Fears of an AI Bubble Seem Overhyped: A Rebuttal to the Goldman and Sequoia Reports

A few weeks ago, Goldman Sachs and Sequoia published reports (here and here) that presented some skepticism about an AI-fueled rally for markets and the economy. They were well-reasoned, nuanced articles, and diverged from the groupthink that has emerged about AI over the last two years — and have helped trigger a shift in mood that has driven a light market pullback around AI.

The papers fundamentally argue that generative AI is overhyped as a driver of market returns because i) the technology is only being used for efficiency gains (which will be competed away as sources of profit) and not driving new revenue streams to justify the investments, and ii) the costs and quality of AI technology aren’t improving fast enough to drive widespread adoption for the use cases imagined.

There were several points they made that I found salient:

• Generative AI has more use cases around automation and cost reduction than use cases that generate revenue. The superpower for this technology is to give humans leverage in automating existing tasks, and every major use case that has gotten traction to date has been about efficiency. As a result, AI users should see the promise of AI as lower costs rather than higher revenue.
• The profit improvements from many cost-reduction initiatives will be competed away over time. Over a few year period, AI laggards will likely catch up with AI early-adopters, and efficiency gains will no longer result in greater profits (the benefits will instead be largely passed along to consumers in most industries). For example, if an airline figures out a great AI call center that reduces its customer service fees, all other airlines will eventually figure this out, and it will ultimately work its way into lower ticket prices rather than higher profitability for airlines.
• Much of today’s AI revenue isn’t yet sustainable: it is either “innovation showcase” (companies and individuals trying to show off innovation rather than real applications) or AI companies buying from AI companies. A variety of issues hold back widespread adoption: accuracy/hallucination issues, price, and the fundamentals of any technology change (internal realignment, alignment of vendor ecosystem, security and regulatory reviews, procurement processes, etc.). I think these constraints are real, and is the most compelling argument against AI hype — as it is fundamentally a timing question of when real use cases will emerge.
• Many of the AI businesses emerging don’t have great defensibility (i.e. “Automation for X” can often be built by many startups simultaneously), and as a result won’t be able to build incredibly valuable new companies that drive new revenue streams.
• AGI is overhyped. Despite the excitement of enthusiasts, there is little evidence that we’re heading towards a world where AI accelerates this much.
• There are going to be micro-bubbles and frenzies along the way. Even if AI as a whole does make a major impact, any of the specific companies hyped will quite possibly fail, and the market will probably go through waves.

I believe all of the points above are correct. In the past, I’ve said that my base case is that there will be a 2–5x surge in American productivity in the next decade as a result of AI, and as a result we should act like this is 1995 — and that we’re in the early stages of a sustained shift in markets. So how do those views align?

What I think was missing in the logic of these papers:

• Unlike many speculative bubbles (eg. crypto, in my view) — the killer use cases are real and near-term. Lots of real use cases are already viable — but take time to become enterprise ready. Many aspects of call center automation, customer service automation, programming, legal document mining, and medical billing automation have clearly-viable automation use cases that will have widespread adoption in the next year or two.
• Moreover, a huge number of automation use cases are on the precipice of viability, and a small change in the accuracy or cost can push them into feasibility. Meanwhile, the underlying technology is continuing to improve rapidly, and a Moore’s Law type progress seems like a reasonable trajectory. So the boom in real use cases a couple years out seems likely.
• While many AI-related efficiency gains will be competed away, they will make existing moats and network effects even more valuable — and distribution channels will become more valuable. For example, if you automated away 5% of the expenses of every insurance company in America by improving back-office processes, while much of that gain would be passed along to consumers, I expect that insurers would capture some of the improved margins based on their existing moats.
• A new set of companies that provide core infrastructure and complimentary components for AI will pick up revenue. This is the tailwind that NVIDIA and big tech is riding, and it also applies to a variety of startups.
• Consumers will accrue most benefits, and that will benefit a wide variety of companies in the market. If consumers become more wealthy, they will buy more of a wide variety of products. So even areas like real estate or retail could end up major beneficiaries of AI.

To recap — in a world where everyone improves productivity by 2–5x, I would expect markets would do well in aggregate, even if AI doesn’t create new revenue streams for incumbents but just helps them boost efficiency. The beneficiaries would be:

1. Consumers will reap most of the gains — and in turn spend more on a variety of non-AI products
2. Incumbents with good moats will keep some of the gains — and the companies that adopt AI ahead-of-market may be able to use the transition period to deepen their moats
3. New infrastructural components for AI use cases
4. Investors that systematically back #2 and #3. In particular, VC (and public markets for big tech) are making big bets or #3, and some PE firms are looking at #2 as a generational “digital transformation” opportunity.

So, ultimately, even though AI’s economic impact may just be about efficiency gains, and even though the majority of the returns might not go to the industries where AI has the biggest impact (as they get competed away), the gains to the economy, consumers, investors, and builders will be very real — and if anything, there is still an under-appreciation in markets of the long-term impact of this.