AI's civil war will force investors to pick sides

Reuters

07-02-2025

LONDON, Feb 7 (Reuters Breakingviews) - In their satirical history of the United Kingdom, '1066 And All That', the authors W. C. Sellar and R. J. Yeatman cast the English civil war of the 17th century as a conflict between the 'Wrong but Romantic' Cavaliers and the 'Right but Revolting' Roundheads. The aftermath of the release last month of Chinese startup DeepSeek's R1 artificial intelligence model, which matches or outperforms existing offerings from U.S. technology titans at a fraction of the cost, has exposed a similar divide among the world's leading innovators in the field of machine learning.
On one side are those who strive for artificial general intelligence (AGI), the point where machines match or surpass human capabilities. Let's call them AI Cavaliers. Facing them are AI Roundheads who are focused on the more mundane goal of solving specific problems as efficiently as possible. Deciding which side to back in this AI civil war will be a defining decision for investors in the world's hottest technology.
The AI revolution that has gripped global stock markets for the past two years is driven by three epochal trends. The first is the generation of vast volumes of machine-readable data by the digitisation of almost every aspect of daily life. The second is the collapse in cost of computing power prompted by ever more efficient chips. The third is a dramatic improvement in machine learning algorithms – the software that computers use to extract the signal from the noise in data sets. Together these developments have sparked a step-change in the accuracy of predictive modelling.
Every technologist agrees that this revolution is a momentous shift for the world. Where they diverge is on the question of where it can be most valuably applied. DeepSeek's dramatic intervention has thrown the differences between two visions for AI into stark relief.
The AI Cavaliers have a romantic vision of what the new machine learning algorithms can achieve. They see it as the royal road to the creation of thinking machines empowered with AGI. Their champions are well-known chatbots such as OpenAI's ChatGPT and Anthropic's Claude. Their weapon of choice is the large language model (LLM), which uses AI's prodigious powers of pattern recognition to predict the next word in a string of text, with mind-bogglingly coherent results. The data they aspire to crunch is no less than the totality of human knowledge – or at least, everything that can be scraped from the internet. Their appetite for computing power is therefore similarly vast. Indeed, in principle, it's limitless.
This intoxicating vision is the stuff of science fiction. So it's no surprise that the launch of ChatGPT's original iteration in November 2022 seized the public imagination and ignited a stock market boom. Yet it is haunted by three big questions. The first is the fiercely disputed technical conundrum of whether LLMs can achieve AGI. The second is the commercial dilemma over whether the models have any enduring competitive advantage. Finally, there is the great financial unknown of how much capital spending these ventures will require in the form of semiconductors, data centres, and energy.
The release of DeepSeek's model amplified all these doubts. Shares in Nvidia (NVDA.O), opens new tab, the leading AI chip producer, dropped 17% in a day, wiping out a record of $600 billion of market value. Energy companies whose stocks had risen on forecasts of galloping electricity demand for AI training also took a hit. In the fortnight since, shares in both Microsoft (MSFT.O), opens new tab and Google owner Alphabet (GOOGL.O), opens new tab sank following trading updates which cast doubt on their vast investments in computing capex. Put bluntly, DeepSeek has cast doubt on the return on investment in AGI.
The AI Roundheads, by contrast, are a very different tribe. They see machine learning in much more practical terms. For them it is a tool to crack specific, well-defined problems, not an end in itself. Rather than force-feeding LLMs indiscriminately with data from the web, their models are trained on specialised, proprietary datasets. In contrast to the unbridled extravagance of the Cavaliers' demands for hardware and power, their models are puritanically frugal.
Their icon is Google DeepMind's AlphaFold2, developed by the search giant's London-based AI research arm. Released in 2022, it applied tailor-made deep learning algorithms to a carefully selected database of 170,000 existing proteins to predict the structures new ones will take, knowing only their constituent amino acids. Training took a matter of weeks and required the equivalent of under 200 chips, opens new tab. That was sufficient to crack a critical practical problem which had eluded scientists for 50 years, and won its inventors a Nobel Prize, opens new tab in chemistry.
For the AI Roundheads, in other words, DeepSeek's latest model confirmed what they already knew. With the right data, carefully targeted algorithms, and a well-defined problem to solve, there is no need for vast investment in high-end hardware. AI's three components – data, computing power, and algorithms - are substitutes. Better algorithms can work with messier data and less computing power. More relevant datasets reduce the need for faster chips and expensive programming talent. The economics of the AI Roundheads' approach thus has a compellingly simple logic. Take aim at a well-defined predictive challenge and choose your inputs to get there at least cost.
Which side should investors take in this battle? It would be foolish to deny that the Cavaliers' dream of AGI is uniquely captivating, or that their LLMs are remarkable things. Yet when it comes to adding concrete economic value, the Roundheads look better placed for victory – just as they prevailed in the English civil war.
In December, DeepMind unveiled GenCast, opens new tab – an AI model trained on historical meteorological data which beat traditional statistical tools at forecasting wind power production and extreme weather events 97% of the time. Meanwhile, DeepMind founder and CEO Demis Hassabis last month suggested, opens new tab that the company's Isomorphic Labs spinoff will have AI-designed drugs in clinical trials by the end of 2025.
Investors do not need AI to spot which is the Big Tech company where the 'Right but Revolting' AI Roundheads hold most sway.
For more insights like these, click here, opens new tab to try Breakingviews for free.