7 segments · 67 companies · as of Sep 18, 2026
Switching & Fabrics
Switches, adapters and software that connect AI accelerators
Where it sits
Fed by: Chip ManufacturingFeeds: Servers & Systems
AI clusters exchange model parameters and intermediate results across thousands of accelerators. Ethernet, InfiniBand and scale-up fabrics combine switching chips, systems, network adapters and software to keep those accelerators working as a coordinated machine.
Why now
102.4T switching and 800G/1.6T ports are entering deployments while NVLink Fusion, UALink and scale-up Ethernet broaden rack designs. Congestion control, cable reach and validation increasingly determine usable cluster performance.
Branded switch systems
Complete leaf, spine and AI back-end switches combine hardware, congestion control and support. Buyers choose Ethernet, InfiniBand or specialist fabrics around accelerator software and operating experience.
Spectrum-X Ethernet and Quantum-X800 InfiniBand switches for GPU cluster back-end networks
leaderEtherlink and 7800R4 systems with up to 576 800G ports; EOS controls cloud AI Ethernet fabrics
leaderNexus and 8000 AI switches use Silicon One; G300-based 102.4T systems target second-half 2026 availability
majorVia Juniper QFX switches and Apstra, supplies Ethernet AI fabrics after the July 2025 acquisition
majorPowerSwitch Z9864F-ON 800G systems form Ethernet back-end fabrics for AI Factory deployments
major7220 IXR fixed and 7250 IXR modular switches pair high-speed AI fabrics with SR Linux
challengerCloudEngine Ethernet switches provide RoCE fabrics for domestic AI and Ascend deployments
majorVia H3C, S12500AI distributed chassis provide 400G/800G ports for Chinese AI data centers
majorRG-S9910-128OC2VS provides 128 800G ports for Chinese cloud and intelligent-computing fabrics
challengerNebula RoCE fabric includes a 576-port 800GE chassis for Chinese intelligent-computing clusters
challengerCN5000 Omni-Path 400G edge and director switches serve AI and HPC clusters through server OEMs
niche
Sources: idc.com, corporate.celestica.com, nvidia.com, ruijie.com.cn, blogs.cisco.com, h3c.com
ODM & white-box switches
Cloud operators and networking brands buy switch platforms from specialist manufacturers. Board design, thermal engineering, firmware qualification and production capacity decide 800G and 1.6T design wins.
DS5000 800G and DS6000/DS6001 1.6T platforms use Broadcom silicon for hyperscale AI networking
leaderVia Edgecore, AIS1600-64O and AIS800-128O open switches use 102.4T Tomahawk 6 silicon
leaderDC9064O offers 64 1.6T ports; also manufactures 51.2T Bailly CPO systems for early hyperscaler deployments
majorQuantaMesh TA064-IXM provides 800G switching; TB064-IXP extends the roadmap to 102.4T CPO
majorS9321-64E 800G and Tomahawk 6 1.6T liquid-cooled switches support disaggregated AI fabrics
challenger51.2T Bailly CPO switch systems have entered limited manufacturing and hyperscaler validation
challengerBroadcom-based 800G switches and a liquid-cooled 1.6T design target AI network OEM customers
challenger800G switch and optical transceiver platform extends communications ODM work into AI data centers
challengerDemonstrated a 64-port 1.6T Tomahawk 6 switch for open AI racks; customer production ramp is undisclosed
emergingEnterprise Business Group builds cloud switch hardware and integrates networking into AI rack manufacturing
nicheSSE-T8164 open Ethernet switch supplies 64 800G ports with SONiC for GPU cluster fabrics
nicheX732Q-T programmable 32-port 400G switches use Intel Tofino2 silicon and its own CX-NOS
nicheVia Ingrasys, develops a Spectrum-4 TOR switch with embedded storage and 400G/800G links, targeting 2027 availability
emerging
Sources: corporate.celestica.com, trendforce.com, deltaww.com, wiwynn.com, supermicro.com, cloudswit.ch
Ethernet switch silicon
Merchant and captive ASICs determine switch bandwidth, radix and congestion behavior. AI buyers need packet delivery that sustains synchronized workloads, with 102.4T generations reducing the number of fabric tiers.
Tomahawk 6 102.4T, Tomahawk Ultra and Jericho 4 ASICs span AI scale-out, scale-up and scale-across
leaderCaptive Spectrum Ethernet and Quantum InfiniBand ASICs anchor its end-to-end AI network platforms
majorSilicon One G200/G300 switching and P200 routing silicon serve branded systems and hyperscale designs
majorTeralynx 10 programmable 51.2T switch ASIC supports 64 800G ports and 112G long-reach SerDes
challengerX2 programmable Ethernet ASIC and E1 DPU provide merchant silicon for disaggregated cloud networks
challengerDomestic Ethernet ASIC portfolio reaches 25.6T switching and 800G ports for Chinese data centers
challengerCaptive Ethernet Network Processor chips implement forwarding and congestion control in CloudEngine AI switches
majorAnnounced ICN Switch 1.0, a captive 25.6T interconnect chip linking up to 64 Zhenwu accelerators
emerging
Sources: broadcom.com, marvell.com, xsightlabs.com, support.huawei.com, alibabacloud.com, static.cninfo.com.cn
NICs, SuperNICs & DPUs
Adapters terminate fabric links and offload data movement, security and virtualization. GPU back-end NICs emphasize RDMA and congestion handling, while hyperscalers also build captive infrastructure processors.
ConnectX-8 and BlueField-3 serve AI fabrics; ConnectX-9 and Grace-based BlueField-4 extend the Rubin-generation roadmap
leaderThor Ultra 800G AI NIC supports UEC and PCIe Gen6; announced sampling, with volume timing undisclosed
challengerPensando Pollara 400 serves Ethernet AI fabrics; Vulcano 800 extends the Helios generation roadmap
challengerOCTEON DPUs offload network, security and storage processing in cloud infrastructure
majorEthernet 800-series adapters and infrastructure processing units serve server front-end networks
majorAnnapurna-designed Nitro v6 networking cards provide up to 400 Gbps per card inside AWS servers
majorAzure Boost offloads host networking and storage for Azure VMs, including H200 and GB300 families
nicheE1 800G programmable DPU combines Arm Neoverse N2 cores, PCIe Gen5 and inline cryptography
challengerNT400 programmable SmartNICs offer deterministic 400G processing for cloud infrastructure workloads
nicheCN5000 single- and dual-port 400G SuperNICs connect AI and HPC hosts to Omni-Path fabrics
nicheGoogle Titanium IPUs, co-developed with Intel, offload networking and storage in C3 and later cloud instances
majorVia Alibaba Cloud, CIPU offloads virtual networking and enables eRDMA in commercially available ECS instances
majorTaihang DPU 2.0 supplies 200G network, storage and security offload for Baidu AI Cloud infrastructure
nicheThunderFjordB Agilex FPGA SmartNIC supplies two 400G ports for programmable cloud packet processing
nicheT7 unified-wire 400G SmartNICs launched in July 2026; programmable DPU production is scheduled for December
niche
Sources: blogs.nvidia.com, cloud.google.com, alibabacloud.com, cloud.baidu.com, silicom-usa.com, chelsio.com
Scale-up fabrics & interface IP
Scale-up links expose tightly coupled accelerator communication across trays and racks. Proprietary NVLink competes with open UALink and Ethernet approaches; several merchant switches remain in early ramps or development.
NVLink 6 and NVLink Fusion connect rack-scale GPUs and partner XPUs through the NVLink ecosystem
leaderTomahawk Ultra provides low-latency Ethernet switching; SUE-T transport and ESUN extend the scale-up architecture
majorHelios uses UALoE Ethernet scale-up links across MI455X accelerators; this is not proof of shipping UALink ASICs
challengerScorpio X-Series 320-lane switches began shipments in 2026; UALink silicon volume is not separately established
challengerXConn PCIe/CXL switches ship within a broader NVLink Fusion and UALink/ESUN development roadmap
challengerUnifiedBus switches connect 384 Ascend 910C NPUs and 192 CPUs inside CloudMatrix384 supernodes
majorSkyHammer UALink/ESUN switch is under development; no production customer socket has been verified
emergingAdopts NVLink Fusion as a design foundation for custom XPUs connected to Nvidia rack-scale systems
emergingUALink controller and 224G PHY IP enable custom accelerator scale-up links and interoperability
majorUALink controller, PHY and verification IP support 100G/200G lanes and memory-semantic operations
majorGoogle Ironwood TPU pods use captive ICI links across up to 9,216 chips for tightly coupled model execution
majorICN 1.0 couples Zhenwu accelerator memory and compute through Alibaba-designed switching and protocol logic
emergingTeraPHY optical I/O joined NVLink Fusion; named production accelerator sockets remain unverified
emerging
Sources: amd.com, broadcom.com, ir.asteralabs.com, marvell.com, cloud.google.com, ayarlabs.com
Network OS & fabric software
Network operating systems and controllers configure routes, telemetry and congestion policy. Open software separates switches from silicon vendors, while integrated stacks simplify validation and operations for AI clusters.
EOS and CloudVision coordinate Etherlink fabrics, telemetry and operations across AI Ethernet switches
leaderCumulus Linux, NetQ and UFM operate and monitor Spectrum Ethernet and Quantum InfiniBand fabrics
majorNX-OS and Nexus management automate AI fabrics, including Cisco systems using Spectrum-X silicon
majorJuniper Apstra intent-based automation and Junos operate validated multivendor AI data-center fabrics
majorOriginated SONiC for Azure; contributes the open network OS used across merchant-silicon switches
majorFBOSS controls disaggregated scheduled fabrics and custom OCP switching in its AI training networks
nicheSR Linux and Event-Driven Automation manage 7220 IXR Ethernet leaf-spine fabrics
challengerArcOS runs on Edgecore 800G systems and provides routing, congestion control and fabric automation
challengerNetwork Cloud AI software builds scheduled Ethernet fabrics on Jericho-based white-box switches
challengerCertified SONiC distributions and ONES support qualify and operate 400G/800G merchant-switch platforms
challengerVia IP Infusion, OcNOS supplies commercial network software for disaggregated Ethernet systems
nicheAsterNOS, built on SONiC, operates merchant-silicon AI fabrics with RoCEv2 forwarding and network management
niche
Sources: engineering.fb.com, arrcus.com, drivenets.com, aviznetworks.com, access-company.com, cloudswit.ch
Network test & precision timing
Signal integrity, error-rate and traffic tests validate faster links before deployment. Low-jitter oscillators and clock distribution provide the timing needed by switch ASICs, SerDes and optical modules.
Ixia traffic systems and electrical/optical test platforms validate 800G/1.6T fabrics and SerDes
leaderSpirent TestCenter HSE assets plus optical instruments validate AI Ethernet fabrics and modules
majorVia Tektronix, DPO70000SX scopes analyze high-speed electrical and optical link signal integrity
majorLeCroy oscilloscopes and PCIe analyzers debug high-speed serial links in AI servers and switches
majorMP1900A BERT platform validates PAM4 and PCIe links; optical instruments support module development
majorML7008F-LFT 224G-per-lane BERT and 1.6T compliance fixtures test next-generation modules and hosts
nicheOptical sampling, spectrum and Ethernet test instruments validate modules and data-center fiber links
nicheSiT9501 low-jitter MEMS oscillators and acquired Renesas timing products synchronize AI network links
leaderZL-series network synchronizers and clock distribution support Ethernet switches and optical transport
majorSi5396/Si5397 jitter attenuators provide programmable low-noise reference clocks for optical and Ethernet links
majorLMK jitter-cleaning and clock-distribution devices supply reference timing for network equipment
majorLow-jitter clock generators and distribution ICs synchronize high-speed converters and network interfaces
nicheLow-phase-jitter differential quartz oscillators target SuperNIC, 800G/1.6T and CPO timing
nicheNZ-series 156.25/312.5MHz differential oscillators clock 800G/1.6T transceivers; 625MHz devices are sampling
nicheSG2016EGN/VGN differential crystal oscillators supply reference clocks for 800G optical transceivers
nicheX-series 312.5MHz low-jitter crystal oscillators entered production in 2026 for high-speed optical modules
nicheZNB3000 vector network analyzers characterize insertion loss and signal integrity in high-speed interconnects
nicheVia acquired Rakon, RXO7050M low-noise quartz oscillators supply timing references for switches and routers
niche
Sources: sitime.com, ndk.com, epsondevice.com, ele.kyocera.com, rakon.com, rohde-schwarz.com
Glossary
- Scale-out
- Networking between servers or racks, usually using Ethernet or InfiniBand.
- Scale-up
- Low-latency links that connect accelerators into a tightly coupled compute domain.
- DPU
- A processor that offloads network, storage and security functions from the host CPU.
- RoCE
- Remote direct memory access carried over Ethernet, reducing host involvement in data movement.
- SerDes
- Circuits that convert parallel chip data into high-speed serial electrical lanes.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗