ARTICLE | 4 MIN | VIEWS FROM THE FLOOR

DeepSeek Breaks the AI Capex Model, But What Comes Next?

January 28, 2025

DeepSeek’s innovation has profound implications for AI capex, adoption, and global competition. No wonder it rattled markets.

On Monday, DeepSeek, a hitherto unknown Chinese AI lab, upended all assumptions about the economics of AI, triggering a $1 trillion rout in US and European technology stocks.

DeepSeek unveiled a new large learning model, claiming it built this in just two months, for less than $6 million, leveraging existing technology and models — compared to the $5 billion OpenAI reportedly spends annually. The company allegedly achieved this with just 10,000 Nvidia graphics processing units (GPUs) — specialised processors optimised for the large-scale computations required in AI— representing a fraction of the hardware deployed by hyperscalers like Amazon and Microsoft, large cloud service providers which operate massive data centres to power cloud computing and AI services.

In our view, DeepSeek has discovered and openly shared a series of clever and novel techniques that drastically reduce the cost of training large AI models, and this will have a profound impact on the tech sector. 

Its transparent approach will likely spark an explosion of new models and drive costs even lower. Cheaper training also improves the ‘return on capex’ for hyperscalers, making AI investments more efficient. While its claims still require independent verification, the lack of pushback from industry commentators, analysts, and the market reaction itself suggests this breakthrough is being taken seriously.

Short-term capex slowdown

The sell-off reflects fears of a ‘deep air pocket’ in AI infrastructure spending. For years, the tech sector has built investment cases on the belief that developing competitive AI models demands massive hardware investments and seemingly unlimited budgets.

Historically, cheaper technology drives much broader adoption, and the same could happen with AI over time. However, this step-change in model efficiency suggests that we are likely to see a meaningful pause in AI capex, as the industry adjusts to the new reality.

One of our most consistent views for the last 12 months has been that ‘necessity is the mother of all innovation’, and although the efficacy and speed (pending verification) have surprised us, the fact that China tech is catching up shouldn't be a surprise. It's been forced to turn to innovation given the constraints around chip supply they have had imposed on them.

By using older hardware and novel techniques, DeepSeek has shown that cutting-edge AI development is no longer the exclusive domain of companies with access to the latest chips. This raises further questions about the sustainability of US leadership in AI technology and the long-term impact of export restrictions.

For us, this validates our long-held concerns about challenges in semiconductor capital equipment and data-centre supply chains, as competition intensifies and oversupply risks materialise. DeepSeek’s breakthrough accelerates this shift.

The opportunity

At the same time, DeepSeek highlights the enormous opportunity that affordable AI presents. By cutting training costs and hardware requirements, it democratises access to AI development, paving the way for a proliferation of new models and applications. This could trigger a significant upgrade cycle for existing compute infrastructure, as generative AI moves from being centralised in hyperscale data centres to running on smaller, more distributed devices.

Inference — using AI models for predictions or tasks after they’ve been trained — could soon run directly on smartphones, PCs, cars, and alternative reality (AR) glasses. This marks a potential ‘Cambrian explosion’ of AI adoption, where its use becomes ubiquitous, transforming industries and everyday life.

On the other hand, this could also make enterprise adoption of AI much more palatable by solving the cost issue. If this represents a step-change in AI scaling laws — rules governing how performance improves with added resources — GPU productivity has significantly improved, enhancing the return on investment (ROI).

Existing installed GPUs could be reused for these more efficient models, lowering the cost barrier to adoption. This could benefit software providers and vertically integrated tech giants like Facebook and Amazon, which can improve ROI on their in-house AI services. Of course, the debate of whether companies will ‘build’ in-house or ‘buy’ externally offered AI solutions remains open, further complicating the competitive software landscape — enhancing the case for more nuanced stock-picking.

The tech sector thrives on constant innovation and self-disruption, but DeepSeek’s breakthrough underscores how quickly assumptions can be overturned. Investors must not only identify trends but also understand where we are in the monetisation cycle, identifying timing and profit pools.

All data Bloomberg unless otherwise stated.

With contributions from Sumant Wahi, Portfolio Manager at Man Group, focusing on technology equities, and Nick Wilcox, Managing Director, Discretionary Equities at Man Group.

For further clarification on the terms which appear here, please visit our Glossary page.