DeepSeek under the microscope: a sanity check

Chinese developer DeepSeek’s achievements with its V3 and R1 AI models has sent shock waves through the industry, suggesting that high performance models can be developed and trained without access to high end GPUs—and at a much lower cost than many of the leading foundation models that dominate the market. This has sparked a storm of commentary and media coverage, with much of it overblown, confusing, and misleading. This blog, written by Eden Zoller, Bradley Shimmin, Lian Jye Su and Alexander Harrowell is based on an extended analysis (DeepSeek Sanity Check) that cuts through the hype to present a balanced perspective of DeepSeek and what its actions mean for the AI industry and enterprises. 

The evolution of DeepSeek the industry did not see coming

DeepSeek has been steadily releasing its family of open-source models over the last couple of years and its innovations are not solely premised on new technology breakthroughs that in reality are very rare. The DeepSeek R1 family of reasoning models leverage existing open-source models, including its own V3, Meta’s Llama, and Alibaba’s Qwen. What the talented DeepSeek team have done is to approach model development in ways that are creative, agile, and highly effective. DeepSeek-R1 trained on large-scale reinforcement learning without supervised fine-tuning, incorporating multi-stage training and cold-start data. This implies that DeepSeek-R1 requires even fewer resources, which explains the excitement around its reasoning and inference performance. According to DeepSeek, R1 beats OpenAI o1 on the benchmarks AIME, MATH-500, and SWE-bench Verified. 

The arrival of DeepSeek-V3 and R1 is being treated as a game changer, but the game has been gradually changing over the last few years—the shift from very large proprietary, general-purpose models into small, task-specific, open-source models. Features that were considered frontier-level in large models, such as multimodality, are now appearing in the 3B–8B range, while basic LLM functionality is growing in the 1–3B range. The arrival of smaller, cost-effective open models is expanding AI innovation and competition in terms of who drives it and the forms it can take, which is good news for developers and enterprises. Small models are particularly beneficial to innovation on edge devices such as smartphones, PCs, automotive, and robotics. Certain versions of DeepSeek models (e.g., DeepSeek-R1-Distill) can be hosted locally, appealing to organizations prioritizing data privacy and vendors focusing on edge AI applications that require ultra-low latency. 

DeepSeek data privacy in perspective  

The news of DeepSeek’s achievements has ratcheted up ongoing geopolitical sensitivities about data access and exploitation, particularly in the US. Geopolitics aside, enterprises interacting with DeepSeek’s hosted AI services are, on the face of it, opening themselves up to potential data privacy loss. But this is the case with all cloud-based AI services, and the only appreciable difference between AI services from Apple, Meta, ByteDance, and DeepSeek rests in who gains access to how much valuable data—the AI service provider, advertising data brokers, or even governmental agencies. It is therefore up to corporate IT to control access to popular GenAI models, presenting users with internally controlled options that are good enough to dissuade users from taking things into their own hands (e.g., shadow IT). DeepSeek is available under a free open-source software licensing model, which is significant because it allows companies to protect their data when using emerging AI services. Open-source software gives companies the option to both build derivations of the DeepSeek-R1 and host those themselves without any involvement from DeepSeek beyond providing attribution. 

DeepSeek will not stop the march of big models 

The financial markets reacted strongly to DeepSeek-V3 and R1, regarding these models as serious challengers to the dominant foundation models from vendors like OpenAI. Concerns also surfaced that DeepSeek’s modeling techniques (using cheaper, less powerful GPUs) would negatively impact the long-term demand for compute-intensive AI chipsets and hyperscale data centers. 

Although DeepSeek’s achievements suggest the path for AI development set by OpenAI and others is not the only way forward, it does not signal the end of large-scale foundation models that require high end AI chipsets and access to powerful data center infrastructure. Like it or not, there will still be a need for large-scale foundation models. Even though smaller-scale models like DeepSeek-V3 perform very well in specific domains, they may lack the broad-spectrum versatility of large models like GPT-4 or Claude 3.5, and this helps sustain demand for the larger models.

Moreover, too many powerful vested interests are driving frontier AI and the infrastructure investments that support this, notably in the US where Microsoft, OpenAI, Google, Amazon, and Anthropic are committed to AGI. The new US administration’s $500bn Stargate Project, aimed at establishing US global AI dominance, includes financial backing from OpenAI and technology support from Oracle, Microsoft and NVIDIA (among others).

DeepSeek has not upended the race for AI supremacy

There have been assertions that DeepSeek’s achievements have ended the race for AI supremacy between the US and China. The US is home to most of the world’s leading AI companies. It has also imposed bans on the export of high-end GPUs to China to weaken that country’s ability to compete in the AI race. In this context, DeepSeek’s ability to produce high performance models using lower-spec GPUs is impressive. Further, it suggests that rather than stymie innovation, the US export ban has acted as an innovation driver. But this does not mean that DeepSeek has stopped the AI race or given China a definitive lead. AI innovation is highly dynamic. As soon as a major break is announced, rivals find ways to incorporate it to drive further innovation. 

However, while DeepSeek may not have ended the race for AI supremacy, it will hopefully encourage AI stakeholders to reconsider what AI supremacy means. An important reflection is that innovation is not dependent on brute force, large-scale models and easy access to powerful, expensive compute infrastructure. DeepSeek has shown that agility, creativity, and the ability to think outside of the box can drive innovation. The arrival of DeepSeek should also prompt AI leaders to question core assumptions that sit behind notions of AI supremacy: that it produces fallout and has negative consequences, especially when it undercuts responsible AI by compromising on safety and equality in the race to get ahead.