Networking Watch · October 13, 2026
Cisco and NVIDIA propose a 400G-per-lane optical module, Arista and Cisco push Ethernet into the rack, and Meta builds its own 102.4T switch: AI networking is being rebuilt around open standards.
Core Research Question
Is networking becoming the main battleground of AI infrastructure, and will open Ethernet standards displace single-vendor interconnects at each scale?
1. Cisco and NVIDIA Target AI Networking Bottlenecks With a New Optical Standard
The development. At the OCP Global Summit on October 12, NVIDIA and Cisco announced plans for Ultra-Dense Small Form Factor Pluggable (USFP), a pluggable optical module designed to support 400 Gb/s-per-lane electrical signaling. The first planned modules pack 16 full-duplex lanes into an OSFP-sized footprint for switches and SuperNICs, with room to grow to 32 lanes or more. The companies say this reaches up to 12.8 Tb/s of capacity, 4 times the bandwidth density and about 1.3 times the power efficiency of 200G-per-lane pluggables, and supports up to 100 watts without a dedicated cooling circuit in the module. These are vendor figures, and no availability date was given.
Companies and technologies. USFP uses a new two-dimensional array connector with a two-step insertion design to improve signal integrity and reduce contact wear. NVIDIA and Cisco are working with TE Connectivity and Amphenol on cages, connectors and modules, and plan an open specification through a community-led multi-source agreement. By our arithmetic, 16 lanes at 400G is 6.4 Tb/s per direction, so the 12.8 Tb/s figure counts both directions; Cisco does not spell this out.
Market context. Cisco's broader OCP announcement the same day included first shipments of G300-powered systems (102.4T), a new 51.2T G302 chip with 1.6T optics planned to ship before the end of 2026, and Silicon One P200 systems now in production at Microsoft and Alibaba, per customer statements in the release. The USFP step follows the 1.6T interoperability work the Ethernet Alliance demonstrated in September.
Why it matters: Switch capacity is rising faster than front-panel optics, so bandwidth per port and per watt now limit how many accelerators a fabric can connect without adding tiers. Our analysis: USFP is a bet that pluggable optics, not co-packaged optics, can carry the next speed step.
What to watch: The multi-source agreement, a first availability date, sampling from TE and Amphenol, and whether Broadcom, Arista, or optics makers back USFP.
2. Open Ethernet Moves Inside the Rack: Arista, Cisco and the Fight With Proprietary Interconnects
The development. Arista Networks said on October 7 that it is bringing Ethernet to AI scale-up, the short-reach network connecting accelerators inside a rack, where Nvidia's NVLink has dominated. Etherlink SU-144 links up to 144 accelerators in one hop and up to 1,024 across racks, built with AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm. Arista says the racks also support ESUN, NVLink and NVLink Fusion. Its claims on bandwidth and footprint are company figures, and it described a design, not a shipping date. Network World's analysis of the announcement, which was not opened for this edition, frames it as a challenge to Nvidia's scale-up fabric.
Companies and technologies. Cisco said on October 12 that it is a member of OCP's ESUN effort and is working on its 1.0 specification for scale-up. NVIDIA, meanwhile, is pushing Ethernet and open transport on its own terms: Spectrum-6 switches, Multiplane networking, and Falcon support, where Google plans Falcon in future NVIDIA SuperNICs and DPUs per press coverage of NVIDIA's OCP announcements. Meta's engineering post also describes Ethernet scale-up switching for its own MTIA accelerators.
Market context. IDC credits Arista with 18.7% of data-center Ethernet switch revenue and Nvidia with 20.4% in the second quarter, so the two are competitors in the same fast-growing market, now also in scale-up.
Why it matters: As AI infrastructure spreads beyond one accelerator vendor, interoperability and vendor choice become competitive advantages. If Ethernet works in the rack, fabric buyers gain leverage over a single-vendor interconnect.
What to watch: ESUN 1.0, Arista's shipping dates, and whether Nvidia opens NVLink Fusion further or competes with Ethernet inside the rack.
3. Meta's Custom AI Switch Shows How Hyperscalers Are Rebuilding Network Infrastructure
The development. Meta's engineering post on October 12 described IcePack, its first liquid-cooled network rack, populated with IceCube switch blades that deliver 102.4 Tb/s (128 ports of 800G) of non-blocking capacity and run Meta's open-source FBOSS operating system. SDxCentral describes the N-IceCube as built on NVIDIA silicon, but Meta says IceCube comes in two variants, one on Broadcom's Tomahawk6C chip and one on NVIDIA's Spectrum-6. NVIDIA, for its part, says Meta is adopting Spectrum-6 to build larger Vera Rubin AI factories, with multiplane networking designed to scale past 100,000 GPUs using fewer switch tiers.
Companies and technologies. FBOSS uses the OCP Switch Abstraction Interface, which lets software control different vendors' hardware. Meta also said it is co-leading a new OCP effort, the Open Switch Platform Architecture, with Microsoft and Alibaba, has submitted specification v1.0, is open-sourcing its MetaRoCE AI transport, and that vendors can onboard their own switch platforms to FBOSS without Meta engineers in the loop. NVIDIA's multiplane design spreads GPU traffic across up to eight independent network planes so traffic reroutes if a component fails.
Market context. The pattern is merchant silicon from multiple vendors, hyperscaler-controlled software, and open standards at the platform layer. Meta announced an expanded multiyear infrastructure partnership with NVIDIA in February, so the Spectrum-6 choice extends an existing relationship rather than replacing Broadcom.
Why it matters: Hyperscalers are combining custom network software with merchant silicon from competing vendors, and using open standards to keep switching suppliers interchangeable. That limits any single chip vendor's leverage.
What to watch: OSPA adoption beyond Meta, Microsoft and Alibaba, which IceCube variant ships at scale, and whether FBOSS gains outside users.
4. Arelion and Nokia Expand Optical Spectrum on a Live Route
The development. Arelion announced on October 8 that it completed a live field trial with Nokia on its 500-kilometer Amsterdam-to-London route, demonstrating about 21% more usable optical spectrum using Nokia's 1830 Global Express Super C and Super L open line system, CHMQ6 transponder modules and ICE-X 400G and 800G ZR+ coherent pluggable optics.
Companies and technologies. Super C and Super L extend the usable wavelength bands on existing fiber, so more capacity can be carried without new routes. Arelion says the capacity could support its managed optical fiber service connecting data centers and cloud environments. No commercial deployment date or investment value was announced.
Why it matters: Long-haul capacity between data centers is becoming a constraint alongside inside-the-building bandwidth. Extending bands is cheaper than laying fiber, but this is a trial result, not a deployment.
What to watch: Commercial rollout timing, and whether neocloud demand turns into wavelength orders.
5. The AI Network Stack, October 2026
Cisco and Meta both now describe AI networking in three scopes. This table maps the week's announcements:
| Scope | This week |
| Scale-up (inside the rack) | Arista SU-144 and ESUN versus NVLink; Cisco joins ESUN; Meta Ethernet for MTIA. |
| Scale-out (between racks) | Meta IceCube (Tomahawk6C or Spectrum-6), Cisco G300 and G302, NVIDIA Spectrum-6 and Multiplane. |
| Optics | NVIDIA and Cisco USFP at 400G per lane; 1.6T OSFP today. |
| Scale-across (between sites) | Cisco P200 routing in production at Microsoft and Alibaba; Arelion and Nokia extended-band optics. |
The CODEW angle: Each scope has an open-Ethernet challenger and an NVIDIA-aligned incumbent, and the same companies appear on both sides.
The CODEW Angle
Networking is a strategic battleground for AI infrastructure, not a supporting layer, and the competition is now about openness as much as speed.
Cisco and NVIDIA are cooperating on optics while competing in switching; Arista and Cisco are backing open Ethernet for scale-up while NVIDIA extends its own fabrics; Meta is using both Broadcom and NVIDIA silicon under its own software. The common thread is hyperscaler and neocloud demand for vendor choice.
Watch which specifications reach multi-vendor shipping product: USFP, ESUN 1.0 and OSPA are all announced; none is yet in volume. Our October 8 Networking Watch covered the switch-market boom behind this.
Sources
→ Cisco: NVIDIA and Cisco introduce new pluggable optics form factor to scale AI factories
→ Cisco: Cisco advances scalability and openness at OCP 2026
→ Engineering at Meta: OCP Summit 2026, open innovations in scale-up, scale-out and scale-across networks
→ NVIDIA: OCP Global Summit, NVIDIA opens Vera Rubin architecture
→ SDxCentral: Meta builds custom N-IceCube AI network switch using Nvidia Spectrum-6 ASICs
→ StorageReview: NVIDIA and Cisco USFP optics target 400Gb/s per lane
→ TokenPost: Nvidia brings Meta and Google into new AI networking push (secondary)
→ Network World: Arista steps up its Ethernet challenge to Nvidia's AI scale-up fabric (not opened)
→ IDC: Ethernet switch market surges 43.4% to $18.9B in 2Q26
→ Light Reading: Arelion partners with Nokia on live network field trial
→ SDxCentral: Arelion and Nokia increase optical spectrum capacity by more than 20%
→ TelecomTV (Arelion release): Arelion partners with Nokia
The CODEW Stat
400G per lane USFP · 102.4T Meta IceCube blades · 21% more optical spectrum · 3 scopes of AI networking
Performance figures are vendor-reported; USFP and Arista scale-up have no announced ship dates.
Editorial Note
Networking Watch is the fast-moving intelligence layer tracking the companies, infrastructure, standards, products and market developments shaping modern networking. It is distinct from Infrastructure Software Watch, Semiconductor Watch and Cloud Computing Watch.
This edition draws on Cisco, NVIDIA and Meta announcements made at the OCP Global Summit on October 12, 2026, Arista and Arelion releases, IDC data and trade press. Developments occurred October 7 to 12; each is dated in the text.
Educational content only. Not investment advice. Performance, density and power-efficiency figures are vendor-stated and not independently demonstrated; USFP, ESUN 1.0 and OSPA are announced plans or specifications, not shipping products. Some details come from secondary coverage and have not been independently verified by The CODEW.
Reviewed by Erwin Castro
on
Tuesday, October 13, 2026
Rating:
No comments: