What Happened

On July 30, 2026, Israeli networking chip startup Xsight Labs announced a $300 million funding round at a $2.8 billion valuation. The round was led by Fidelity Management & Research Co., with participation from a broad syndicate including Intel Capital, Battery Ventures, T. Rowe Price, Valor Equity Partners, Atreides Management, Maverick Capital, and others.

Founded in 2017 and headquartered in Tel Aviv, Xsight develops two core products: the E1 data processing unit (DPU) and the X2 Ethernet switch. The E1 pairs two 400G ports with 64 Arm-based Neoverse N2 cores on a 5nm process, with PCIe Gen5 connectivity and line-rate encryption. Its distinguishing feature is that all 64 cores sit directly on the data pathway, enabling routing, telemetry, and packet inspection on every packet — rather than offloading to a smaller exception-handling engine. Xsight claims this delivers four times the performance per watt of previous-generation designs.

The X2 switch provides 12.8 Tbps of full-duplex bandwidth at under 200 watts, with first-bit-to-first-bit latency under 700 nanoseconds across 128 lanes of 100G PAM4 SerDes. According to the company, it draws 40% less power than comparable 12.8 Tbps switches and features a reprogrammable data plane for on-the-fly reconfiguration.

Xsight confirmed that SpaceX has selected the X2 as the networking core for Starlink V3 satellites, citing both performance and the chip's ability to operate in radioactive and thermally extreme environments. The company also stated its products are being evaluated by multiple Tier-1 hyperscalers, though it did not name them.

Why It Matters

The AI infrastructure stack has a networking problem. As training clusters scale to tens of thousands of accelerators, the network fabric connecting them becomes both a performance bottleneck and a power-consumption liability. In AI data center racks, accelerator chips consume the bulk of the power budget — every watt spent on networking is overhead that directly reduces the compute density achievable per rack.

Xsight's value proposition targets this exact pressure point: lower power consumption per terabit of switching, with programmability that lets customers adapt the data plane to their specific workloads rather than relying on fixed-function silicon. The E1 has achieved SONiC-DASH Hero 800G validation, sustaining over 14 million connections per second with zero dropped packets — a benchmark that matters for hyperscalers standardizing on open networking software.

The broader strategic question is whether Ethernet can close the gap with proprietary interconnects like NVIDIA's InfiniBand in AI training environments. The Ultra Ethernet Consortium published its 1.0 specification last year, defining a remote-memory-access transport with multipath routing and congestion control specifically designed for AI training and inference traffic. Xsight says the X2 is among the first switches to meet this standard. If the UEC specification gains adoption across hyperscalers, it could create a viable open alternative to vendor-locked interconnects — and Xsight is positioning itself as a hardware beneficiary of that shift.

Investors are backing this thesis with real money: the round implies a $2.8B valuation on a company whose most notable public customer is SpaceX (not a hyperscaler), and whose Tier-1 hyperscaler evaluations are still ongoing. The bet is that the AI data center networking market will balloon to $150 billion by 2028 and that Xsight's power-efficiency advantage will win meaningful share.

Who Is Affected

Hyperscale data center operators are the primary audience. Xsight's products are in active evaluation with Tier-1 players, and if those evaluations convert to design wins, it would signal that open Ethernet is gaining ground over proprietary interconnects in the most demanding AI training environments.

GPU cloud providers and AI infrastructure startups building their own training clusters should monitor Ethernet-based fabric as a potential cost and power-efficiency lever. The performance-per-watt argument is particularly relevant for operators facing power-constrained data center footprints.

Enterprise network hardware buyers running large-scale inference or data processing workloads may benefit from increased competition in the switching and DPU market, which is currently dominated by Broadcom, Cisco, and Marvell. A viable programmable Ethernet alternative could apply pricing pressure across the board.

Strategic Implications

For AI startup founders: The Ethernet-vs-InfiniBand decision for cluster networking is becoming a genuine strategic choice rather than a default. If you're building infrastructure, track Ultra Ethernet Consortium adoption milestones and hyperscaler design wins. If UEC gains traction, your networking cost per token could drop meaningfully within 12-18 months as open Ethernet alternatives scale.

For developers/operators building with AI APIs: This is an infrastructure-layer development with no immediate impact on API-based workflows. However, if you operate your own inference clusters, programmable Ethernet switches with sub-700ns latency and lower power draw could reduce infrastructure overhead. Watch for SONiC-DASH compatibility as a benchmark signal for whether these chips fit into open networking stacks.

For non-technical business owners evaluating AI tools: No immediate action required. This is a deep infrastructure play that may eventually reduce the cost of AI services as networking competition increases. The broader signal worth noting: investors are projecting $150B in AI networking spend by 2028, which means the hardware layer of AI is still in rapid flux — and cost structures for AI services are not yet stable.

What to Watch Next

Monitor for announcements of Xsight's Tier-1 hyperscaler design wins — these would be the strongest signal that open Ethernet is gaining ground in AI training environments. Also watch for Ultra Ethernet Consortium adoption milestones from other hardware vendors, as broader ecosystem support would validate the standard's viability against proprietary interconnects.

Frequently Asked Questions

Q: What makes Xsight Labs' networking chips different from existing solutions?

A: Xsight's chips emphasize power efficiency and programmability. The X2 switch delivers 12.8 Tbps at under 200 watts — reportedly 40% less power than comparable switches — and features a reprogrammable data plane. The E1 DPU places all 64 Arm cores directly on the data pathway for per-packet processing, which Xsight claims delivers 4x performance per watt over previous-generation designs.

Q: Will Ethernet replace InfiniBand for AI training clusters?

A: It's too early to say definitively, but the Ultra Ethernet Consortium's 1.0 specification — designed specifically for AI training and inference traffic — is a meaningful step toward making Ethernet competitive with proprietary interconnects. Xsight's X2 is among the first switches to meet the UEC standard, and hyperscaler evaluations are ongoing. Adoption will depend on whether open Ethernet can match InfiniBand's latency and congestion management at scale.