Xsight Labs funding round raises over $300M at $2.8B valuation
Xsight Labs raised more than $300 million to expand programmable Ethernet and DPU chips for AI data center networking.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Xsight Labs funding has passed more than $300 million in a new round that values the Tel Aviv networking chip startup at $2.8 billion, the company said. The capital is meant to push its programmable Ethernet switches and 800-gigabit data processing units further into cloud and AI infrastructure, where networking power draw and latency have become cost constraints.
Fidelity Management & Research Co. led the round. Xsight said Aliya Capital Partners, Atreides Management, Artisan Partners, Battery Ventures, Diagonal Capital, Intel Capital, Key1 Capital, Maverick Capital, Sienna, T. Rowe Price, Union Group and Valor Equity Partners also participated. The company did not disclose revenue, headcount, ownership sold in the round or whether the stated valuation is pre-money or post-money.
Xsight, founded in 2017, says it will use the money to add engineering capacity in Israel, the U.S., Europe and Asia. It also plans to increase manufacturing and supply chain capacity as it seeks orders from what it described as tier-1 customers. The company did not name those prospective customers.
What does Xsight Labs make?
Xsight develops two main data center networking chips: the E1 data processing unit and the X2 Ethernet switch. Programmability is the pitch: customers can alter software behavior for different networking uses rather than buying fixed-function silicon for every workload.
The E1 DPU combines two 400G ports with 64 Arm Neoverse N2 cores on a 5-nanometer process, 32 megabytes of shared cache, PCIe Gen5 connectivity and line-rate encryption engines, according to Xsight. The company says the chip places those 64 cores directly in the packet path so routing, telemetry and packet inspection can run across all packets instead of being handled by a smaller exception engine.
Xsight claims that design delivers four times the performance per watt of earlier-generation designs. It also said the E1 has achieved SONiC-DASH Hero 800G validation, with more than 14 million connections per second and no dropped packets in that benchmark.
The X2 switch provides 12.8 terabits per second of full-duplex bandwidth while using less than 200 watts, according to the company. Xsight also claims first-bit-to-first-bit latency under 700 nanoseconds, 128 lanes of 100G PAM4 SerDes and power consumption 40% below other 12.8 Tbps switches.
Why AI data centers care about Ethernet chips
AI clusters spend much of their power and cooling budget on accelerators. Xsight’s argument is that networking silicon should consume less of that budget, particularly as model training and inference systems scale across more racks.
The company and its backers are also positioning around a larger market shift. Xsight cites expectations that AI data center networking could become a $150 billion market by 2028, and says new generations of its E1 and X2 products will support standards from the Ultra Ethernet Consortium. The consortium’s 1.0 specification, published last year, defines a remote-memory-access transport with multipath routing and congestion control for AI training and inference traffic.
Xsight said its X2 is among the first switches to meet the new Ultra Ethernet standard. That matters if Ethernet continues to gain ground against proprietary interconnects used in many AI accelerator clusters, although buyers will still compare performance, supply assurance and software maturity against incumbent vendors.
The company’s named customer is SpaceX. Xsight said last year that SpaceX selected the X2 as the high-speed networking core for Starlink V3 satellites, citing more than 1 Tbps of fronthaul traffic and about 160 gigabytes per second of uplink capacity. Xsight said the switch was chosen for performance and for operation under radiation and temperature extremes.
Xsight said the E1 and X2 are also deployed by multiple data center operators worldwide in edge and network infrastructure, and that tier-1 hyperscalers are evaluating its products. It did not identify those operators or hyperscalers. The company is entering a field where Cisco, Broadcom and Marvell already have large customer bases and broader product lines, so the new funding buys Xsight time and capacity, not a guaranteed opening.
This story draws on original reporting from SiliconANGLE.
