The AI economy is hard to read but opportunities may have arisen for selective investors
The AI rout on Thursday was caused specifically by the blue wave in the US midterm elections and the implication of more hawkish policy towards the data centre buildout and AI in general. Many Democrats had campaigned on stricter regulations for the construction of data centres, and with Democrats winning in important data centre states such as Texas, Georgia and Ohio, investors fear that construction will slow sharply or even be banned.
The market has taken a typical ‘shoot first, ask questions later’ view, aggressively selling off those companies providing equipment for data centres or the chips that go into them - wiping approximately US$2.3trn in market value on Thursday alone. The market went from rotational to panic mode, but investors appear to be overlooking the strength of this earnings cycle. Indeed, this week's sell-off coincided with earnings growth this quarter matching last quarter's incredible 34%; at the margins, with analysts reducing their earnings growth expectations for next year (to 20% rather than 21%), but this still falls far short of justifying stocks' wholesale decline.
The reality for AI regulation will likely be less dramatic than feared. Campaign promises are often diluted once elections are over, and job growth remains important for politicians. State regulation will tighten in some areas, especially regarding power and water consumption, but the end result is almost certain to be a slowdown in data centre construction rather than an abrupt halt.
Our core criticism of the AI trade does not stem from the potential political blockade of data centre construction - on the contrary, this could even play into the hands of the hyperscalers who have the balance sheets and permitting relationships to navigate a tighter regulatory environment, while more speculative builds are less likely to get permissioned. This could even be a healthy outcome for the market, removing some of the froth from the AI infrastructure buildout.
Our deeper concern comes from the interconnected nature of the AI economy: the largest companies in the world are investing effectively the GDPs of entire countries to build out compute capacity on the basis that two unprofitable start-ups - each with potential valuations of well over a trillion US dollars - will be able to fulfil their vast commitments to the hyperscalers for compute.
Anthropic's quarterly revenue approached US$17bn in the third quarter, a roughly 45 times increase over six quarters, yet the company still lost US$23.3bn. Currently, some sources suggest that up to 75% of compute commitments come from OpenAI and Anthropic, and therefore the bulk of this AI cycle is based on the growth of these two companies. Anthropic has signed US$600-$800bn in medium-term computing capacity contracts (with varying contractual obligation/pricing). That makes Anthropic’s expected upcoming Initial Public Offering particularly significant. We continue to expect it to proceed, but further clarity this week could provide an important early test of public-market appetite for the valuations underpinning the AI investment cycle.
Despite Anthropic committing vast sums to compute relative to its current revenues, this does not necessarily mean that the mathematics of the AI economy cannot work. Anthropic’s revenues have grown by more than 1,100% in 12 months for instance. For the Large Language Models (LLM), (referring to the AI developers), projecting profitability is a balancing act between the level of diffusion of AI through the economy, the amount the LLMs can charge for their services (predominantly measured by 'token prices' - a single token being about three quarters of a word of output from an LLM) and the cost to the LLM for compute. This logic does not consider the improvement in AI services with agentic AI yet to be widely adopted.
On this basis, demand for AI is accelerating on a massive scale; according to tokensperday.com at the end of 2024 9.4 trillion tokens per day was the estimated global daily use, currently the estimated daily use is 461 trillion (these estimates are highly uncertain).
Meanwhile (and to make matters more complicated) token prices have collapsed by circa 60% from their summer peak. This divergence reflects intense competition among AI providers absorbing margin to win volume, but it does not necessarily make the AI economy less feasible (and this is well supported). The narrative is that as token prices fall the adoption rate increases and revenue counterintuitively increases. This is known as Jevons' Paradox. The original formulation comes from the Victorian economist William Stanley Jevons, who observed in 1865 that improvements in the efficiency of coal-burning steam engines - which reduced the cost of coal consumption per unit of work - led not to less coal being burned overall, but dramatically more, as the lower cost made coal economical for a far wider range of uses. The bear counterargument is that Jevons' Paradox has limits: when deflation runs at roughly 47% per quarter, the volume response required to sustain revenue growth becomes implausibly large - and as Amazon's CEO Andy Jassy has noted, AI inference remains approximately ten times too expensive for truly pervasive business adoption even after the price declines of the past year.
Finally, we come to compute prices; again this is incredibly difficult to measure and a moving feast. Compute prices have travelled in the opposite direction of token prices, rising circa 90% in the last 12 months. The market for compute remains exceptionally tight. What makes this even more complex is that compute and token prices are not directly comparable: compute is sold as capacity per hour, while tokens are essentially the output from an LLM.
All of this is to say that the AI economy is incredibly complicated and hard to read, and even though we can admit to offering little concrete certainty at this highly uncertain moment we believe that opportunities have arisen from this volatility. The build out is real, the investment is real and the cash flow is real, making the valuations of some of the companies at the receiving end of those cash flows relatively compelling. The Philadelphia Semiconductor Index now trades at just 13.7x forward earnings - cheaper than the US equity index and at levels not seen since January 2019, before the AI boom began.