Artificial Intelligence – Circularity Debate

Bubble hunters float a hot new topic.

• As the numbers become ever more incredulous, the deals being struck are increasingly looking circular, but analysis shows that they are more creative than circular, leading me to think that this is not a sign that the bubble is about to pop as the bubble hunters would have us believe.
• In fact, so many people are now worried about a bubble that a spectacular collapse is now pretty unlikely, paving the way for an orderly correction as outlandish expectations get reset to reality.
• A circular arrangement is where a vendor lends a customer the money to buy its product at very favourable rates to enable the customer to be able to afford the product.
• This has been used as something to perpetuate unrealistic demand and to prop up flagging growth, and was endemic in the Internet bubble of 1999 and 2000.
• Here, telecom equipment vendors issued bonds and lent operators the money to buy fibre and install it in the belief that internet traffic would materialise to light up the fibre.
• Unfortunately, it took 5 years or more for the traffic to materialise, and many telecom equipment vendors collapsed under the weight of the debt they had raised as revenues and profits fell.
• This looks pretty bad, but it is worth remembering that this practice has been going on in the automotive industry for decades and lending money to consumers to buy cars is often the most profitable part of the industry.
• Hence, I don’t see this practice on its own as a sign of a bubble, but it does tend to occur close to the top of a trend, fad or bubble.
• In the current AI frenzy, there are a number of deals that have made headlines and look circular from a distance and include:
  • First, AMD and OpenAI: where AMD has issued 160m warrants with a strike price of $0.01, which, if exercised, would give OpenAI a 10% stake in AMD in return for a cash consideration of $1.6m.
  • These warrants are only exercisable at certain share price thresholds and upon the deployment of each GW of AMD chips for which OpenAI will be paying in cash.
  • However, if the warrants have no lock-up in terms of selling the shares bought for $0.01 each, then OpenAI could pay for almost all of the 6GW by selling the shares it gains from the warrants.
  • This is not circular, as it is not AMD giving or lending the money to OpenAI, but it is the shareholders of AMD and the market purchasers of AMD shares that are paying for the chips.
  • This is an elegant solution, but it will quickly unravel if AMD’s share price does not significantly increase, meaning that OpenAI will have to raise the cash to buy something like $60bn in AMD chips.
  • However, there is no firm commitment to buy the chips, meaning that if the share price fails to rally, OpenAI can simply not fill the order.
  • Second, Nvidia and OpenAI: where Nvidia will end up with $100bn in equity in OpenAI in return for supplying $100bn worth of chips for both training and inference.
  • This is less circular than the AMD deal as it simply boils down to OpenAI paying for chips with shares in the company as opposed to cash.
  • It will look the same in Nvidia’s financial statements, except that cash flow will be negatively impacted by the outlay to purchase the shares from OpenAI, which will offset any cash income received for the chips.  
  • There are a number of other transactions, such as OpenAI’s deal with Oracle and Coreweave, that have raised questions, but these are more akin to money circulating around the industry, given the cross holdings and have raised fewer eyebrows.
  • The net result is that while these deals look circular at a very high level, they are more about addressing the fact that OpenAI needs to invest a vast amount of money in compute while it has negative cash flow.
• Hence, I would classify them as creative rather than circular, but it is worth noting that if demand dries up, then these deals will grind to a halt as there is no firm commitment to fill the orders in their entirety.
• Furthermore, the big difference between these deals and what happened in the Internet bubble is that back then, demand for fibre had not materialised.
• By contrast, demand for compute is currently outstripping supply and many data centres that have not yet broken ground are already pre-sold.
• Hence, while there is certainly an overbuild going on, it is nothing compared to the Internet bubble, and so when the correction comes, it is likely to be an orderly correction rather than a mad rush for the exits.
• This could typically include valuations going sideways for several years while revenues and profits catch up.
• Those that will be most severely impacted will be the LLM companies like OpenAI, Anthropic, Mistral and so on, whose products are commoditising and who need to generate huge profits to justify their current valuations.
• This is why OpenAI is rapidly pivoting to become a consumer digital ecosystem, which is what its developer day (see here) last week was all about.
• If it can gain traction here before the correction hits, then it may be able to survive on its own, but I suspect that the likes of Anthropic and Mistral will rapidly be acquired by the AI laggards, such as Amazon and Apple.