DeepSeek: Sputnik moment or sobering reminder?

DeepSeek’s late January reveal certainly made some waves, with many dubbing it a “sputnik” moment, representing a seismic global shift in power and prestige when it comes to AI. Markets followed in kind: The S&P 500 fell by 1.5% while the Nasdaq Composite fell by 3%. Nvidia, the GPU company thought to be indispensable to the development of AI in the west, had $600 billion wiped off its value.

What’s the big deal?

After a quiet development and explosive reveal, DeepSeek showed that it could create an AI model of comparative power to its world-leading competitors with far lighter data and hardware requirements.

This both shocked the market and great multitudes on social media and in the press. However, we should all take a deep breath. This is actually a rather predictable evolution, rather than a watershed moment.

Much like previous advances in computing, AI products are becoming smaller, faster and cheaper. It just so happens that DeepSeek is a more efficient product, with a superior price-to-performance ratio. In fact, the Chinese startup announced that it trained its model for less than $6 million and over only two months, using far fewer GPUs than competitors like ChatGPT.

There are important lessons here though. AI’s explosion into public consciousness in the last few years has precipitated a flood of investment into AI projects, imagining that the payoff will be quick and enormous. But that hasn’t quite happened yet. AI projects come with huge costs in terms of capital and energy, yet a truly transformative use case has yet to emerge.

AI costs are huge….

The development of an AI system is an extraordinarily expensive task. The hardware required to build up computational power is significant, crucially relying on masses of expensive GPUs to scale. From there, massive amounts of data are required to train those models, which is not only expensive but increasingly hard to acquire given the ever growing demands of these models. On top of that come enormous energy costs. Tech giants are reporting large increases in their emissions, likely driven by their AI demands. Meanwhile the International Energy Agency (IEA) - who reported that data centers used up around 2% of global energy in 2022 - predict that will double by next year.

It’s expensive now, and it's getting more expensive every day. One report from last year announced that the computational power needed for training an AI model doubles every year, in turn that grows other costs like data needs and energy. That’s largely been driven by the heated race that technology giants are currently engaged in, with early mover advantage being critical. Where that race is headed, or whether it will justify the costs, is unclear.

…but returns are elusive

Last year, Goldman Sachs released a report entitled “Gen AI: Too much spend, too little benefit?” which underlined the potential trap that many AI hopefuls are now in. Darren Acemoglu, an Economics Institute Professor at MIT, is interviewed within the report - says that only about 5% of AI exposed tasks will be cost-effectively automated within the next decade. Ultimately, Acemoglu thinks that US productivity will only be improved by half a percentage point over the next decade. It’s not nothing, but it's hardly the imminent “revolution” that many herald.

Also interviewed within the paper is Jim Covello, Goldman Sachs’ Head of Global Equity Research, who points out that the incredible costs of developing and operating AI technology may eclipse its profitability. He notes “AI infrastructure buildout will cost over $1tn in the next several years alone, which includes spending on data centers, utilities, and applications. So, the crucial question is: What $1 trillion problem will AI solve?” Covello points to the fact that Goldman Sachs had discovered that AIs can update the firm models from historical data faster than a human but at six times the price. He notes “AI technology is exceptionally expensive, and to justify those costs, the technology must be able to solve complex problems, which it isn’t designed to do.”

DeepSeek rings the alarm

DeepSeeks explosive reveal should remind both VC investors and AI companies that Return on Investment Capital (ROIC) matters. Many companies have gone all in on AI but early gains in the area are now slowing, or coming in as less than expected and performance is plateauing. Many may opt to throw more GPUs and data at their AI models in order to expand their processing power, but that will add incredible expense and won’t necessarily solve the problems they want to solve. AI companies are now building things which are exciting, but hard to really monetise, much less recoup the investment given the incredible costs of developing such a project. If DeepSeek’s example shows anything, it's that efficiency should be a cardinal virtue in AI development.

The recent reports of the jailbreaking susceptibility of DeepSeek's latest model reminds us that these systems also present new security challenges. It is more important than ever for companies to protect their dynamic runtime application environments, especially when AI is being utilised.

Judging from headlines about AI, it's easy to see how these things can get confused. There is simply so much doomsaying, dreaming, marketing hype and press attention around this technology that it's often very hard to separate fact from fiction. The reality is we’re just too early to know what, if any fundamental productivity improvements AI will bring. As Robert Solow - the great MIT economist - said during the PC revolution - “you can see computers everywhere but in the productivity statistics.”