Until now, IT leaders have needed to consider the cyber security risks posed by allowing users to access large language models (LLMs) like ChatGPT directly via the cloud. The alternative has been to use open source LLMs that can be hosted on-premise or accessed via a private cloud.
The artificial intelligence (AI) model needs to run in-memory and, when using graphics processing units (GPUs) for AI acceleration, this means IT leaders need to consider the costs associated with purchasing banks of GPUs to build up enough memory to hold the entire model.
Nvidia’s high-end AI acceleration GPU, the H100, is configured with 80Gbytes of random-access memory (RAM), and its specification shows it’s rated at 350w in terms of energy use.
China’s DeepSeek has been able to demonstrate that its R1 LLM can rival US artificial intelligence without the need to resort to the latest GPU hardware. It does, however, benefit from GPU-based AI acceleration.
Nevertheless, deploying a private version of DeepSeek still requires significant hardware investment. To run the entire DeepSeek-R1 model, which has 671 billion parameters in-memory, requires 768Gbytes of memory. With Nvidia H100 GPUs, which are configured with 80GBytes of video memory card each, 10 would be required to ensure the entire DeepSeek-R1 model can run in-memory.
IT leaders may well be able to negotiate volume discounts, but the cost of just the AI acceleration hardware to run DeepSeek is around $250,000.
Less powerful GPUs can be used, which may help to reduce this figure. But given current GPU prices, a server capable of running the complete 670 billion-parameter DeepSeek-R1 model in-memory is going to cost over $100,000.
The server could be run on public cloud infrastructure. Azure, for instance, offers access to the Nvidia H100 with 900 GBytes of memory for $27.167 per hour, which, on paper, should easily be able to run the 671 billion-parameter DeepSeek-R1 model entirely in-memory.
If this model is used every working day, and assuming a 35-hour week and four weeks a year of holidays and downtime, the annual Azure bill would be almost $46,000 a year. Again, this figure could be reduced significantly to $16.63 per hour ($23,000) per year if there is a three-year commitment.
Less powerful GPUs will clearly cost less, but it’s the memory costs that make these prohibitive. For instance, looking at current Google Cloud pricing, the Nvidia T4 GPU is priced at $0.35 per GPU per hour, and is available with up to four GPUs, giving a total of 64 Gbytes of memory for $1.40 per hour, and 12 would be needed to fit the DeepSeek-R1 671 billion-parameter model entirely-in memory, which works out at $16.80 per hour. With a three-year commitment, this figure comes down to $7.68, which works out at just under $13,000 per year.
A cheaper approach
IT leaders can reduce costs further by avoiding expensive GPUs altogether and relying entirely on general-purpose central processing units (CPUs). This setup is really only suitable when DeepSeek-R1 is used purely for AI inference.
A recent tweet from Matthew Carrigan, machine learning engineer at Hugging Face, suggests such a system could be built using two AMD Epyc server processors and 768 Gbytes of fast memory. The system he presented in a series of tweets could be put together for about $6,000.
Responding to comments on the setup, Carrigan said he is able to achieve a processing rate of six to eight tokens per second, depending on the specific processor and memory speed that is installed. It also depends on the length of the natural language query, but his tweet includes a video showing near-real-time querying of DeepSeek-R1 on the hardware he built based on the dual AMD Epyc setup and 768Gbytes of memory.
Carrigan acknowledges that GPUs will win on speed, but they are expensive. In his series of tweets, he points out that the amount of memory installed has a direct impact on performance. This is due to the way DeepSeek “remembers” previous queries to get to answers quicker. The technique is called Key-Value (KV) caching.
“In testing with longer contexts, the KV cache is actually bigger than I realised,” he said, and suggested that the hardware configuration would require 1TBytes of memory instead of 76Gbytes, when huge volumes of text or context is pasted into the DeepSeek-R1 query prompt.
Buying a prebuilt Dell, HPE or Lenovo server to do something similar is likely to be considerably more expensive, depending on the processor and memory configurations specified.
A different way to address memory costs
Among the approaches that can be taken to reduce memory costs is using multiple tiers of memory controlled by a custom chip. This is what California startup SambaNova has done using its SN40L Reconfigurable Dataflow Unit (RDU) and a proprietary dataflow architecture for three-tier memory.
“DeepSeek-R1 is one of the most advanced frontier AI models available, but its full potential has been limited by the inefficiency of GPUs,” said Rodrigo Liang, CEO of SambaNova.
The company, which was founded in 2017 by a group of ex-Sun/Oracle engineers and has an ongoing collaboration with Stanford University’s electrical engineering department, claims the RDU chip collapses the hardware requirements to run DeepSeek-R1 efficiently from 40 racks down to one rack configured with 16 RDUs.
Earlier this month at the Leap 2025 conference in Riyadh, SambaNova signed a deal to introduce Saudi Arabia’s first sovereign LLM-as-a-service cloud platform. Saud AlSheraihi, vice-president of digital solutions at Saudi Telecom Company, said: “This collaboration with SambaNova marks a significant milestone in our journey to empower Saudi enterprises with sovereign AI capabilities. By offering a secure and scalable inferencing-as-a-service platform, we are enabling organisations to unlock the full potential of their data while maintaining complete control.”
This deal with the Saudi Arabian telco provider illustrates how governments need to consider all options when building out sovereign AI capacity. DeepSeek demonstrated that there are alternative approaches that can be just as effective as the tried and tested method of deploying immense and costly arrays of GPUs.
And while it does indeed run better, when GPU-accelerated AI hardware is present, what SambaNova is claiming is that there is also an alternative way to achieve the same performance for running models like DeepSeek-R1 on-premise, in-memory, without the costs of having to acquire GPUs fitted with the memory the model needs.
