DeepSeek – what just happened in AI?

29 January 2025

Stockmarkets sold off with AI infrastructure players losing up to 10% of their value, as Chinese AI model DeepSeek showed results on a par with OpenAI at just 5-10% of the cost. Mark Hawtin, Head of the Liontrust Global Equities team discusses DeepSeek and what this means for AI, markets and investors.

Stockmarkets sold off hard with AI infrastructure players losing up to 10% of their value as the relatively little-known Chinese AI model DeepSeek jammed the airwaves as its latest head-to-head model results showed results on a par with OpenAI at just 5-10% of the cost.

Just days after the US announced its AI infrastructure project ‘Stargate,’ with plans to spend $100 billion and rising up to $500 billion over the next four years, focus shifted to the relatively obscure DeepSeek – a Chinese AI company building open source Large Language Models (LLMs). If true, does this imply the $500 billion US spend is really worth far less or, put another way, doesn’t require so many dollars to achieve the same results.

Before exploring the implications of what DeepSeek has achieved, it’s important to first understand the technical foundation of the company.

What is DeepSeek?
DeepSeek was founded in 2023 as an offshoot of a quant hedge fund and has quickly risen to fame with its rapid release of efficient open-source AI models challenging the Western tech giants. Its DeepSeek-V3 model, released in December 2024, was highly regarded for its cost effectiveness, which was trained with only 2.8 million graphic processing unit (GPU) hours, approximately 2,000 Nvidia chips and spend on compute power of approximately $6 million. Compare that to similar models typically costing upwards of $60 million and Meta’s Llama using 30.8 million GPU hours.

In our view, the DeepSeek buzz is justified and more notable when considering the US imposed restrictions on Nvidia’s AI chips to stall AI advancement in China in an “AI Arms Race.” However, this looks to have forced efficiency and created novel architectures such as a Mixture-of-Experts (MoE) which only activates the necessary parameters and optimises communication between GPUs. It employs techniques like DualPipe algorithms to reduce training costs.

DeepSeek also uses an approach called multi-head latent attention, which compresses attention mechanisms to minimise memory usage during inference, enhancing speed without sacrificing accuracy.

When combining with low-precision training techniques like FP8, this can reduce GPU memory demands by 40%. More excitingly, DeepSeek uses reinforced learning, mimicking human learning of self-improvement rule-based reward systems through trial and error – differentiating itself from competitors relying on supervised fine-tuning.

By cutting back on the compute needed and being able to handle tasks with just 5-12% of the energy consumption used by state-of-the-art models and an estimated 90-95% cheaper than OpenAI’s offerings, this raises the question of what now for capex spend and AI infrastructure names?

We have been making the point for some time now that the while the AI opportunity is huge, the capex cycle was running an asymmetric risk profile and, as a result, we have been reducing exposure to the capex names.

To be clear, we don’t think it changes the capex spend dynamic in the short term − the hyperscalers will still invest heavily. Meta stated last week that its capex for 2025 is expected to be $60-65 billion vs. the market expectation of $51billion.

At the margin, if it makes everything more accessible and cheaper then it will benefit as well. However, we could see Nvidia impacted – as a start, it probably loses the inflated margins it has been achieving over the last two years – that’s a 10-15 percentage point risk.

Democratisation of AI
DeepSeek’s open architecture isn’t just about cheaper AI – it’s also a power shift. We could witness an explosion of new applications by decoupling AI advancement from high capital costs.

Open models enable customisation for niche-use cases that are often overlooked by Big Tech’s one-size-fits-all strategy, fostering creativity across all industries. Consider the impact of the iPhone and the app store back in 2007, no one could have predicted the level of innovation and disruption that the app store would bring.

As the costs of AI continues to decrease, we are poised for a new wave of innovation and disruption.

What next – where to invest?
We have been crystal clear that the opportunity for AI is immense and nothing in the DeepSeek news changes that. In fact, it can only help the speed of adoption as the infrastructure becomes cheaper and cheaper and potentially a commodity.

It’s worth flagging Jevons Paradox. When Jevons studied coal consumption in relation to steam engine efficiency in the mid-19th century, he noticed that as steam engines became more efficient, coal consumption increased rather than decreased. As AI gets more efficient and accessible, we could likely see AI skyrocket and turn into a commodity.

Set against this backdrop, our investment thesis is reinforced. AI is definitely the right thesis. AI infrastructure players may well become victims of the right thesis, wrong valuation mantra. However, there will be a broadening out of the opportunity set. Users of AI to drive productivity and stronger moats will prevail. This is early in the cycle and there will be some amazing investment opportunities. Any turmoil around the theme in the short term should be seen as a chance to position portfolios to AI use case winners.

Past performance is not a reliable guide to future returns. You may not get back the amount originally invested, and tax rules can change over time. The writer’s views are their own and do not constitute financial advice. 

This information should not be relied upon by retail clients or investment professionals. Reference to any particular investment does not constitute a recommendation to buy or sell the investment

Main image: ryan-quintal-US9Tc9pKNBU-unsplash

Professional Paraplanner