chip-to-chip and rack-to-rack links toward production. The round equals 14.5% of that valuation, a rich wager that the next constraint in AI infrastructure is not another accelerator but the bandwidth and power needed to keep accelerators fed.
The roads between chips now command unicorn money
The Series C announcement says Seligman Ventures led the financing, with Cisco Investments and optical-components maker Lumentum joining the cap table. Eliyan will use the money for electro-optical interconnect development, manufacturing scale, and ecosystem partnerships. The company says customers across hyperscale, accelerator, memory, and infrastructure markets are evaluating or deploying its technology, but it names none of them. That omission is the line between a compelling architecture thesis and a proven production business.
The financing arithmetic is the first useful operator signal. Divide the $145 million round by the $1 billion valuation and the new capital equals 14.5% of the valuation. Those figures come from the company release; GamesBeat’s interview with co-founder Patrick Soheili supplies the competitive context: Nvidia’s NVLink, AMD’s Pensando assets, and Broadcom all occupy nearby territory. Investors are pricing Eliyan as a strategic layer before it discloses revenue, unit volume, or named deployments.
The cap table explains some of the premium. Eliyan says it has 12 strategic investors, and the latest two—Cisco and Lumentum—sit on the networking and optical sides of the problem. In January, AMD, Arm, Coherent, and Meta joined a $50 million strategic financing, alongside existing Samsung and Intel investors. Supplier, customer, and potential competitor categories now overlap. That is useful validation, but strategic checks are not purchase orders.
Eliyan’s pitch spans three distances: die to die, chip to chip, and rack to rack. Its NuLink PHYs and NuGear chiplets address the “memory” and “I/O” walls that appear when accelerators cannot move data quickly or efficiently enough. The company’s technology overview describes electrical links inside packages and systems, with optical links extending the reach. SiliconANGLE frames the round around AI clusters stalling while interconnects fail to feed compute. The operator takeaway is simple: accelerator utilization depends on the fabric around the accelerator.
The historical funding trail shows how much capital that fabric consumes. Eliyan raised $40 million in 2022 and a $60 million Series B in 2024. Add those rounds to the new $145 million Series C and the disclosed named-round total reaches $245 million. That calculation uses the three company announcements and excludes any financing already folded into a later round, so it is best treated as a gross disclosed-round sum rather than net cash. GamesBeat says total funding is “close to $250 million,” which corroborates the order of magnitude.
This is why the story belongs beside Microsoft’s 46% capex ratio. Microsoft can buy tens of billions of dollars of short-lived compute, but poor data movement still turns expensive chips into idle inventory. The thesis extends the earlier warning that AI infrastructure’s fragile supply chain reaches beyond accelerators: packaging, memory, optics, and links can each strand the component with the largest invoice. Builders procuring rack-scale systems should ask for delivered application throughput and energy per useful token, not a component benchmark that assumes perfect feeding.
Demand named deployments before designing around the promise
Who should switch? Accelerator designers, memory vendors, and rack builders at the start of a new architecture cycle should benchmark open chiplet and electro-optical options before hardening around one proprietary fabric. Teams buying ordinary cloud instances should not redesign applications around Eliyan today; their leverage is to demand that cloud and system vendors disclose utilization, network oversubscription, and workload-level performance.
The cost is integration, not merely silicon. A new interconnect touches packaging, retimers, optics, thermal design, firmware, board layout, validation, and manufacturing yield. A link that wins on bandwidth can lose on latency, reliability, availability, or engineering schedule. Standards compatibility matters because a system that saves power but creates a single-source dependency can become more expensive over its useful life.
The Series C therefore funds a crossing Eliyan has not yet publicly demonstrated: from silicon-proven components to repeatable volume deployment. GamesBeat notes that the near-term focus is scale-in and scale-up, while scale-out still requires more work. Operators should preserve that distinction. Connecting components inside a system is not the same problem as coordinating thousands of racks across a data center.
What could break the thesis? Nvidia can keep NVLink attractive by bundling the fabric with the accelerators customers already want. AMD and Broadcom can integrate competing links into larger portfolios. Co-packaged optics schedules can slip, manufacturing yields can disappoint, and customers can evaluate without deploying. Better model efficiency could also reduce pressure on the fabric, although rising agent workloads may consume every efficiency gain.
The evidence that would change the verdict is concrete: named production customers, shipping volumes, pJ/bit, bandwidth density, latency, reliability, package yield, and measured accelerator-utilization gains against incumbent fabrics. Eliyan’s release provides more than 100 patents and 12 strategic investors, but patents and cap-table logos cannot substitute for those operating measures. Buyers should request them under NDA before committing a platform generation.
A disciplined pilot compares one representative workload on the incumbent fabric and the proposed Eliyan path. Measure completed jobs per rack, power at the wall, failure recovery, software changes, and total integration engineering. Price the exit route too. If Eliyan improves useful throughput enough to repay integration inside one hardware generation, it has found a durable layer. If the benefit exists only in PHY specifications, the unicorn valuation is financing a roadmap rather than buying operators a result.
For now, the round is a market-power signal. Cisco and Lumentum joining AMD, Arm, Meta, Samsung, and Intel says the industry expects interconnect to capture more of the AI bill of materials. It does not say who wins. The rational operator response is to benchmark the roads before buying more cars—and refuse to confuse strategic investors with production traffic.