WAIC 2026 just wrapped up, and Huawei's 1024-card "big cabinet" might be the toughest piece of hardware at the entire show. On July 20th, the four-day 2026 World Artificial Intelligence Conference (WAIC) came to a close in Shanghai. Over 1,100 companies exhibited, with 349 products making their global debut—the largest scale in the event's history. From agentic phones to space-based computing satellite constellations, from humanoid robots to weather AI models, there was no shortage of fresh innovations to check out. But when it comes to the standout...

💡 What You Will Learn

WAIC 2026 just wrapped up, and Huawei's 1024-card "big cabinet" might be the toughest piece of hardware at the entire show. On July 20th, the four-day 2026 World Artificial Intelligence Conference (WA

📜 Table of Contents

• 1,024 Cards in One Machine — What Does That Even Mean?
• Not Just Stacking Hardware — Air-Cooled Version Is Here Too
• 750+ Commercial Deployments — Not Just on Paper
• From "Chasing NVIDIA" to "Building Your Own System"

WAIC 2026 Just Wrapped Up — Huawei's 1024-Card "Big Cabinet" Might Be the Hardest Piece of Hardware on the Floor

On July 20th, the four-day 2026 World Artificial Intelligence Conference (WAIC) wrapped up in Shanghai.

Over 1,100 companies exhibited, 349 products had their global debuts, and the scale broke all previous records. From agentic phones to space-based AI computing constellations, from humanoid robots to weather AI models, there was certainly no shortage of new and exciting things to see. But when it comes to the hardest piece of hardware on the floor, many would point to that big cabinet at Huawei's booth — the Ascend Atlas 950 SuperPoD, 16 compute cabinets combined into one, packed with 1,024 Ascend cards, on real display.

This wasn't a PowerPoint slide; it was the actual hardware. A spec sheet was posted next to the booth, and several engineers immediately pulled out their phones to take photos — which isn't common at WAIC, and it shows this machine is genuinely impressive.

1,024 Cards in One Machine — What Does That Even Mean?

Let's look at the specs first:

Each cabinet holds 64 Ascend cards, and the on-site display combined 16 cabinets into one unit, totaling 1,024 cards. FP8 compute hits 1 EFLOPS, doubling to 2 EFLOPS with FP4, and 256TB of global unified memory — all cards share the same "brain," no need to shuffle data back and forth. Inter-card NPU interconnect bandwidth reaches TB-level, with latency as low as 3 microseconds.

And this isn't even the ceiling — it can scale up to 8,192 cards, more than 20 times the scale of the previous-generation Ascend 384 supernode.

There's a key concept here called a "supernode": it's not just stringing a bunch of servers together with network cables and calling it a cluster. Instead, a high-speed interconnect protocol lets all cards share the same memory space, enabling compute pooling and zero-copy data transfer. The Ascend 950's "Lingqu" (UnifiedBus) interconnect protocol is exactly what does this.

What was the biggest pain point for domestic AI chips before? Single-card performance could catch up, but scaling fell apart — once you added more cards, communication latency ate away at the compute. That's what Lingqu solves: TB-level bandwidth plus 3-microsecond latency lets 1,024 cards work together like a single computer, rather than 1,024 small machines each doing their own thing.

Built on this supernode, the Atlas 950 SuperCluster can scale to 500,000 cards — already at the top tier globally.

Not Just Stacking Hardware — Air-Cooled Version Is Here Too

Another challenge with large clusters is cooling.

The previous Ascend 384 supernode used liquid cooling. Performance was fine, but it wasn't friendly to existing data centers — retrofitting for liquid cooling requires civil engineering and adding piping, and many data centers run by telecom operators and financial institutions simply can't be modified.

Huawei also showcased the Atlas 850E air-cooled supernode this time, which can scale to 96 cards for commercial deployment and mounts directly into standard racks — no liquid cooling retrofit needed. This is great news for financial institutions, telecom operators, and the education sector — no need to spend millions retrofitting data centers just to upgrade equipment. The barrier to AI compute adoption just dropped another notch.

750+ Commercial Deployments — Not Just on Paper

The Ascend 950's debut isn't just flexing muscles in a lab. The previous-generation Ascend 384 supernode has already seen 750+ commercial deployments, covering eight industries including internet, telecom, finance, education, healthcare, transportation, and manufacturing. It's also the only domestic supernode that has trained SOTA models.

On the software ecosystem front, since the CANN heterogeneous computing architecture went fully open-source in 2025, the community has gathered 67 projects, with over 12.44 million lines of open-source code, and monthly active developers surpassing 3,500. It supports mainstream open-source frameworks like PyTorch, vLLM, SGLang, and Triton on the upstream side, so developers don't need to learn an entirely new toolchain from scratch.

From "Chasing NVIDIA" to "Building Your Own System"

What's most striking about this machine isn't how high the specs are — it's how the narrative around domestic compute is shifting.

For years, the story of domestic AI chips has been "catching up to NVIDIA" — Has single-card performance caught up? Has memory bandwidth caught up? Has the software ecosystem caught up? Every question implicitly used NVIDIA as the reference point.

But looking at the Ascend 950's specs, that reference point might need to change. A 1,024-card supernode, 1 EFLOPS, a 500,000-card cluster, a self-developed "Lingqu" interconnect protocol — domestic compute has moved from "matching a specific metric" to "building its own technology stack."

The supernode concept itself is a track Huawei defined. NVIDIA has NVLink, Huawei has Lingqu; NVIDIA has DGX, Huawei has Atlas SuperCluster. Both are pursuing full-stack closed architectures, but Huawei's path is "heavier" — from chips to interconnect to software to clusters, all done in-house. This path requires heavy upfront investment and delivers results slowly, but once you've walked it, no one can choke you.

Of course, ecosystem building is still a work in progress. CANN's 3,500 monthly active developers still lag far behind CUDA's millions, and there's still a gap between benchmark scores and real-world deployment. But the direction is clear: Huawei isn't building a "domestic replacement for NVIDIA" — it's building a full-stack system from chips to interconnect to software to clusters.

This path is harder, but once it's walked, it can't be choked. A domestic solution with a 500,000-card cluster means Chinese large-model companies no longer have to wait in line for NVIDIA's production capacity — there's now another path of their own.

Do you think the Ascend 950 can go head-to-head with NVIDIA? Drop your thoughts in the comments.