DeepSeek AI challenges ChatGPT with low-cost innovation

The Chinese artificial intelligence company DeepSeek has developed a Generative AI model that is accurate, comparable to ChatGPT, and significantly cheaper to operate.

Devendra Sharma, Chief Data & AI Officer at Qodea, has outlined several factors contributing to DeepSeek's cost-effective operation. He said, "DeepSeek has done three very clever things that have allowed its model to operate at a fraction of the cost and compute power of other GenAI models, and thus completely upend the AI market."

DeepSeek's training efficiency sets it apart as it requires substantially fewer resources. "DeepSeek is much more efficient when training. Most LLMs require anywhere from 80,000 to 100,000 GPUs for training. DeepSeek conducted its training with only around 2,000 GPUs and can use lower-powered chips," Sharma noted. "While other models are spending USD $100 million in the region, DeepSeek managed to do the same for just USD $5 million."

The model employs less precise numerical computations, significantly reducing the demand on memory and costs. "Most current AI models—ChatGPT, Gemini, etc.—perform numerical computations up to 32 decimal places. However, DeepSeek decided that this level of precision was unnecessary, and instead, their model only calculated up to 8 decimal places. By using a lower 'floating-point precision', DeepSeek can use up to 70% less memory than other models, massively reducing costs," Sharma stated.

Another cost-saving measure involves strategically activating parameters, differentiating it from traditional models. Sharma explained, "DeepSeek significantly reduces the number of active parameters. Imagine that you're sitting in front of ChatGPT and inputting a question. In traditional AI models, it's like having thousands or even millions of experts sitting there awake and ready to answer you at all times – which, of course, consumes huge amounts of power and resources. DeepSeek takes a different approach. Instead of keeping all those 'experts' active at the same time, it only activates them when needed. This is a huge cost-saving move."

Despite its advantages, integrating the model into existing business operations can be challenging. Google's Gemini is cited as more user-friendly.

Adam Casey, Director of Cyber Security and CISO at Qodea, highlighted concerns regarding the cybersecurity risks associated with DeepSeek. "There are no free meals in this world. While the low costs of DeepSeek will be tempting for many, I'd urge extreme caution," Casey warned.

Casey pointed to broader geopolitical concerns about technology originating from China, remarking, "DeepSeek is extremely problematic from a data security and compliance perspective, for obvious reasons. China has a history of espionage, and most Western governments, therefore, consider technologies from China to be high risk." He drew parallels to other Chinese companies, such as TikTok, Hikvision, and Huawei.

Further emphasising data privacy issues, Casey added, "In addition, under GDPR there is a requirement around data transparency that DeepSeek cannot meet. You need to know how a system works, where the data is processed, and whether you have consent for that data to be shared – it's a minefield."

He cautioned that while bans on technology applications like DeepSeek could be enforced, there are challenges in their effectiveness, especially with third-party use within supply chains. "We are already seeing that AI technologies are being introduced by stealth – by employees and through third parties – without organisations being aware," Casey said. "While you may not be using the technology directly, your service and third party software providers are incentivised to lower the cost of delivering their solutions and therefore may be tempted to use technology like DeepSeek."