Artificial Intelligence – Crazy Economics

The business model of compute needs an urgent fix.

• Unless one is a huge technology company with cash to burn, the economics of compute is an existential issue that, if not addressed, will force a significant part of the AI industry into bankruptcy.  
• OpenAI has inadvertently let slip (and echoed by CoreWeave’s financial statements) that the realisable revenue per GW of compute is not growing despite advances in silicon that greatly improved the total amount of compute that a GW can produce.
• Furthermore, the costs to build and deploy 1 GW continue to rise, meaning that unless something changes, the business model of compute is becoming less and less viable as the industry moves from one generation to the next.
• OpenAI published a blog post (see here) aimed at proving that it can grow revenues in line with compute capacity, which is supposed to justify its outlandish plans to build 30GW+ of compute capacity over the next 5 years or so.
• Here, it states that in 2023 it had 0.2GW of capacity from which it generated $2bn in revenue ($10bn/GW), in 2024 it had 0.6GW from which it generated $6bn ($10bn/GW) and in 2025 it had 1.9GW from which it generated $20bn ($10.5bn/GW).
• My back-of-the-envelope calculations for CoreWeave indicate a very similar profile.  
• This is a precarious state of affairs because the cost to build 1GW of capacity is growing and now sits at roughly $35bn-$50bn/GW from around $22bn/GW in 2023, largely as a result of rising costs of silicon, networking and power/cooling technologies.
• Hence, the costs of building compute are rising rapidly, while the revenue remains static, indicating that the economics of compute are worsening, which is precisely what the dark corners of the bond market are pricing in.
• For many years, the bond market has viewed Oracle as unlikely to default on its debts, as the cost to insure oneself against a default in the next 5 years (credit default swap) by Oracle has typically been around 0.05% of the sum insured, which is in line with IBM, which is also around 0.05%.
• However, this all changed when Oracle signed a $300bn deal to supply OpenAI with compute as once the euphoria wore off, the shares fell by 40%, and the credit default swap widened to 0.14%, which is roughly where it remains today.
• This is an increase of 180% or 2.8x, indicating that the bond market is uncomfortable with the business model of compute, which is very similar to what the bond market was signalling in the months before Lehman Brothers went bankrupt, triggering the 2008-2009 financial crisis.
• OpenAI’s own data confirms that the bond market is on to something, as the cost to build a GW or data centre is rising rapidly while the revenue that one can generate from it is static.
• This is an indicator that the business case for offering compute for sale is worsening and doing so at an increasingly rapid pace.
• Consequently, either the cost to build a GW of compute needs to fall (as opposed to rise), or the revenue that a GW of compute can generate needs to rise.
• Failure to address this issue means that OpenAI, Anthropic, Mistral et al are not going concerns, which the market may realise once they file for bumper IPOs and these financial issues are presented in black and white.
• When WeWork filed for its IPO, it did not take very much to realise that its business model was broken, which is what caused the IPO to fail (see here).
• The bigger OpenAI becomes, the more money it loses, meaning that it must find a credible way to monetise the 700m+ users who do not pay it anything but cost it a fortune in free compute.
• Failure to do this could cause the IPO to be pulled, which in turn would cause a loss in confidence in the business model for AI.
• As all of these companies are now connected to one another, this could easily set off a domino effect where the valuations of all concerned take a big hit, and some go out of business or are acquired.
• This is why the business model of compute needs to be fixed as just throwing more and more money at the problem does not seem to be having the desired effect.