OCP 2026: The AI Battleground Shifts from Compute to Connectivity

·        Astera Labs - showcased Orion, its  new PCIe-based rack-scale AI reference design. Orion is a modular rack design with production-grade connectivity that lets customers choose their own GPUs and CPUs, rather than locking a rack to one vendor’s platform.

Accelerators plug in as PCIe cards and a 320-lane PCIe switch fabric scales a pod from 8 to 160 cards so that operators can deploy the rack once and swap compute as prices, lead times and workloads change. The scale-up fabric runs on PCIe 6 today and Astera Labs says UALink is on the roadmap. The fabric also adds In-Network Compute and Hypercast for collective traffic.

Orion ships as reference designs that can be bought as a complete rack or one at a time. The Orion-CT compute tray with eight accelerators and CPU, GPU, or XPU options while the switch fabric tray links trays across scale-up and scale-out topologies. An Orion-MT memory tray will become available soon for fabric-attached KV cache offload near the GPU. On the software side, COSMOS Rack Manager and Fabric Manager handle discovery, telemetry, diagnostics, RAS, fabric routing, firmware and lifecycle management.

The target market is inference for neoclouds and enterprises. Astera argues that operators must commit to accelerator capacity long before they know which part will be available or affordable. This means that a rack tied to one platform carries that risk for its whole life. OEM and ODM partners will manufacture and sell the racks, with the first designs due in Q1 2027.

Alongside it, Astera announced Aries 7, a PCIe 7 retimer and optics family, which are expected to sample to select customers in Q4 2026

Figure 1:  Astera Labs Orion PCIe-based Rackscale Infrastructure

·        Upscale - unveiled its proprietary SkyFabriX switch silicon, which delivers 115.2 Tb/s of scale-up capacity. The hardware supports industry open standards such as OCP’s E-SUN and UALoE (UALink over Ethernet). In addition, the company showcased its Token Fabric Architecture, which unifies traditionally distinct scale-up (intra-rack) and scale-out (inter-rack) networks into a single, cohesive infrastructure. Also on display was its Nvidia Spectrum-X-powered 400G, 800G and 1.6T scale-out switches using UpScale’s software.

·        Cisco and NVIDIA - announced plans for an Ultra-Dense Small Form Factor Pluggable (USFP), a new high-density 400Gb/s-per-lane optical module designed to equip the hardware ecosystem for the next generation of AI factories using pluggable transceivers.

This announcement follows a previous announcement a few days earlier about the formation of the 400G Optical MSA to establish interoperable, single wavelength 400G optical link specifications. Members of the optical MSA include Cisco, Nvidia, Broadcom, MACOM and Semtech.

·        Micas Networks - announced three 102.4 Tb/s open Ethernet switch platforms designed for next-generation AI factories. All three are built on Broadcom Tomahawk 6 ASICs and includes a 4RU 800G air-cooled switch, a 2RU 1.6T liquid-cooled switch and a 2RU 1.6T Davisson CPO-based switch. Production of the 2RU liquid-cooled and 4RU air-cooled Tomahawk 6 platforms is scheduled to begin in Q4 2026. Production of the CPO platform is expected to start in  Q1 2027.

·        Ayar Labs – showed off its TeraPHY optical engine paired with SuperNova™ light source inside WiWynn’s AI rack. This demonstration is a significant step forward from the companies' previous rack-level architecture collaboration, advancing to platform integration and functional validation. The system integrates GUC's multi-chip packages (MCPs) with TeraPHY™ optical engines (OEs), powered by the Ayar Labs SuperNova™ light source.

·        Samtec – showcased its PCIe 5.0 over FireFly optical cable assemblies, Si-Fly HD 224 Gbps PAM4 co-packaged and near-chip systems as well as an early proof-of-concept Si-Fly HD demo targeting 448 Gbps PAM4 and PAM6.  However, the centrepiece was the Si-Fly HD CPX architecture, whose SFCM connector mounts directly on the package substrate and mates with Samtec’s cable or compatible optical assemblies. The company ran a 224 Gbps demo with Si-Fly HD CPX copper and optics modules to validate the CPX concept and MSA architecture.

·        LinkTel – showed of its next-generation high-speed optical connectivity portfolio including its 12.8T XPO pluggable due to go into volume production in 2027. Also on show was its 200G/lane 800G and 1.6T transceiver products and its NPO and CPO optical engines using both EML and TFLN configurations.

·        BizLink – showcased numerous integrated solutions across power, high-speed copper, optical interconnects, as well as rack monitoring and liquid cooling. Particularly noteworthy was the collaboration with ams OSRAM combining the latter’s addressable microVCSEL technology with BizLink’s fiber array assembly, optical packaging, testing and high-volume manufacturing, plus co-packaged optics, 800G/1.6T, and XPO optical transceiver interconnects.

·        Point 2 Technologies – hosted demonstrations of its e-Tube™ RF-over-plastic dielectric waveguide technology, highlighting ultra-low-latency, energy-efficient signal transmission designed as a copper-and-optics alternative for AI data factories. Point2's recently closed a $136 million Series B funding round (backed by Arm, LB Investment, Maverick Silicon, NVentures and UM).

·        Genuine Optics – showcased its 3.2T OSFP Module featuring a pioneering 450 Gb/s per lane signalling speed.  Built on a Thin Film Lithium Niobate (TFLN) platform paired with an industry-leading 2nm DSP, it is designed to achieve exceptionally low power consumption for hyperscale AI networks. The company also highlighted its advanced packaging capabilities via previews of its 6.4T NPO Optical Module.

·        Luxshare Tech - showed its Xuanyuan 3.0 AI rack, a full-rack platform combining HVDC power, liquid cooling, NPC/CPC and XPO/NPO technologies, with GPU, switch and CPU nodes plus cable box trays. It uses an orthogonal cabling architecture aimed at signal reach and easier on-site maintenance.

Analyst Viewpoint

·        Most of the industry is rallying behind open Ethernet-based alternatives to Nvidia, with Meta, Microsoft and a wave of startups pushing standards that allow any company’s chips to work together. Companies such as Astera Labs and Upscale have developed systems that allow data center operators to mix chips from different suppliers instead of locking them into one vendor - a direct challenge to Nvidia’s all-in-one approach. Interestingly, Nvidia is hedging its bets as well - it is an investor in Upscale.

·        Pluggables and alternative NPO and CPO-based optics were also a major theme. Integrated optics via NPO/CPO is expected to be used in scale-out switches first before migrating to the scale-up network connecting GPUs to GPUs. As speeds transition from 200G/lane to 400G/lane, NPO/CPO will proliferate at the expense of copper for scale-up and pluggable optics for scale out. However, copper and pluggables are not going away quietly. Both will coexist for some time: copper continuing to be used inside servers with optics elsewhere, while new pluggable form factors such as XPO and USFP means that pluggable optics will continue to be used in scale-out networks for some time.

To learn more, contact Counterpoint Research to access our latest newsletter: Optics, Interconnects & Networking Monthly Supply Chain & Market Pulse Intelligence, Oct 2026.

Counterpoint’s Optics, Interconnects & Networking – Monthly Supply Chain & Market Pulse Intelligence tracks the technologies, vendors and supply-chain developments shaping next-generation AI data center connectivity.

Covering the ecosystem from substrate to rack, each edition analyzes scale-up, scale-out and scale-across interconnects, highlighting technology transitions, supply constraints, competitive dynamics and emerging opportunities.

Combined with Counterpoint’s quantitative trackers, the service helps clients understand how large the market is, who is positioned to capture it, why the market is changing and what to watch next.