Alibaba Cloud at OCP 2026: 9 Sessions on Open AI Network & 3 Scaling Paths

From Open Network to AI Infrastructure

The OCP Global Summit 2026 took place in San Jose from October 12–15, attracting over 14,000 industry participants. Alibaba Cloud's intelligent network R&D team delivered nine technical sessions spanning AI network architecture, data center interconnect (DCI), optical transport networks, SONiC open-source software, switch chip evaluation, and automated testing. These practices illustrate Alibaba Cloud's technical layout across different network scales — Scale-up, Scale-out, and Scale-across — and its ongoing exploration of open collaboration to improve network engineering and verification systems.

Three Extension Directions for AI Networks

1. Scale-up: Open Ethernet Toward AI Super-Nodes, Optical Interconnect Driving High-Density Expansion

At OCP 2026, Ethernet-based Scale-up emerged as a key direction for AI networking. Vendors including AMD, Microsoft, Meta, OpenAI, and Google showcased related products and achievements, signaling Ethernet's expansion from traditional data-center Scale-out networks into high-bandwidth, low-latency interconnects inside super-nodes.

The SONiC Scale-Up Working Group, jointly driven by Alibaba Cloud and Microsoft, made significant progress: Ethernet-based Scale-up networks have moved from architecture design and standardization toward device implementation and test verification. Working group members gathered for a milestone photo at the summit.

Alibaba Cloud's TU1-based Ethernet Scale-up switch also debuted at the summit and completed a live power-on demo at the Keysight booth, demonstrating actual device operation. This marks a step from standard promotion to device demonstration and validation, reflecting collaboration among Alibaba Cloud, Microsoft, Broadcom, and Keysight.

As super-node scale grows, traditional copper cabling faces transmission-distance and wiring-density limits, making optical interconnect a critical technology. Alibaba Cloud senior director Yao Zhiping compared pluggable optics, near-package optics (NPO), in-package interconnect variants, and co-packaged optics (CPO), and demonstrated a 102.4T switch based on an IPO (in-package optics) approach. This design routes 512 lanes of 224G high-speed signals from the top of the switch ASIC package. Test results showed signal integrity better than NPO, approaching CPO, while retaining pluggability and multi-vendor collaboration advantages.

This practice explores combining high-density optical interconnect with open switch architecture, providing a new technical option for AI network evolution from 200G to 400G. From Ethernet-based Scale-up device demonstrations to optical interconnect technology exploration, Alibaba Cloud is advancing AI super-node networks toward higher bandwidth, larger scale, and greater openness through open-source community collaboration and product practice.

2. Scale-out: Decoupled Routing Distribution to Support 100k-GPU Clusters

Ultra-large GPU clusters demand greater control-plane scalability. Alibaba Cloud director Ruan Yixing shared a routing architecture for 100k-GPU clusters. In a public cloud serving many tenants, propagating full business routes across the entire fabric could generate tens of millions of BGP path states.

Alibaba Cloud decouples infrastructure routes from business routes: infrastructure routes continue propagating via BGP among upper-layer switches, while business routes are redistributed among access switches by a route-service cluster composed of multiple route reflectors (RRs), using access-switch loopback addresses for recursive resolution. This architecture reduces per-access-switch BGP business peers from hundreds to a few, compressing routing state from tens of millions to four hundred thousand.

Additionally, Alibaba Cloud collaborated with Cisco to introduce BGP Group Graceful Restart, reducing the impact of RR cluster failures or restarts on business traffic. The solution has been deployed on two 100k-GPU clusters and is evolving toward a next-generation HPN Pivot architecture targeting 500k GPUs.

3. Scale-across: Making Cross-Data-Center Networks Adapt to AI Jobs

As single data centers hit power and space limits, cross-region collaboration becomes a key AI infrastructure expansion direction. Alibaba Cloud WAN architecture director Su Yuanchao and Cisco Fellow Clarence Filsfils jointly presented DCI/WAN traffic engineering practices for AI traffic.

Alibaba Cloud builds its wide-area network on the eCore architecture, using multiple independent networks, multiple routing domains, and multi-plane redundancy to limit fault blast radius. A unified SONiC, SRv6 technology stack and white-box devices are employed, with equipment covering 12.8T and 51.2T switching capacities; the backbone uses 64×800G ports to support high-bandwidth cross-DC interconnect.

On traffic engineering, the two parties explored AI-job-aware scheduling: using reserved bits in the RDMA BTH header to carry traffic-engineering identifiers, enabling the network to perform path scheduling at AI-job granularity. Combined with SRv6 Binding SID, cross-DC path intent can be expressed with fewer instructions, adapting to NICs with limited capabilities and providing finer-grained traffic control and performance assurance for AI training.

4. Optical Transport Network: Extending SONiC to OTN

Cross-DC interconnect relies not only on IP networks but also on open, orchestratable optical transport infrastructure. Alibaba Cloud director Ruan Yixing introduced the SONiC for OTN project.

The SONiC-OTN Working Group, with participants including Alibaba, Microsoft, Molex, Nokia, Cisco, and Accelink, leverages the OCP SAI extension framework to define vendor-neutral abstract interfaces for optical amplifiers, variable optical attenuators, and other devices. SONiC's containerized architecture and multi-ASIC design manage multiple optical line cards.

The project proceeds in phases: Phase 1 covers optical amplifiers and variable optical attenuators; Phase 2 targets a complete open line system; Phase 3 plans support for transponders and optical channels, plus a SONiC OTN KVM image runnable without physical hardware.

Alibaba Cloud also shared AI-DCI optical network practices: single-chassis single-fiber 25.6 Tb/s FST full-spectrum transmission, hollow-core fiber (HCF) expanding data-center site selection, lossless traffic scheduling via IP-optical layer coordination (LD/RD), and Echo fiber active sensing technology — pushing optical transport toward openness and intelligent O&M.

Open-Source Engineering & Test Verification: Building a Reliable Delivery System

Rapid AI network evolution demands not only architectural innovation but also higher development efficiency, hardware verification, and system reliability for open-source software. At the summit, Alibaba Cloud, Microsoft, Google, Keysight, and other partners shared multiple engineering practices.

1. Bazel: Boosting SONiC Build Efficiency

Alibaba Cloud senior technical expert Shi Shao and Google's Yunang Chen presented the Bazel-based reconstruction of the SONiC build system.

Traditional SONiC builds rely on a multi-layer Make system where incremental compilation often redundantly rebuilds many dependencies. With Bazel, orchagent incremental rebuild time dropped from ~6 minutes to 10 seconds. The solution introduces bazel-remote with object storage for cross-team build cache sharing, and uses locked inputs, isolated execution, and deterministic outputs to improve build reproducibility. A phased Make-and-Bazel coexistence migration aims to bring full-image build time under 15 minutes.

2. Scale-up Test Standard: Establishing a Unified Verification Method

Alibaba Cloud senior technical expert Zheng Haiyang and Keysight's Daisy Sun reported progress from the SONiC Scale-Up Test Sub-Working Group.

The test standard combines UEC specifications with hardware-level SAI verification, following a "configure, inject, observe, verify" flow. Test equipment injects LLR (link-layer retry) and CBFC (credit-based flow control) traffic, validated against ASIC hardware counters. The demonstrated platform passed 22 tests covering full-frame line-rate forwarding, low latency, FEC error handling, incast fairness, virtual-channel isolation, and credit limiting. Discovered SAI attribute gaps fed back into specification refinement. Test standards, plans, and results are now open for community review in SONiC.

3. Tier 1 Ranking Test: Vendor-Neutral Switch ASIC Evaluation

Alibaba Cloud senior technical expert Zhu Shenglin introduced the SAI-based Tier 1 Ranking Test framework.

The framework bypasses the not-yet-ready SONiC control plane to directly evaluate switch ASICs in bare-metal environments, covering basic functionality, hardware specifications, control-plane programming performance, and data-plane throughput. By measuring time near SAI calls, it reduces interference from RPC and other overhead on chip programming performance assessment, without modifying vendor SAI code. Currently, the framework includes 420+ basic functional test cases and 60+ specification and performance test cases, hosted in sonic-mgmt and open for community contribution, providing a unified method for cross-vendor switch ASIC comparison.

4. Formal Methods: From Protocol Specs to Automated Verification

Alibaba Cloud senior technical expert Liu Mengqi and Google's Ali Kheradmand presented the SONiC Formal Methods Working Group's open-source infrastructure and latest progress.

Traditional handwritten unit tests suffer from limited behavior coverage and incomplete state verification, often missing bugs triggered only by specific configuration, runtime state, and input combinations. The working group adopts an "executable model + symbolic execution" approach: starting from network protocol and module behavior specifications (HLD/RFC), they write formal models; symbolic execution explores feasible paths, automatically generating test cases executed on real code, providing mathematically systematic behavior coverage guarantees.

The group open-sourced modeling libraries and test frameworks in sonic-formal-infra, and reported progress with the SONiC community on BGP message error handling (RFC 7606), BGP-LS (RFC 9552), and graceful restart (RFC 4724). The roadmap extends formal methods from control plane to SAI and data plane, drawing on Google's P4-BAR experience. Jointly driven by Alibaba Cloud, Google, and Microsoft, the working group turns "from protocol spec to test" verification capabilities and protocol models into reusable open-source infrastructure that can continuously evolve with SONiC.

5. vSONiC: Building a Large-Scale Digital Twin Test Environment

Zhu Shenglin also introduced Alibaba Cloud's vSONiC test system. vSONiC replaces physical ASICs with chip simulators running in KVM, while keeping the rest of the stack — kernel, FRR, orchagent, syncd, and vendor SAI/SDK — identical to production hardware.

Alibaba Cloud built a four-layer test system: single-device community test cases, multi-node SRv6 tests, 1:1 production environment simulation, and a full digital twin including controllers and monitoring components. Currently ~80% of Alibaba switch software tests run on vSONiC, comprising 80+ Jenkins jobs totaling over 200 hours of test time, with more than 100,000 test scheduling runs. The next step is to introduce AI-driven automated diagnosis, repair, and regression testing, collaborating with the SONiC community to achieve an end-to-end one-hour automated repair loop.

Open Collaboration, Co-Building the AI Network Foundation

From Ethernet-based Scale-up to Scale-out routing decoupling, to Scale-across traffic engineering and open optical transport, Alibaba Cloud demonstrated technical practices for different network scales at OCP 2026. Simultaneously, work on Bazel, Ranking Test, Scale-up test standards, formal methods, and vSONiC further strengthens the open network's software engineering and verification ecosystem.

These directions serve actual large-scale AI infrastructure needs while being codified as reusable community assets through open projects like SONiC, SAI, and OCP. Through collaboration with Google, Microsoft, Cisco, Broadcom, Keysight, and other industry partners, Alibaba Cloud continues to drive open interfaces and co-evolution among network software, switching hardware, and optical transport systems.

Facing the AI era, open network development requires not only higher bandwidth and larger scale, but also more flexible architectures, more reliable verification, and more efficient software delivery. Alibaba Cloud will keep combining production practice with global open network ecosystem building, providing an open, scalable, and sustainably evolvable network foundation for AI infrastructure.