DeepSeek is a Chinese artificial intelligence (AI) research lab that spun out of a Chinese hedge fund called High-Flyer. The launch of DeepSeek’s R1 model was accompanied by a research paper detailing several key innovations which allowed the model allegedly to be trained for far less cost than leading Western counterparts, while delivering comparable cutting-edge performance. The costs of running live data through a trained AI model to then make a prediction or solve a task (known as inference costs) when using the model were also significantly cheaper.
R1 sent shockwaves through equity markets as investors scrambled to understand the implications for future AI infrastructure spending, and NVIDIA lost c$600bn of its market capitalisation (-17%) in one trading session – the most of any company, ever. Markets soon recovered their footing as the ‘true’ training costs of the DeepSeek model were estimated to be perhaps north of $1bn when considered on an apples-to-apples basis to Western counterparts. Within days, upward revisions to Meta Platforms (Meta), Google, Amazon and Microsoft’s plans for capital expenditure on AI infrastructure calmed investor nerves.
The collapsing cost of using models soon after their introduction has been a constant feature of the AI landscape. The cost of inference had fallen 20x even before DeepSeek was revealed. Indeed, the open-source nature of Meta’s Llama models (from which DeepSeek was reportedly ‘distilled’) is intended to create these types of innovation at the ‘trailing edge’, refining, advancing and ultimately commoditising the leading-edge work.
The more important issue for investors to consider is the broader implication of ever-cheaper and ever more powerful AI. Satya Nadella, CEO of Microsoft, wrote of the DeepSeek news: “Jevons' Paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning into a commodity we just can’t get enough of.” So what is Jevons' Paradox? Writing in 1865, economist William Stanley Jevons argued that the increasing efficiency of coal use (e.g. steam engines that extracted more work from the same amount of coal) would – paradoxically - lead to more coal being consumed. The phenomenon is applied today to describe situations where technological developments that increase efficiency paradoxically result in an increase in overall use of the technology rather than a decrease.
At Polar Capital, we believe AI will prove to be the next general-purpose technology in the mould of electricity or the internet and we expect every industry to be reimagined in its wake. Existing profit pools will be reallocated and huge new industries will be created. PhD-level intelligence will likely soon be available on demand for a negligible incremental cost, the implications of which will be far-reaching as new applications become economical. These might include ubiquitous real-time AI in every device, embedded in everyday objects, providing constant personalised assistance. It may mean models that were previously deemed too computationally expensive to deploy in production (e.g. extremely large language models for niche applications, hyper-personalised recommendation systems) become feasible. We could see an explosion of AI applications in every sector of the economy and in our daily lives because using AI has become so much more efficient.
In the Polar Capital Artificial Intelligence Fund, which mainly invests in non-technology AI beneficiaries, we are lucky to have a front row seat for the early adoption phase and are already seeing a wide range of companies adopt AI. They should stand to benefit as AI gets cheaper and more efficient (as Jevons suggested), and can be applied economically to an ever broader, more complex set of use cases.
As AI gets more efficient and accessible, we will see its use skyrocket, turning into a commodity we just can’t get enough of.
Axon Enterprise, manufacturer of TASERs and police bodycams, has a new AI tool called Draft One. This takes the audio feed from a police officer’s camera and auto-drafts incident reports, saving officers 82% of the time it would normally take. This is not a trivial saving given report writing consumes up to 40% of an officer’s time, leading to one third of UK police officers considering leaving the police service.
We are close to reaching real-time AI assistance from the complete range of medical disciplines on Intuitive Surgical’s robots, or a researcher having access to cross-disciplinary scientific expertise in real time as they conduct experiments. Low-cost inferencing of more advanced AI will allow companies such as Walmart, Tesco and Dick’s Sporting Goods to solve the most complex inventory management and logistics problems to reduce wastage and maximise sell-through.
Amara’s Law states: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run”, but this may not be correct in the case of AI. We are extremely optimistic on the speed of AI adoption given the cost to use a given level of AI falls around tenfold every 12 months. The comparable figure for semiconductor advances (Moore’s Law) saw costs fall c2x every 18 months.
There has never been a more exciting time to be a technology investor, and DeepSeek should be seen as Jevons himself would have seen it: “The more we render it efficient and economical, the more will our industry thrive, and our works of civilisation grow."
Ali Unwin
Fund Manager, Polar Capital Global Technology
Polar Capital LLP Polar Capital LLP is authorised and regulated by the Financial Conduct Authority (“FCA”) in the United Kingdom, and the Securities and Exchange Commission (“SEC”) in the United States. Polar Capital LLP’s registered address is 16 Palace Street, London, SW1E 5JD, United Kingdom, Company number OC314700.
The value of securities and the income from them can fall as well as rise. Past performance should not be seen as an indicator of future returns. All views expressed are those of the author and should not be considered a recommendation or solicitation to buy or sell any products or securities.
