Parts of the stock market, notably those with exposure to artificial intelligence (AI), experienced considerable downward pressure on Monday (27 January). The reason behind this is DeepSeek, a Chinese lab, shocked the tech world after producing a large-language model (LLM) or advanced reasoning model, called DeepSeek R1.

This model delivered comparable performance to the world’s best chatbots at seemingly a fraction of the cost. And over the weekend it was downloaded more than ChatGPT.

How DeepSeek shocked tech

DeepSeek-R1’s an open-source model with comparable performance to Western AI systems like ChatGPT. The model’s success has sparked concerns about the future dominance of US tech giants, causing a sell-off of US tech stocks. This development’s seen as a potential equaliser in the AI field, particularly benefiting researchers and developers with limited resources.

Diving deeper, DeepSeek-R1 impresses because of its cost effectiveness. Trained for just $5.6m, it operates at 95.3% less cost than Anthropic’s Claude 3.5 Sonnet and charges users only 2% of OpenAI’s O1 model rates. Its four-stage training process, including large-scale reinforcement learning on reasoning problems, has achieved O1-level performance at a fraction of the cost of its peers.

However, there’s another angle. DeepSeek R1’s claimed to be built on “less advanced” or “last generation” Nvidia GPUs (Graphics Processing Units), in line with export controls on the most advanced chipsets.

Some analysts are questioning whether this is possible and it may be a matter of time before we find out more. DeepSeek published its methodology, meaning it can be copied and evaluated. Some analysts suggest that China has evaded export restrictions and R1 could have been built on higher-end chips.

What does this mean for tech?

This development challenges the assumption that cutting-edge AI requires massive investments in advanced hardware and infrastructure. In addition to companies like Nvidia, stocks in the data centre segment have taken the biggest hit. These are the firms responsible for building the AI infrastructure that the sector has been clamouring for over the past year. This includes rack-scale providers, cooling companies and those in networking.

As I write, Celestica‘s (NYSE:CLS) among the hardest hit stocks. Sadly, it’s one of my favourites with a price-to-earnings-to-growth (PEG) ratio consistently under one. Celestica’s a leading provider of electronic manufacturing services (EMS) and operates multiple manufacturing facilities worldwide. The company specialises in producing complex electronic components and systems for various industries.

Celestica’s business model heavily relies on serving large technology companies, particularly hyperscalers like Amazon and Google who are investing heavily in AI and data centres. In Q3 2024, hyperscaler revenue increased 54% to $761m, accounting for 30% of total revenue. Thus, if we’re seeing a massive breakthrough in AI cost-efficiency, Celestica may see a slowdown in sales growth.

However, there’s certainly a possibility all of this is being overplayed. After all, greater LLM efficiency combined with the rapid expansion of data centres could represent another quantum leap forward for AI, making it more accessible and compounding performance gains. Moreover, cheaper AI should hasten its adoption, particularly among SMEs and the broader market. This could become apparent in the coming months.

I may have considered buying more Celestica, but it’s already my largest holding by some distance. I don’t want to become over-concentrated.