Oracle and Nvidia – Canaries?

The cure for low prices is low prices.

• Oracle’s precarious financial arrangements and Nvidia’s admission that it is not committed to investing in OpenAI go hand in hand with the financial markets signalling that the foundation of the AI boom is not secure.
• The problem is simple in that it costs $50bn to build 1GW of capacity, which is expected to last for 5 years before it needs to be replaced.
• From this, one can generate $10bn a year in revenues, and it does not take a genius to figure out that this model leaves no room for profit, paying back debt or equity investors.
• Hence, it is clear that the business model is broken and that something needs to change.
• The promise has always been that with each successive generation of chips, the return that one can earn from one’s investments would improve, but this has not come to pass.
• Take OpenAI, for example, which has been able to generate around $10bn per GW of capacity that it operates over the last 3 years, which is unchanged despite significant technology advances (see here).
• I suspect that this is not because Nvidia or AMD have failed to deliver on their promises of more compute at a lower price, but because the price of compute tokens has been falling more quickly.
• This is not unexpected, and in fact is a requisite for AI to really take off across the economy, but unless money is made along the way, the whole proposition will grind to a halt while the market figures out the economic way forward.
• This is most likely to take the form of the endless supply of money drying up, meaning that the uneconomic companies go out of business, leading to a large slowdown of growth in the supply of compute.
• This, in turn, will lead to higher prices and allow those who are left to make money from selling compute tokens, and it is from this more stable foundation that the AI industry can develop.
• The situation in which we find ourselves today remains unstable as the business model of compute upon which the AI boom is being built is not viable, meaning that something needs to change.
• This is why the credit default swap (the cost of insurance against going bankrupt) on Oracle remains 300% above its historical average and why all of the “circular” deals that have been struck are little more than statements of intent.
• This means that anyone can change their minds about how much compute capacity they build, when they build it and how much they invest at any time with no penalties.
• This is what Jensen is signalling when he said over the weekend that the $100bn investment in OpenAI was “never a commitment”.
• It is also why the bond market is demanding a higher interest rate from those who wish to borrow to build data centres, as it perceives that the risk of not getting its money back is rising.
• This is something with which I agree, and something needs to change in the business model of compute to prevent economics from getting in the way of the rollout of the AI industry.
• I think that the change will be a slower rollout of compute, leading to slower growth in the supply of tokens, meaning that the price that can be charged for them rises, bringing everyone into profit.
• However, while everyone continues to be willing to throw money into AI with hardly a consideration of whether they are backing an economic proposition, supply will continue to grow, and the price for tokens will continue to fall.
• Eventually, the money will dry up as losing money on trillions of tokens gets pretty annoying after a while, meaning that the cure for low prices of compute is the low price of compute, meaning that the lower it goes, the more likely it is to trigger a reset.
• At what level the money gets fed up, I don’t know, but keeping an eye on the bond market will be a reasonable indicator.
• At the moment, it is flashing yellow, indicating that there is a problem, but those who are investing the big money in uneconomic data centres have yet to pay attention.
• I continue not to want to be anywhere near this when the money finally dries up.