The Thesis Behind the Spectacle
Huawei doesn't want to build a better Nvidia chip. They want to build a bigger machine.
That was the unmistakable subtext of Eric Xu's keynote at HUAWEI CONNECT 2025 in Shanghai, where the Rotating Chairman unveiled what Huawei calls the world's most powerful SuperPoDs and SuperClusters — computing systems that shelve the idea of the GPU as hero unit and reframe AI hardware architecture as a question of interconnect topology and aggregate scale.
The logic is almost audacious in its simplicity. Huawei can't access TSMC's leading-edge process nodes. US export controls saw to that after 2019, when the Entity List cut off access to American semiconductor technology and the Commerce Department tightened the screws on advanced chip manufacturing equipment. Instead of chasing density on a single die, Huawei's answer is to wire hundreds of thousands — eventually over a million — Ascend NPUs together into what they describe as a single logical machine that can "learn, think, and reason as one."
Whether that bet pays off depends on questions far harder than raw teraflops. But the roadmap, as published, is the most aggressive domestic AI hardware architecture proposal to come out of China. And it isn't vague vaporware. Huawei has shipping hardware today.
The Numbers on the Table
Let's get the specs down before the analysis, because they're the thing everyone's going to quote.
The Atlas 950 SuperPoD packs 8,192 Ascend NPUs. Huawei claims 1 EFLOPS of FP8 performance and 7.3 EFLOPS at FP4 precision. Shipping window: Q4 2026. The Atlas 960 SuperPoD doubles the NPU count to 15,488 and roughly quadruples the math to 4.9 EFLOPS FP8 and 30.9 EFLOPS FP4, planned for 2027.
Stack SuperPoDs into clusters and the numbers go vertical. The Atlas 950 SuperCluster aggregates over 500,000 Ascend NPUs. The Atlas 960 SuperCluster targets more than one million. Huawei's explicit claim — made in the company's own press materials, which is a genre that usually under-promises — is that these systems "are poised to outperform all other computing clusters on the market."
More pointedly, Huawei says it leads Nvidia's latest publicly announced roadmap by roughly 40% in computing performance, three-fold in interconnect bandwidth, and 60-fold in scale. The comparison isn't apples-to-apples, Nvidia's NVL144 and the upcoming Rubin Ultra platform take different architectural stances on how to build a frontier training system, but the directness is telling. Huawei is no longer positioning itself as a domestic fallback. It's positioning itself as a global pacesetter.
One detail that matters for the "can they actually scale this?" skeptics: Huawei claims the Atlas 950 SuperPoD can scale "unconditionally to 1 million cards" via what they call a 3D Matrix Network with orthogonal interconnect topology. Nvidia's NVL144 tops out at 72 GPUs. Nvidia's Rubin Ultra tops out at 144. Huawei's architecture, in theory, has no hard ceiling, the topology is designed to grow along three axes (level scaling, cluster scaling, PoD scaling) without hitting a switching wall.
Theory is the keyword.
LingQu and the Interconnect Problem
The single hardest engineering challenge in any multi-thousand-chip training cluster isn't the silicon. It's the wires between the silicon.
Huawei's answer is LingQu (officially "UnifiedBus"), a full-stack interconnect standard that spans chips, computing modules, SuperPoDs, and SuperClusters. At HUAWEI CONNECT 2025, Xu announced that Huawei has published the LingQu technical white paper as an open standard, the first such Chinese interconnect architecture to be offered to global partners. The stated ambition: "one standard, one ecosystem" over the next five years.
On the Atlas 950 SuperPoD, each Ascend NPU gets 16K UnifiedBus links. Huawei claims this roughly doubles interconnect bandwidth per NPU compared to the prior generation while cutting latency by 10x. At the PoD level, the 3D Matrix Network delivers microsecond-level cross-node communication, fast enough to treat 8,192 chips as a unified memory domain for training workloads.
Compare that to the industry baseline. Most AI clusters today bottleneck at the interconnect layer: NVLink is fast but proprietary; InfiniBand is fast but complex; Ethernet is scalable but has to be tortured into behaving like a lossless fabric. If LingQu delivers on its latency claims, and if Huawei can attract third-party adoption, it addresses the architectural bottleneck that Hu, in his keynote, explicitly named as one of three barriers to the intelligent era (alongside computing gaps and ecosystem fragmentation).
Three pillars. Computing gaps. Architectural bottlenecks. Ecosystem fragmentation. The LingQu announcement isn't just a product launch, it's an ecosystem play. Huawei has 300+ commercial customers running Atlas 384 SuperPoDs across 20+ industries today. Those are the seeds of an installed base that could pull partners into the LingQu orbit rather than Huawei trying to build a coalition from scratch.
Rethinking AI Hardware Architecture for Inference
Training gets the headlines. Inference gets the bill.
Huawei's UCM (Unified Computing Memory) inference architecture is a direct response to the memory wall problem that's become acute with reasoning models and long-context serving. The architecture splits into two approaches: UCM-S for speculative cache optimization across device memory, host memory, and local SSD; UCM-A for intelligent offloading of KV caches into an external large memory pool.
The claim: UCM can reduce inter-token time by up to 10x on certain workloads. If even partially true, that's the kind of efficiency gain that changes inference economics for production deployments, and it sidesteps the HBM capacity race that currently locks cloud providers into Nvidia's supply chain.
This is the part of Huawei's strategy that I find genuinely interesting beyond the marketing superlatives. Everyone can point at an EFLOPS number. Building a coherent inference layer that decouples memory from compute is a harder, smarter problem, and one where the incumbents (Nvidia's H100/H200/GB200 roadmap) have a vested interest in keeping everything bundled.
The Ecosystem Reality Check
Here's what keeps me cautious.
Huawei's claims, and the press release is thick with them, rest heavily on comparisons to "publicly announced product roadmaps from peers." Nvidia's roadmap has a long history of shifting. Huawei's own history with the Entity List has been a series of workarounds built on SMIC's constrained process nodes, where each generation of Ascend has squeezed more from less. The Atlas 850 NPU upgrade is slated for 2026; the Atlas 960 depends on a "next-generation NPU" whose specs and manufacturing path Huawei hasn't disclosed.
That last point isn't academic. SMIC's most advanced public process node, their equivalent of 7nm-class, has no EUV lithography access. Huawei can wire more chips together. They cannot make individual chips faster or more power-efficient on the trajectory that TSMC and Samsung can. A million-card cluster running on less efficient silicon will draw serious electricity.
The broader question of whether LingQu gains international traction is equally open. Huawei's openness pitch, publishing white papers, inviting partners, is credible as a strategy. Whether it survives the gravitational pull of CUDA's sixteen-year software moat is a different question entirely.
Where This Lands
Huawei's HUAWEI CONNECT 2025 keynote wasn't a product demo. It was a position paper for AI hardware architecture as a systems problem rather than a chip problem. The bet is that the next five years will reward the player who connects the most chips together fastest, not the one who builds the single best chip.
The Atlas 384 has proven Huawei can ship at meaningful commercial scale. The Atlas 950 and 960 SuperPoDs, if they land on the Q4 2026 and 2027 timelines, would represent a genuine leap, not just for China's domestic compute independence, but for how the global industry thinks about scaling AI infrastructure beyond the GPU paradigm.
Whether the numbers hold when you run real training runs on them? That's the question that matters, and it's one Huawei hasn't answered yet. Give them time. But don't confuse the roadmap with the receipt.
