Chinese AI lab DeepSeek has reportedly ordered 160,000 Huawei Ascend 950DT AI accelerators, according to TechPowerUp and Wccftech, with Wccftech citing a Bloomberg report. The chips are intended for large-scale inference operations and will be deployed in a single data center in Inner Mongolia, which Wccftech describes as a 1GW facility.
The Ascend 950DT features 144 GB of Huawei's own HBM memory design, with TechPowerUp citing nearly 4 TB/s of bandwidth and Wccftech specifying 4.0 TB/s bandwidth alongside a 2.0 TB/s GPU-to-GPU interconnect. TechPowerUp states each chip can deliver about one PetaFLOP of FP8 compute or roughly two PetaFLOPS at FP4, and Wccftech notes the chip also supports FP8, FP4, and Huawei's proprietary HiF8 low-precision formats. Wccftech adds that the chip is built on Huawei's third-generation DaVinci v3 architecture using SMIC's N+3 process, an advanced 7-nanometer-class technology.
At the system level, TechPowerUp reports the full DeepSeek deployment will deliver a peak of 160 ExaFLOPS at FP8 or about 320 ExaFLOPS at FP4, with roughly 23 PB of total HBM memory, enough to serve thousands of concurrent users. The chips connect via Huawei's Lingqu interconnect and optical links, with up to 8,192 chips per Atlas 950 SuperPoD; TechPowerUp notes Huawei's architecture can theoretically scale to 64 SuperPoDs (524,288 chips total), far beyond the current order. Wccftech separately notes the Atlas 950 SuperPoD is rated for 8 EFLOPS of peak AI compute per 8,192-chip unit.
On pricing, Wccftech reports the Ascend 950DT sells for around 111,000 Yuan (about $16,000) per unit, compared to $15,000–$25,000 for Nvidia's H20, putting DeepSeek's total order at roughly $2.56 billion. Wccftech also notes Huawei's production capacity is constrained to a few hundred thousand units per year, meaning the order could take up to a year to fulfill.
Wccftech frames the order as a snub of Nvidia's China-specific H20 GPUs despite what it describes as a roughly 4x performance gap and a year-long delivery delay, citing earlier comments from DeepSeek CEO Liang Wenfeng that Huawei's supernode can match Nvidia's GB300 in capability and latency, though requiring about four Huawei GPUs per Nvidia GPU and lagging by two years. The piece also cites commentary from Sphere Semi CEO Steven Glinert suggesting Nvidia views China as a market where Huawei now holds an unassailable lead.
Both sources note Inner Mongolia's growing role as a hub for Chinese AI infrastructure. Wccftech points to the region's cold, arid climate as suited for solar-powered data centers, citing Alibaba's Ulanqab facility with power costs around $0.05 per kWh, and notes ByteDance is reportedly in talks to add 5-6GW of capacity in the region, with each 1GW of capacity costing around $23.8 billion.