Resolight

**Product and the concrete problem it solves.** Resolight is attacking the single most expensive and power-hungry chokepoint in modern AI infrastructure: the network that moves data between GPUs. As training and inference clusters scale to tens of thousands of accelerators, more of a data center's cost, power, and latency budget is consumed not by compute but by the interconnect that stitches the accelerators together. Today that interconnect is fundamentally electrical: optical fiber carries light between racks, but at every switch the light must be converted to electrons, switched by silicon, and converted back to photons (optical-electrical-optical, or OEO, conversion). Each conversion burns power, adds latency, and requires a transceiver that can fail. Resolight's founder describes the status quo vividly — you build a high-speed optical highway and then install a stoplight in the middle of it. Resolight's answer is all-optical switching: a photonic system that keeps data in light form end to end and routes it without converting back to electronics. The company claims this delivers roughly 10x higher bandwidth, about 90% less network power consumption, microsecond-class latency, and up to 10x fewer switches, while eliminating large numbers of transceivers and their associated failure points.

**Core technology and how it actually works.** The heart of Resolight is a photonic processing engine that manipulates data in bulk, directly in the optical domain, rather than routing packets through electrical switch silicon. By avoiding OEO conversion at each hop, the architecture removes the per-switch power and latency tax and collapses the number of discrete switching elements a signal must traverse. Resolight positions its approach as a "scale-everywhere" flat fabric: rather than the rigid, topology-constrained tiers (leaf/spine, scale-up vs. scale-out) that force GPUs into fixed clusters, an all-optical fabric can deliver more uniform bandwidth inside and outside the rack and allow GPU clustering to be defined in software rather than dictated by cabling. Importantly, the company frames itself as compatible with — not opposed to — the industry's emerging optical roadmap, including co-packaged optics (CPO) and linear-drive pluggable optics (LPO); its argument is that even those approaches still ultimately switch traffic electrically, whereas Resolight keeps the switching itself in the optical domain. All-optical switching is often called a "holy grail" of networking precisely because it is hard: prior generations of optical-switching companies struggled with buffering, contention, and ecosystem fit, so the central technical question for Resolight is whether it has genuinely solved the problems that defeated its predecessors.

**Market, customers, and go-to-market.** Resolight is aiming at one of the largest and fastest-growing infrastructure markets of the decade: the interconnect and networking layer of AI "factories" operated by hyperscalers, neoclouds, and large enterprises. As accelerator counts and model sizes explode, networking has become a strategic bottleneck, and power and cooling constraints are now first-order limits on how big a cluster an operator can build. A technology that credibly cuts network power by a large fraction while raising bandwidth speaks directly to the binding constraints of the buildout. The company's go-to-market is business-to-business and infrastructure-level: it must sell into extremely conservative buyers (hyperscale and neocloud operators, and the GPU and switch ecosystems around them) whose qualification cycles are long and whose switching costs are high. Resolight combines an Israeli engineering base (it is registered as an Israeli company and appears in Startup Nation Central's database) with a U.S. commercial presence in Silicon Valley, reflecting the reality that the buyers, standards bodies, and capital for AI networking are concentrated in the United States.

**Traction, funding, and third-party validation.** Resolight emerged from stealth in early 2026, with its public launch and CEO interviews clustered around April 2026 and appearances tied to the OFC optical-networking conference. Publicly disclosed financing includes a seed round from **Entrée Capital** reported in February 2026; some third-party aggregators list substantially larger cumulative totals and a Series A, but those figures are not corroborated by primary sources and should be treated as unverified. The company has also been recognized within Israel's innovation ecosystem (it appears as a grant/program entrant under the Israel Innovation Authority). The strongest external validation is arguably the founder's track record and the fact that credible trade press (SiliconANGLE, Futuriom) took the launch seriously as a genuine attempt to revive all-optical switching for the AI era. The honest calibration: this is an early, largely pre-revenue company whose headline performance numbers are company-stated and not yet independently field-proven, and whose full capitalization and customer roster are not disclosed.

**Founders and team background.** Resolight was founded in May 2024 by **Ofer Shapiro** (co-founder and CEO) and **Raanan Ben Shahar** (co-founder). Shapiro is a serial deep-tech entrepreneur best known as a co-founder of Vidyo, the video-conferencing pioneer; he is credited with inventing multicast/scalable video technology used widely across the industry and holds a large patent portfolio (public profiles cite on the order of 52 patents), and he has been named a World Economic Forum Technology Pioneer and recognized with industry innovation awards. That pedigree matters in a category where the technical risk is high and investor conviction leans heavily on founder credibility. Startup Nation Central lists the company in the 11–50 employee band, consistent with an early, IP-centric photonics team; the depth of the photonic-hardware and manufacturing bench beyond the founders is not fully documented in public sources and warrants direct diligence.

**Competitive dynamics.** Resolight is entering one of the most fiercely contested arenas in computing, against incumbents with enormous resources and startups chasing the same photonic thesis. (1) The dominant incumbents — **Nvidia** (InfiniBand and Spectrum-X, plus a co-packaged-optics roadmap), **Broadcom** (Tomahawk/Jericho merchant switch silicon and CPO), and **Marvell** (custom AI silicon and optical DSP, which acquired photonic-fabric startup Celestial AI) — own the electrical-switching status quo and the customer relationships. (2) **Cisco** and other established networking vendors bring silicon-photonics portfolios and deep enterprise/data-center channels. (3) Photonic-interconnect startups such as **Ayar Labs** (in-package optical I/O) and **Lightmatter** (photonic interconnect and compute) are pursuing adjacent bets on moving optics closer to the compute die. Resolight's differentiation is that it targets the switching itself — keeping traffic optical through the switch, not merely optical between switches — and claims compatibility with the CPO/LPO ecosystem rather than requiring buyers to abandon it. Its levers are: (i) a genuinely distinct all-optical-switching architecture; (ii) a credentialed, patent-rich founding team; (iii) large claimed power/bandwidth advantages that map onto operators' binding constraints; and (iv) an Israeli-plus-Silicon-Valley footprint. The risk is that the incumbents' co-packaged-optics roadmaps mature and consolidate the market before an unproven startup can field qualified, at-scale hardware.

**Defense, security, and resilience dual-use relevance.** The dual-use case here is a real but adjacency-level one, and it should be stated as such: Resolight is building commercial AI-data-center infrastructure, not a fielded defense system, and it has disclosed no defense program or contract. That said, the strategic relevance is genuine. (1) **Sovereign AI compute:** efficient, scalable interconnect is a gating factor for national and allied AI-compute capacity; a technology that materially cuts the power and cost of large clusters is strategically meaningful for resilient, sovereign compute independent of a single vendor's stack. (2) **Defense and government data centers:** military and intelligence organizations run large, power- and cooling-constrained compute estates where the same efficiency and scalability benefits apply. (3) **Photonic substrate adjacencies:** all-optical routing and photonic signal handling are technologies with long-standing relevance to defense communications, low-SWaP resilient networking, and photonic signal processing used in sensing and electronic-warfare research, though translating a data-center product into those domains would be a separate, unproven effort. The appropriate framing is that Resolight is a critical-infrastructure/AI-resilience play whose defense applicability is a credible adjacency, not a demonstrated capability.

**Growth stage, trajectory, and key diligence risks.** Resolight is an early-stage, founder-driven deep-tech company: founded in 2024, out of stealth in 2026, with a disclosed seed round, a small team, first-generation technology, and no publicly documented revenue or named production customers. The trajectory it is betting on — that AI networking's power and bandwidth walls force operators toward all-optical switching — is plausible and well-timed, but the execution risk is severe. Key diligence risks: (1) **category risk** — all-optical switching has a graveyard of prior attempts undone by buffering, contention, and ecosystem misfit, so the burden is on Resolight to prove it has solved them; (2) **incumbent risk** — Nvidia, Broadcom, and Marvell have vastly deeper resources and control the customer relationships and roadmaps; (3) **adoption risk** — hyperscaler/neocloud qualification cycles are long and conservative, and switching costs are high; (4) **claims risk** — the 10x-bandwidth and 90%-power figures are company-stated and not independently field-validated; (5) **capital risk** — disclosed funding is a seed round, and aggregator reports of larger totals are unverified against a capital-intensive photonic-hardware roadmap; and (6) **manufacturing/scale risk** — producing photonic switching hardware at data-center volumes is hard and unproven. The bull case is a credentialed team riding a genuine infrastructure bottleneck with a differentiated architecture; the bear case is an elegant approach that cannot cross the qualification-and-scale chasm against entrenched, better-funded incumbents.