Huawei has announced a 64PB all-flash storage tier — the OceanStor M900 — built to feed KV caches to its Atlas 960 SuperPoD accelerator racks, and it claims numbers that put it well past rival AI storage systems from DDN, IBM, MinIO, VAST Data and similar vendors.

Specifically, Huawei says the M900 cluster delivers 40TB/sec of aggregate bandwidth, which the company calls 1.5 times higher than peer solutions without naming them. Those peers — per our reading of the market — cluster in the 10–25TB/sec range. That's the headline, and it lands squarely in Nvidia's territory.

Why does KV cache need its own storage?

Nvidia has already defined the category with its AI Data Platform reference design: GPUs, BlueField NICs, Spectrum-X switches and Dynamo software layered over a tiered memory hierarchy. GPU HBM sits at the top. Server DRAM follows, then local SSDs, then BlueField-4-connected NVMe flash acting as an L3.5 tier — one network hop from HBM, microsecond-class latency.

Huawei is copying the playbook with its own silicon. Its SuperPoDs target 10-trillion-parameter models with million-plus-token context windows. At that scale, no single accelerator's high-bandwidth memory can hold all the key-value tokens. You need a caching scheme, and you need flash close to the compute.

The Atlas 960E SuperPoD scales to 4,096 Ascend NPUs, delivering 8 EFLOPS of FP8 compute with up to 1PB of HBM and a 256TB unified memory pool. Connectivity comes via 5,500 Hi-ONE units using UnifiedBus near-packaged optics. Huawei says that arrangement cuts power draw by over 550kW versus the 48,000 800G optical modules a conventional fabric would require — and doubles fault-free operating time to 99.8 percent system availability.

What does the M900 actually deliver?

The M900 provides up to 4PB of shared L3.5 KV cache capacity per cluster, carried over optical networking at that 40TB/sec figure. Per NPU, that works out to terabytes of cache capacity — enough, Huawei claims, to double an inference cluster's token throughput and halve time to first token in typical AI programming scenarios. Those are vendor numbers, not independent test results.

The box itself integrates the CPU, network controller and NAND controller in one unit, so SuperPoD NPUs get a direct one-hop connection to the SSDs. Huawei says this cuts access latency from milliseconds to roughly 60 microseconds.

David Wang, Huawei's Deputy Chairman and Rotating Chairman, framed the endurance story as the differentiator. "OceanStor M900 also uses hybrid media and an optimized retention algorithm, extending SSD read/write lifespan by 16-fold," he said. "This ensures a higher KV cache hit rate alongside long-term stability and reliability from the ground up."

Under the hood, the M900 runs what Huawei calls KV-aware adaptive storage. It predicts the expected lifetime and value of each piece of KV cache data, then schedules that data across on-chip memory, DRAM and SSDs. Huawei claims this placement strategy lets the drives absorb up to 24 drive writes per day — a strikingly high rating — stretching SSD endurance 16 times and supporting three years of stable operation with fewer replacements.

What don't we know yet?

Plenty. Huawei hasn't published M900 data sheets or technology backgrounders. There's no word on rack unit size, controller counts, drive counts, per-node capacity or cluster node counts. Pricing, availability dates and which flash Huawei sources all remain open questions — the last one politically loaded, given US export controls on China's access to leading-edge NAND and the domestic supply chain Huawei has built around its Ascend NPU line.

Is this a real SuperPOD rival?

The compute side is unambiguous: recent products like the Atlas 350, using the Ascend 950PR chip, and the upcoming Atlas 960 SuperPoDs position Huawei as the domestic alternative to Nvidia GPUs in China. The NPUs use Huawei's Da Vinci architecture — Cube matrix cores plus Vector cores — and support FP8 and FP4 low-precision formats for training and inference. Strung together over UnifiedBus all-optical interconnect, they scale to hundreds of thousands of NPUs acting as one logical machine.

The storage tier was the missing piece. Now Huawei has a full-stack answer to Nvidia's AI Data Platform: compute, interconnect, and flash, all under one roof. Until datasheets arrive and third parties can test the 40TB/sec and 2x-throughput claims against DDN, VAST and IBM deployments, the benchmark gap stays an open question — but the competitive pressure on Western AI storage vendors selling into China just became concrete.