The Chinese AI could spur adoption of the technology, increasing demand for chips from the likes of Nvidia
Alok Sama
February 2nd, 2025
The Sunday Times
A week ago you might have been excused for thinking that a DeepSeek was an unusually cerebral Sikh gentleman. That was before its founder, the enigmatic Liang Wenfeng, wiped out almost $1 trillion in market value globally. DeepSeek’s emergence followed hot on the heels of President Trump’s announcement of the $500 billion Stargate artificial intelligence infrastructure project— which itself added a comparable amount to already frothy tech valuations. This is the type of volatility one encounters at the apex of market bubbles and forces one to ask some, well, deep questions.
To start with, a cynical view. Liang’s day job is as a hedge fund manager, his inspiration the late Jim Simons — the man whose quant models “solved” the market. Did Liang short Nvidia stock and then release the source code for DeepSeek’s R1 reasoning model to set up the greatest bear trade ever? That, too, on the eve of the Chinese year of the snake. And what does one think of someone who, presumably unironically, names his outfit High-Flyer Investment Management?
Is this all part of (yet another) Chinese conspiracy to tap into everyone’s data, a chatbot version of TikTok? Is DeepSeek simply a lightweight and plagiarised version of OpenAI, created using a technique the tech bros call “distillation”, which is explicitly forbidden by OpenAI’s terms of service when used to create a competing large language model? Did it exfiltrate — a geeky euphemism for “steal” — data from OpenAI? That would be delectably risible given OpenAI’s own copyright infringement battles with content creators of all varieties.
All of this makes me sound like I’m long Nvidia, or still employed by SoftBank, neither of which is true. The success of TikTok was an early indicator that the ingenuity of Chinese AI talent should not be underestimated. My former boss Masayoshi Son identified early the brilliance of the TikTok algorithm, which promotes content not based on the follower count of the creator, but on the quality and appeal of the content itself, allowing new creators to go viral. This was a profoundly elementary premise. As is DeepSeek’s efficient computations: for example, its AI model recognises that calculating complex sums to eight rather than 32 decimal places is good enough.
There’s even more to DeepSeek, including the equally obvious notion that while searching information in a library, one needn’t scan the entire library but rather start by making some judgments about which section is most likely to provide an answer. All of which implies, as Microsoft chief executive Satya Nadella concedes, a dramatic increase in the efficiency of AI models. The prominent Silicon Valley venture capitalist Marc Andreessen was more effusive, calling DeepSeek “one of the most amazing and impressive breakthroughs I’ve ever seen”.
On a recent trip to India, Sam Altman, the OpenAI chief executive, was asked by an Indian venture capitalist whether startups could build foundational AI models such as ChatGPT. Altman dismissively responded that this was a “totally hopeless” endeavour. That a Chinese upstart has done so must be jarring for Altman, particularly so on the heels of the announcement of the grandiose Stargate datacentre infrastructure project, labelled by some — tastelessly, given Altman’s partner and financier is the Japanese SoftBank Group — as the new Manhattan Project.
DeepSeek has shown that running AI models does not always require power-hungry cloud computing infrastructure but can be done at the edge of the network, for example on laptops and handheld devices. And access to capital and the latest generation of Nvidia chips are not the competitive moats one believed they were only a week ago. Sure enough, Nvidia stock took a one-day 17 per cent hit on the DeepSeek news. The basic premise of Stargate seemed questionable, and in the UK, some suggested that the Labour government’s new AI policy, with the expansion of datacentre infrastructure at its core, needed to be re-examined.
But here is an alternative take, and one I subscribe to. In 1765 James Watt’s steam engine dramatically improved efficiency, reducing coal consumption by about 75 per cent relative to legacy technology, in this case Thomas Newcomen’s engine of 1712.
A hundred years after Watt’s breakthrough, William Stanley Jevons, the British economist, analysed total coal consumption and posited a paradox that has been a central feature, along with network effects, of most disruptive technologies in the digital era. The decline in cost promotes mass adoption which, in the case of AI, will probably lead to a rise rather than fall in required compute capacity, both in datacentres and at the edge of the network. (The latter, incidentally, favours the British company ARM, which rules when it comes to edge processing with its energy-efficient chip designs.)
In his keynote address at the Consumer Electronics Show, the formidable Jensen Huang, chief executive of Nvidia, highlighted two AI mega trends to watch in 2025: “agentification” of AI, and physical AI. The former refers to software bots that can make our lives easier: for example, OpenAI launched ChatGPT Operator, a tool for repetitive tasks such as filling forms, booking travel and scheduling. I suspect it is clumsy to begin with but gets better by the day; these things always do. As for physical AI, I took my first Waymo (Google’s self-driving platform) ride in San Francisco recently and experienced my first “aha” moment in tech since I first had an iPhone in my hand in 2007.
All of which means that I’d be wary of short selling Nvidia. For now anyway.