Eliyan’s $145M/ $1B chiplet-memory bet turns NVLink into a testable “standard,” not a monopoly

Eliyan Eliyan (private) announced on July 29, 2026 that it raised a $145M Series C at a $1B valuation to ease bottlenecks in moving data between AI chips in data centers. In parallel, Eliyan’s CEO described the mechanism bluntly: most AI GPUs are limited by how fast they can receive data to process—leading to materially underutilized compute capacity.

Eliyan Eliyan describes NuLink as a die-to-die PHY optimized for high-speed serial connections, with extensions into die-to-memory (D2M) and rack-scale optical fabrics. Crucially for investors tracking Nvidia’s ecosystem lock-in thesis, Eliyan explicitly states NuLink is protocol-agnostic and fully supports industry standards including UCIe and BoW, rather than only a proprietary NVIDIA-style interconnect plane.

Reuters reports Eliyan’s CEO arguing that data movement is the gating factor: GPUs may process data faster than they can receive it, leaving compute waiting. That reframes the “AI memory wall” as a transport-and-protocol problem: if the interconnect between compute tiles and memory tiles cannot sustain the required read/write patterns (or the coherency/translation overhead is too costly), adding more HBM per GPU won’t fully eliminate the idle time.

Eliyan’s announcement and technology page describe electro-optical interconnect, chiplet connectivity, and advanced package-level integration. That touches multiple layers in the AI server value chain: (1) board and optical/photonic transport components, (2) foundry and packaging capacity that can actually assemble die-to-die links at the required bump pitches and power budgets, and (3) downstream hyperscaler and OEM server designs that must validate coherency and end-to-end bandwidth.

Nvidia’s NVLink story has often been framed as a moat because coherent die-to-die fabrics are tightly integrated into platform design choices. Eliyan Eliyan doesn’t disprove that outright. Instead, it attacks the monopoly assumption by positioning a third-party interconnect technology that claims standards support (UCIe/BoW) and targets both die-to-die and die-to-memory bandwidth efficiency—making it possible for system builders to evaluate alternative architectures on common interfaces and packaging regimes.

Key risk: even if the PHY layer claims are strong, interconnect usefulness depends on platform coherency behavior, memory protocol integration, and total system power/thermals. Eliyan’s funding buys time; the market repricing only follows if those integration details work in shipped systems and show clear utilization lift.

To avoid turning this into pure narrative, we anchor the investable takeaway to listed-company fundamentals that relate to the interconnect supply chain. Nvidia’s revenue scale has surged to a TTM level of $253.49B, which increases the economic value of cluster utilization improvements. That makes any validated alternative to proprietary interconnect planes potentially important enough to matter to multiple hardware layers—especially those that bottleneck physical assembly and optical reach.