DeepSeek Plans Massive Huawei AI Chip Deployment in Inner Mongolia

Hangzhou-based artificial intelligence organisation DeepSeek is preparing a significant deployment of Huawei Ascend 950DT accelerators. The plan involves installing at least 160,000 of Huawei's next generation of chips at a new data centre in Inner Mongolia. This facility is being constructed on a gigawatt scale site, with chip and memory supply dictating delivery pace for this extensive model inference cluster.

Credit: UNSPLASH

The proposed installation, if completed at its projected scale, would represent one of the largest known clusters built around Huawei artificial intelligence silicon. This move, according to individuals familiar with the matter cited by Bloomberg, signifies a notable increase from prior Ascend installations. Previous deployments have been reported in the low tens of thousands of cards, making DeepSeek's current plans a substantial expansion in capacity. The data centre itself remains under construction, with the accelerators intended for its operational phases.

But DeepSeek intends to utilise the 950DT cards primarily for inference, which entails serving production models to end-users. This differs from the more demanding work of training new artificial intelligence systems. While DeepSeek has previously experimented with Huawei hardware for training purposes, the organisation continues to depend on Nvidia accelerators for that specific stage of development. The Ascend 950DT is advertised for both decode and training operations; however, DeepSeek's immediate strategy separates these tasks onto different hardware stacks, highlighting the current strengths of domestic accelerators for high volume serving applications.

The chosen location and the power requirements are as noteworthy as the sheer number of chips involved. The Inner Mongolia campus is described as operating at a gigawatt scale, indicating sufficient power draw at full operational capacity to supply approximately 750,000 homes. The availability of inexpensive western power and favourable cooling conditions aligns with the "eastern data, western compute" model frequently adopted by many Chinese operators for intensive artificial intelligence workloads. The planned 160,000 Ascend 950DT units are projected to occupy only a portion of the site's overall capacity, leaving a significant segment available for other accelerator configurations.

And the precise timing of delivery is contingent upon Huawei's capability to manufacture the 950DT units in significant quantities. Industry observers note that shortages of high end memory and competition from other customers are expected to limit 950DT output to the low hundreds of thousands in the current period. Consequently, fulfilling DeepSeek's substantial order may extend over an extended duration. DeepSeek has reportedly sought governmental support to secure a larger and quicker allocation of Huawei's premier cards, amidst intense competition from other major cloud and internet buyers for the same production line.

This substantial build forms part of a wider industry movement towards employing domestic artificial intelligence accelerators for large scale serving operations. Several Chinese firms have already established smaller Ascend clusters. Artificial intelligence labs across the country continue to combine local silicon for inference tasks with imported or externally rented capacity for training when required. The overall narrative for operators observing China's artificial intelligence computing landscape concerns the resilience of the Ascend ecosystem.

So this involves assessing whether Ascend supply, its networking infrastructure, power provisions, and cooling systems can collectively sustain the demands of model traffic at a true data centre scale. DeepSeek and Huawei declined to provide comment regarding the reported plans. The situation presents a significant test for the domestic supply chain and its ability to meet escalating computational needs for artificial intelligence applications.

• DeepSeek plans to install at least 160,000 Huawei Ascend 950DT accelerators at a new data centre in Inner Mongolia.

• These chips are intended primarily for artificial intelligence model inference, not for training new systems.

• The Inner Mongolia site is designed to operate at a gigawatt scale, with power capacity sufficient for 750,000 homes.

• Delivery of the chips faces potential delays due to shortages of high end memory and competing demand from other buyers.

• The deployment reflects a broader industry shift towards domestic artificial intelligence accelerators for large scale serving operations.

Source: Pandaily