Artificial Intelligence – Bubble Debate pt. III

The real problem is economics.

• The credit markets and recent sell-off point to a bubble, while demand points to the opposite, meaning that the deciding factor is likely to be the economics, which currently need to be improved.
• The debate about whether we are in an AI bubble continues to rage, which is bleeding into both the equity and debt markets and, as usual, there are arguments in both directions.
• For the bubble
  • First, Overbuild: which is based on the view that spending trillions of dollars on AI capacity over the next 5 years will result in an excess of supply.
  • This, in turn, will lead to a collapse in price, meaning that forecasts are missed and the valuation proposition collapses.
  • This is the main reason why the argument for overbuild is being made and follows from OpenAI’s disclosure of its plan to build 30GW+ of capacity over the next 5 years.
  • Second, Valuations: which in many cases are outlandish.
  • OpenAI at $500bn and Anthropic at $183bn, despite burning billions of dollars every quarter, are the most well-known, but this is prevalent across the entire sector.
  • Both of these companies are being priced with the expectation that they will invent artificial super-intelligence and earn very large returns from a unique position in the market.
  • Unfortunately, almost all the evidence points to super-intelligence not being achieved using LLMs, meaning that these valuations are, without doubt, too high.
  • Third, Causality, which the machines continue to struggle with and is what prevents them from becoming truly intelligent.
  • The nature of an LLM is that it predicts tokens based on statistics, meaning that it has no real understanding of the task that it is being asked to perform.
  • This means that it is unable to truly reason and will fail some of the simplest tests that a 2-year-old would get right.
  • Consequently, super intelligence is not going to be reached anytime soon, meaning that the predictions of the cool-aid drinkers will remain unfulfilled.
  • Fourth, Power: where outside of China and the Middle East, there is nothing like enough.
  • OpenAI is proposing to consume 2% of the entire USA generation capacity on its own, which adds up to a lot more when you include everyone else.
  • Already, wholesale power prices are starting to move as AI datacentres come online, and power generation capacity is not being increased anything like quickly enough.
  • There is much talk of small modular nuclear reactors, which are likely to be important in the long term, but there is only one design that has a licence to operate, with everyone else buried under a pile of paperwork.
  • This is where China has a big advantage, as it was already building out a lot of power generation and is on track to add 100GW of nuclear power by 2050.
  • A lack of power means that the capacity gets delayed, targets are missed, and disappointment and disillusionment set in.
• Against the bubble
  • First, Demand: which continues to be off the chart.
  • Both Microsoft and Google stated that they were unable to keep up with demand for their cloud services this quarter.
  • This goes hand in hand with OpenAI, which now has 800m weekly active users and constantly struggles to meet demand for its services with the capacity that it currently has.
  • This is a stark contrast to the internet bubble, where there was very little demand and where the hope was that if they built it, the demand would materialise.
  • However, when it comes to demand, one also needs to consider the effect of the price of AI services, which is covered below.
  • Second, Awareness: where everyone is aware that the AI boom may well be overblown, which in itself is a contradiction.
  • This is because when one is living in a bubble, most people are unaware of that fact until it is too late.
  • The very fact that this debate is being had is an indication that things are somewhat different this time around.
  • Third, Frame orders: which means that all of the forecasts, plans and investments are not 100% committed.
  • An example of this is Nvidia and AMD’s combined $160bn deal with OpenAI, where only a small fraction is guaranteed to go ahead.
  • These are the builds that are planned for 2026, with the rest contingent on demand and market conditions as and when the next phases become due.
  • Hence, if there is a softening of demand in 2026, then the 5-year rollout could easily become 10 or 15 years with no financial penalties on either side.
  • This would also give more time for the power situation to be sorted out.
• The net result is that the current forecasts are unlikely to be met, but that is unlikely to result in a disorderly collapse, and I think an orderly reallocation of capital is more likely.
• The biggest problem we have at the moment is economic in nature.
• The main reason why there is insatiable demand is that the vast majority of all AI requests are free to access.
• When a good or a service is free, then economics would predict that demand for it will be almost infinite.
• Most of the AI services are free because there is a mad scramble going on to get as many users as possible before anyone else, so that one has scale when it becomes time to earn a return.
• The other problem is that the prices being charged for compute also look too low to me.
• Take CoreWeave, which by the end of 2025 will have 820MW of capacity from which it will earn a run rate of $6.3bn in revenue.
• Round this up and one ends up with a price of $7.7bn per year per GW of capacity that costs $40bn – $50bn to build out and lasts for 7 years using the longest depreciation schedule.
• This means that one would earn $54bn in revenues over that time, just covering the rollout cost, which is clearly not an economic proposition.
• Hence, either the price per GW has to come down a lot or the price of compute has to rise to make this an economic proposition.
• In the mad scramble to build and win customers, neither of these seems likely, and so some form of correction or reallocation of capital is required to ensure that AI is an economically viable proposition.
• This is not necessarily a bubble popping and causing a panic, but it will be painful enough not to want to be anywhere near it.
• The defence is to be in the lower valuation end of the industry, which will correct the least or in the picks and shovels, which are already making money now.
• AI is going to happen, but I suspect it is going to take longer than expected to figure out how to make money, meaning that there is a bumpy road ahead.