Bye bye Nvidia? Chinese cloud providers aggressively cut down AI inference costs by using Huawei’s controversial accelerators and DeepSeek’s tech

DeepSeek’s V3 and R1 models are available through Huawei’s Ascend cloud service
They are powered by the Ascend 910x accelerators banned in the US, EU and UK
The pricing is much lower than offered by Azure and AWS who have started trialing DeepSeek

DeepSeek recently massively unsettled global markets with the launch of its open reasoning LLM, which was built and trained for a fraction of the cost of models from much larger US competitors, although OpenAI has since accused DeepSeek’s developers of using its models to train theirs.

A new paper had claimed DeepSeek’s V3 LLM was trained on a cluster of just 2,048 Nvidia H800 GPUs – crippled versions of the H100 designed to comply with US export restrictions to China. Rumors around DeepSeek’s newer reasoning model, R1, suggest it may have been trained on as many as 50,000 Nvidia “Hopper” GPUs, including H100, H800, and the newer H20, although DeepSeek hasn’t – and likely won’t – confirm this. If true, it raises serious questions about China’s access to advanced AI hardware despite ongoing trade restrictions, although it’s no secret there’s a thriving black market for advanced Nvidia AI hardware there.

Now, in a move that’s going to further shake Western firms, the South China Morning Post reports Huawei Technologies’ cloud computing unit has partnered with Beijing-based AI infrastructure start-up SiliconFlow to make DeepSeek’s models available to end users for an incredibly low price.

Powered by Huawei hardware

This collaboration, which was worked on during the Chinese Lunar New Year holidays, provides efficient, cost-effective access to DeepSeek’s V3 and R1 models through Huawei’s Ascend cloud service, which is powered by Huawei’s own homegrown solutions, including the controversial Ascend 910x accelerators which are banned in the US, UK and Europe.

Huawei has made no secret that it wants to become the Chinese Nvidia, and Huawei Cloud claims its performance levels are comparable to those of models running on premium global GPUs.

SiliconFlow, which hosts the DeepSeek models, has come out swinging with some aggressive pricing, offering it for 1 yuan (approximately US$0.13) per 1 million input tokens and 2 yuan for output tokens with V3, while R1 access is priced at 4 yuan and 16 yuan.

Microsoft added DeepSeek to its Azure AI Foundry a few days ago, and Amazon swiftly followed suit, adding the LLM to its AWS’ Bedrock managed service. AWS showcased the AI model using an ml.p5e.48xlarge instance, powered by eight Nvidia H200 GPUs delivering 1128GB of GPU memory. It’s early days for both cloud offerings though, and they work out much more expensive than SiliconFlow’s super-low pricing.

The collaboration between Huawei, SiliconFlow and DeepSeek highlights China’s broader strategy to strengthen its domestic AI capabilities while reducing reliance on Nvidia hardware.

The South China Morning Post notes, “The move to launch DeepSeek’s models on a homegrown hardware backbone highlights China’s progress in cutting dependency on foreign technology and bolstering its domestic AI industry amid growing efforts by the US to choke off China’s access to high-end chips that the US government said could be used to advance military aims.”