The bear case for AI

Hello, Free Lunch readers. I’m Tej Parikh, the FT’s economics leader writer, and I am standing in for Martin Sandbu this week. In the same vein as my last piece — where I took on the “hot US economy” — I play contrarian again, this time with artificial intelligence.

“Narratives are a major vector of rapid change in culture, in zeitgeist, and in economic behaviour,” wrote Nobel laureate Robert Shiller in his 2019 book Narrative Economics.

Today’s dominant economic and market narrative is the transformative potential of AI. Although US interest rates have risen to their highest in two decades, and economic momentum is easing, the S&P 500 has been pushing higher, driven in part by the frenzy for AI-linked stocks.

But narratives can get ahead of themselves, and euphoria can be blinding. That makes it worthwhile to actively look for evidence that can raise doubt on conventional wisdom. (Notably, in recent weeks there have been murmurings of AI scepticism.) So, I trawled the latest research and spoke to a few “AI bears” for data points that challenge the bullish outlook. Here’s what I found.

1) It is still early days

AI is still in the so-called picks and shovels phase, when upfront capital expenditure is taking place before any major productivity gains can be reaped. This is evident from stock performance.

AI stocks can be grouped into three buckets: the infrastructure enablers (eg Taiwan Semiconductor Manufacturing Co, Arm), the software companies (eg Salesforce) and the adopters. Recently, semiconductor groups have had the most gains in their value, then the cloud, software and services companies. While some early adopters in information, manufacturing and technical fields have seen gains, valuations for businesses in industries with upside productivity potential remain quite tame.

So what? Well, AI has not yet proven to be adoptable at scale across the economy. That does not mean those gains will never arrive — most analysts forecast greater business integration of AI over the coming decade. But it is a reminder that the hype right now is driven mostly by the enablers of the technology, while its upside for business productivity — which will drive economic growth — is still largely theoretical, however optimistic it may look.

If the productivity gains do not come into view soon, it could derail the upward march of the enablers. At the end of June, Nvidia shares tumbled, and insider selling by top executives at the company took place at the fastest pace in years.

As AI bear Jim Covello, head of global equity research at Goldman Sachs, put it recently in a research note: “AI bulls seem to just trust that use cases will proliferate as the technology evolves.”

2) Where is the killer application?

That leads nicely to a key question: what if the end adopters do not benefit as much as the bulls think they might?

Earlier this year I spoke to Erik Brynjolfsson, a professor, author and senior fellow at the Stanford Institute for Human-Centered AI for an FT Economists Exchange. He was optimistic about the potential economy-wide productivity gains from AI adoption. But he warned about what he called the “Turing trap”.

The Turing test was introduced by Alan Turing in 1950. The idea was to set out criteria to measure a machine’s ability to exhibit intelligent behaviour equivalent to a human. But Brynjolfsson reckons it has inadvertently inspired a generation of researchers to make machines that emulate human abilities. “I think it is becoming apparent that it was the wrong goal all along and that we should be thinking how to augment humans and extend our capabilities,” he said.

That leads me to another Erik. Erik Hoel, an American neuroscientist, posits that the industries AI are disrupting are not all that lucrative. He coined the phrase “supply paradox of AI” — the notion that the easier it is to train AI to do something, the less economically valuable that thing is.

“This is because AI performance scales based on its supply of data, that is, the quality and size of the training set itself,” said Hoel. “So when you are biased towards data sets that have an overwhelming supply, that, in turn, biases the AI to produce things that have little economic value.”

Hoel raises an interesting point. Generative AI’s current applications include writing, image and video creation, automated marketing, and processing information, according to the US Census Bureau’s Business Trends and Outlook Survey. Those are not particularly high value. Using specialist data, sophisticated models could do deeper scientific work, but that data can be in short supply or even restricted.

The point is that with the AI infrastructure buildout cost projected by some to be more than a trillion in the coming years — what trillion-dollar problem will AI actually solve? To cite Covello: “Replacing low-wage jobs with tremendously costly technology is basically the polar opposite of the prior [lucrative] technology transitions.”

3) Do the capex plans even add up?

Right, so how farfetched do the projected AI capex and AI revenue figures seem? For measure, a few analysts have done back-of-the-envelope calculations, using various assumptions.

David Cahn, a partner at Sequoia, is not an AI bear but thinks revenue expectations will need to pick up. He has tried to reconcile the gap between the revenue expectations implied by the AI infrastructure buildout and actual revenue growth in the wider AI ecosystem.

He took Nvidia’s run-rate revenue forecast, and doubled it to cover the cost of AI data centres. “GPUs are half of the total cost of ownership — the other half includes energy, buildings, back-up generators,” he noted. He doubled that figure again to incorporate a 50 per cent gross margin for the final graphic processing unit user. That leads to a rough and ready figure of $600bn in AI revenue needed to pay back the upfront capital investment. (This excludes margin for cloud vendors, which would make the revenue requirement higher).

Barclays came to a similar conclusion, using a different approach. It estimates cumulative incremental AI capex between 2023 and 2026 of $167bn across top players in the industry. It reckons that is enough to “support over 12,000 ChatGPT-scale AI products”. But it is unsure that there is enough consumer and enterprise demand to absorb this amount.

Another factor here is competition. “LLM [large language models] . . . have become increasingly indistinguishable from one another,” noted Peter Berezin, chief global strategist at BCA Research. “They may end up functioning more like highly competitive airlines with thin profit margins rather than monopolistic social media platforms.”

The point? It is basic maths — with numerous assumptions — but it does point to capex spending today far exceeding the potential returns.

4) The macro impact remains unclear

There have been numerous studies over the past 18 months that estimate the size of the potential AI productivity growth gain. Two have stood out, partly because they end up at different ends of the spectrum.

First is from Goldman Sachs economists Joseph Briggs and Devesh Kodnani, who last year forecast a 9 per cent rise in total factor productivity and 15 per cent increase in US GDP following full adoption.

Second is MIT economist Daron Acemoglu’s forecast this year of just a 0.5 per cent increase in TFP and a 0.9 per cent rise in GDP in the next 10 years.

The difference comes down to three differences in modelling:

i) The share of automatable jobs: Acemoglu assumes GAI will automate only 4.6 per cent of total work tasks in the next 10 years, whereas Goldman’s baseline is 25 per cent over the long run.

ii) The effects of labour reallocation or the creation of new tasks: Goldman estimates the uplift from displaced workers being re-employed in new occupations made possible by AI-related advances and new tasks that boost non-displaced workers’ productivity. Acemoglu’s modelling focuses on cost savings primarily.

iii) Cost savings: Goldman is more bullish here in part because it expects AI automation to create new tasks and products.

This underscores how differing assumptions of AI’s automatable potential, and its ability to create new activities and lower costs, can drive swings in its projected impact on national-level productivity. While we are getting more clarity on each element, a lot of uncertainty remains. Most investment today is based on firm-level studies of potential productivity gains, but that does not always extrapolate well to the national or global level.

Building on this, ING Research says larger sectors may not even be in a position to use AI, thereby limiting the technology’s near-term economic impact. Its economists argue that the more digitalised European sectors, which tend to be the smallest relative to the economy, are in a better place to implement AI, and experience productivity improvements.

5) The enabling environment

Even if a killer AI application is found, there is still no guarantee that its economic impact will be transformative. As my conversation with Brynjolfsson highlighted, the broader economic, social and legal environment also needs to shift to allow economies to harness the technology’s benefits, and minimise its harms. “Our understanding of the skills, the organisations and institutions needed is not advancing nearly as fast as the technology is,” he said. Here are a few factors that will determine both the pace and level of AI transformation:

i) Energy. The AI industry could consume as much energy as a country the size of the Netherlands by 2027. With net zero targets, that energy must also be clean. Grids need to be rapidly connected, and permitting needs to be swift to get the infrastructure up alongside the AI capex.

ii) Regulation and governance. AI can also be harmful. Deepfakes, privacy violations, market volatility (caused by AI trading for instance) and cyber crime can be counter-productive. The problem is that regulation is running far behind the technology, and at different paces globally.

iii) Society. How AI interacts with society also matters. For instance, GAI has been tipped to capture revenues from creative sectors. But there is opposition both from those employed in these sectors, and the public, who still want a human touch in some industries. Hollywood writers, for example, were able to set up guardrails for how AI is used in the industry. And even then if there are significant automation-related job losses, social unrest and inequality could stymie growth, particularly if retraining initiatives are not widespread.

iv) Skills. Job postings mentioning “natural language processing”, “neural networks”, “machine learning” or “robotics” have picked up. But skillsets will take some time to match the demand. The IBM Global AI Adoption Index 2023 found limited AI skills and expertise as the top barrier hindering businesses’ successful AI adoption today.

The point is that AI’s potential productivity impacts do not matter if the enabling economic and legal environment cannot be put in place to take advantage of it — the AI transition relies on more than just the AI innovators.

These should all add at least a hint of doubt on the thus far exuberant AI outlook. Free Lunch would be interested in your bearish findings too.

Of course, it is early days, new AI applications will arise and adoption should become easier. Nor is the explosive capex necessarily a bad thing. Bubbles can be destructive, but must be weighed against the overall impact on economic capacity — the railroad bubbles in the 19th century burst painfully, but left valuable infrastructure. Perhaps the euphoria is a necessary vehicle to get money into a potentially transformative, but not yet proved, technology.

Either way, it does little harm to step back and reassess one’s assumptions. Narratives are by design appealing, but could be meaningless if they cannot stand up to scrutiny.

Other readables

The troubles of Europe’s battery industry reveal what is wrong with EU green industrial policy, writes Martin Sandbu.

Who is the UK’s new chancellor of the exchequer? Read the FT’s in-depth profile of Rachel Reeves. And Chris Giles explains why you should pay attention to Reeves’s fiscal statement later this month: it may reveal a lot more about how the Labour government will run the economy than yesterday’s King’s Speech.

Ahead of a plenary meeting of the Chinese Communist party’s Central Committee, the country’s official growth rate is slowing, and below the government’s target. That seems to be fuelling a multi-faceted social crisis and rising popular frustration with unfairness and inequality.

More on the similarities and differences between far-right parties in different European countries, from our very own John Burn-Murdoch.