Building the Arteries of the Agentic World: Huawei Upgrades Stellar AI Network With NPO Switches, Quantum-Safe WAN and Embedded AI Guardrails

Under the neon of HUAWEI CONNECT 2026 in Shanghai, Huawei’s Data Communication product line made a case that the AI story of 2026 is no longer just about models — it is about the network plumbing underneath them. The company unveiled a comprehensively upgraded Stellar AI Network Solution, spanning AI Fabric, AI WAN and AI Campus, built around a single philosophy it repeated like a drumbeat: secure, intelligent connectivity for the “Agentic World.”

The backdrop is a number that should stop any infrastructure planner: Huawei says daily token consumption has surged 260-fold in a single year. When inference traffic multiplies by two orders of magnitude, the switches, WAN links and campus networks built for human-scale browsing become the bottleneck. Huawei’s answer is a full-stack refresh that pairs its Ascend compute silicon with purpose-built networking — and bakes security models directly into the forwarding path.

Stellar AI Fabric 2.0: 11-microsecond networks and in-house optics

The data-center layer got the deepest technical overhaul. Huawei introduced the SF9300 series UBG switch, which pairs with the Ascend A5/A6 AI accelerators to form what the company calls a full-series “compute-network integrated” solution. The design choices are aggressive:

• A two-layer multi-plane architecture eliminates the core switch layer entirely, which Huawei says cuts network deployment costs by 30%.
• A unified protocol reduces end-to-end latency from 20 μs to 11 μs — roughly halving the time a token spends traversing the cluster network.
• Link-Layer Retransmission (LLR) delivers zero packet loss during link flaps, which Huawei credits with reducing annual cluster service downtime by 100 hours per year.

The more strategically significant reveal was the CloudEngine XH9300-EN series NPO switches — fully in-house co-packaged optics. NPO (near-packaged optics) moves the light source away from the switch silicon and is the industry’s answer to the power wall in large-scale AI clusters. Huawei claims three breakthroughs:

1. A centralized light-source design lowers interconnect power consumption from 1,000 W to 600 W per unit — a 40% cut in cluster electricity costs at the interconnect layer.
2. A near-packaged architecture reduces optical-electrical conversion latency by 180 ns, optimizing single-node latency by 26%.
3. A pluggable snap-fit structure enables on-site replacement, shortening fault recovery from days to hours — a big deal when a dead optical module can idle an entire training row.

The in-house angle matters beyond performance. With export controls still constraining what Huawei can buy from Western suppliers, an NPO switch line designed and manufactured internally extends the company’s stack-of-one strategy from Ascend accelerators (as showcased in the Atlas 960 SuperPoD earlier this week) down into the optical interconnect itself.

The WAN story: computing power delivered “lossless” over 1,000 km

The second pillar targets the emerging geometry of AI infrastructure, where compute is scattered across regions and jurisdictions. Huawei’s Stellar AI WAN solution claims “lossless delivery of computing power over 1,000 km” through three mechanisms:

• The Starnet lossless algorithm eliminates long-distance packet loss, pushing remote computing efficiency above 95%.
• The XH computing-network appliance uses layerwise model partitioning — running a model’s first and final layers locally and transmitting only high-dimensional vectors over the network. The effect is twofold: raw data never leaves the premises (a data-sovereignty feature), and required branch xPU costs drop fourfold.
• Computation-communication overlapping cuts bandwidth idle time by over 80% and reduces required bandwidth fivefold.

For multinational enterprises navigating data-residency regimes — or Chinese provinces competing to host inference capacity far from the coastal hyperscalers — the ability to federate inference over distance without a performance cliff is a genuine capability, not just marketing.

Security at the WAN layer leans on three tiers: intrinsic security boards that trace threats within minutes at over 95% accuracy and visualize up to 100-hop attack paths; built-in QKD (quantum key distribution) that adds quantum-safe links without extra devices, cutting construction costs by over 60% while extending transmission distance to 80 km; and APN6-based data fencing for network-wide control of sensitive traffic paths.

Campus networks learn to defend themselves

The third pillar, Stellar AI Campus, addresses the quiet operational crisis of agentic AI inside enterprises: with agent adoption approaching 50% of enterprises this year by Huawei’s estimate, manual network operations cannot keep up with 24/7 demands.

Huawei’s upgrade pushes autonomy across data, models and tools. Proprietary network-awareness technology samples over 25 exclusive data types on a 100 ms cycle; a Deep Causal Pruning (DCP) algorithm cuts inference errors by over 20% while lifting troubleshooting accuracy to 95%; and iFlow 2.0 distributed simulation continuously verifies inference paths. The net result, Huawei claims, is network-wide AI-powered operations with MTTR slashed to under 5 minutes.

The security half is where it gets novel. Huawei’s Asset Security 2.0 embeds AI micro-models directly into switches, cutting dumb-terminal threat detection time from 5 minutes to under 10 seconds, while AI Core small models and NPU-accelerated semantic large models reconstruct agent interaction scenarios to catch malicious commands at a claimed 95% detection rate. The frame: endpoint-network collaborative defense against both compromised IoT gear and rogue agent behavior.

Guardrails for the token factory

Tying the layers together is Huawei’s Three-Layer Security Guardrail for the AI production chain itself — defense against hashrate theft, prompt injection and model poisoning. A network-embedded cryptojacking detection model claims a 95% detection rate with millisecond-level blocking of malicious traffic; an Agent Guardrail targets prompt-injection attacks at 95% accuracy under 150 ms latency; and a Host Guardrail monitors malicious processes while keeping CPU overhead below 1%. Huawei cited one real-world customer site where the stack surfaced 415 high-risk vulnerabilities and 322 threat events in its first seven days of operation.

The bigger picture

Two readings of the announcement are worth holding simultaneously.

The first is technical: Huawei is systematically converting the pain points of hyperscale AI — interconnect power, cluster latency, link reliability, distance-induced loss, agentic attack surface — into product lines with quantified claims. The numbers (30% cheaper deployment, 40% less interconnect power, 100 fewer downtime hours) are the kind that go directly into TCO spreadsheets, and they arrive alongside the Ascend-based Atlas 960 SuperPoD shown at the same event, forming a coherent full-stack alternative for buyers who cannot or will not build on Nvidia networking.

The second is geopolitical. Every layer of this release — Ascend-paired switches, in-house NPO optics, QKD links, data-sovereign WAN partitioning — reduces dependence on Western suppliers and aligns with China’s just-released Five-Year Plan targeting 9,800 EFLOPS of intelligent computing by 2030. Huawei’s own forecast that agents will drive over 90% of global AI processing traffic by 2035 frames the bet: whoever architects the network for the agentic era defines the substrate everyone else must interoperate with.

For buyers outside China’s orbit, the release is still a signal worth reading: co-packaged optics moving from conference talks to shipping product, security models running in-switch at line rate, and WANs designed around vector traffic rather than raw data. The agentic world is being built now — and increasingly, its arteries are intelligent.