OpenAI – Token Economics

Revenue per GW has to rise.

• OpenAI is exiting the video generation business in a sign that the economics of generating tokens is rapidly becoming a problem, as those pouring billions in will eventually ask for a return on the investments they have made.  
• OpenAI is exiting its Sora video generation app and with it, the $1bn deal that it signed with Disney, which would have allowed Disney IP to be used in AI-generated videos.
• OpenAI’s commentary implies that it has struggled to get enough traction, and as the deal for Disney to invest $1bn never closed, it was an easy decision to take.
• I suspect that the real reason for Sora’s closure is economics, as video generation consumes a large amount of tokens, and the price that can be charged for Sora means that $-per-token is lower than for other services.
• With compute remaining constrained and an IPO coming up, OpenAI needs to prioritise where it allocates its tokens and maximising $-per-token will be crucial in ensuring that the IPO succeeds.
• This is also a sign that OpenAI has recognised that the business model of compute is challenged and that something needs to change for the industry to show a positive return on investment.
• As it stands today, generative AI services generate around $10bn per GW, cost around $40bn to build and last for 5 years before the equipment needs to be replaced.
• With electricity costing $1.2bn per GW per year (US average), it is not hard to see how the business model is challenged financially, and RFM Research has calculated that the 5-year return is 1%.
• This means that any rational investor would prefer to invest in US treasuries, which are “risk-free” and return around 4%, compared to a data centre, which is far from risk-free and is returning around 1% over 5 years.
• This is the problem that OpenAI has to overcome, as it has been stuck at $10bn per GW of capacity since 2023, and moving to Blackwell from Hopper, which promised economic improvement, does not seem to have helped.
• This is why Nvidia’s promise that Vera Rubin should allow revenue per GW to increase 5x is so important because if it can deliver on that promise, it will transform AI data centre economics.
• If the past is prologue, this would lead one to be extremely cautious about putting faith in this, but something needs to change, or investors will eventually get wise to the problem and stop funding the rollout.
• There are already signs of this in the bond market, where the credit default swaps on Oracle, CoreWeave, and now SoftBank are all elevated, and the share price of SoftBank has declined significantly as a result of its intent to raise debt to invest in OpenAI.
• Hence, OpenAI needs to do everything it can to maximise revenue from all of the tokens that it produces, and it is clear that video generation is not financially viable.
• It also highlights the advantage that Google (and Meta) have over OpenAI, which is that they have very large cash flow from existing business that allows them to invest with less constraint.
• This is problematic for OpenAI as the consumer AI ecosystem is a must-win if it is ever to justify the $1tn+ valuation that will be asked for at IPO.
• Without a convincing story on this front, the market may refuse to pay the asking price, which would trigger a significant loss in confidence and potentially an industry-wide correction.
• This is why the rollout of Vera Rubin is so important, and I will be looking for improvements in revenue per GW as the indicator of whether things are getting better or headed for a correction.