Deepseek’s V3 and R1 models are available through Huawei’s Ascend Cloud service.
Deepseek recently launched Open Reasoning LLM and launched large inference LLM to large global markets built and trained for some of the costs of models from large competitors. While it’s becoming massively unstable in a global market that’s unstable at scale, Openai accuses Deepseek developers of using models to train their models.
The new paper claimed that Deepseek’s V3 LLM was trained on a cluster of 2,048 Nvidia H800 GPUs. Rumors about Deepseek’s new inference model, R1, suggest that it may be trained on 50,000 Nvidia “Hopper” GPUs, including the H100, H800 and the new H20, but Deepseek confirms this Not done. . If true, it raises serious questions about China’s access to advanced AI hardware despite continued trade restrictions, but that’s not a secret, but there’s the advanced Nvidia AI hardware. There is a prosperous black market.
Now, in a move that further shakes Western companies, South China Morning Post has announced that Huawei Technologies’ cloud computing unit will partner with Beijing-based AI infrastructure startup Silicon Flow to create a DeepSeek model with end users They report that they have made it available at an incredibly low price. .
Equipped with Huawei hardware
Working on the New Year’s Holiday in China, the collaboration provides efficient and cost-effective access to Deepseek’s V3 and R1 models through Huawei’s Ascend Cloud Service. It is prohibited in the US, UK and Europe.
Huawei has no secret that he wants to be China’s Nvidia. HuaweiCloud claims its performance level is comparable to the level of models running on premium global GPUs.
SiliconFlow, which hosts the DeepSeek model, comes out swinging at aggressive pricing, with V3 output tokens at 1 yuan (approximately 0.13 USD) per million input tokens, and R1 access at 4 yuan and priced at 4 yuan It will be provided. 16 yuan.
Microsoft added DeepSeek to Azure AI Foundry a few days ago, and Amazon quickly tracked it and added LLM to AWS’ Bedrock Managed Service. AWS introduced the AI model using an ML.P5E.48XLARGE instance with eight NVIDIA H200 GPUs delivering 1128GB of GPU memory. However, it is an early age for both cloud products, and is much more expensive than the ultra-low price of silicon flow.
The collaboration between Huawei, Siliconflow and Deepseek highlights China’s broader strategy to strengthen its domestic AI capabilities while reducing its dependence on Nvidia hardware.
The South China Morning Post said: “The move to launch Deepseek’s model on its own hardware backbone is to reduce China’s dependence on foreign technologies and strengthen its domestic AI industry as China’s increasing efforts to access high-end chips. It highlights China’s progress. The US government said it could be used to advance military objectives.”
