Why Anthropic and OpenAI must go public now. The paradox, that lead to 75% discounts.

Short-opinion article that justifies why the big labs should go public now, and not wait to increase revenue. I believe any increase in revenue will lower the valuation, by lowering the multiple.

This time is going to be a short-opinion article. This opinion justifies any long on Google under the new revenue line that increases both revenue number and multiple; and how I view the IPO-event for the big labs. They must go now, or they will be squeezed on IPO.

Training becomes more and more expensive as these big labs grow. To keep increasing their training, they have to convince private investors to give them more capital to train their models. To create that justification possible, they have to generate revenue and prove their ability to monetize their product.

These big labs have one unique way (at least for now) to originate revenue: they sell inference.

The big paradox is that to sell inference, they have to spend their scarce input: compute. As they grow the inference share of compute, which leads to higher revenue, which justifies the investment. They are reducing training compute. By reducing training compute gives the market the sense that their ability to create smarter models is dropping, where extra compute to training isn't interesting to support. As if new models are no longer worth the investment. Investment interest drops.

Looking ahead, we could say that, eventually, the marginal benefit of training new models eventually will drop to zero, it could take 1-year or 10-years, but eventually will happen. Where the marginal incentive that comes from a smarter model is not positive, or more training do not end up with smarter models. A point where using compute to train, rather than use it on inference, will not be justifiable to any investor capital.

But in this period, without labs owning any application built on top, or any other product vertically or horizontally integrated, these labs end up as cloud providers.

This means the multiples sign for Anthropic and OpenAI could be pushed from the current 30x P/S multiples, to the Alphabet/Microsoft multiple of 10x P/S, and eventually to below 10x P/S as they reach (only) the cloud provider status, with inference share of their own compute above 80/90%.

I believe the inference share of the big labs, will determine the IPO multiples sign by the public/private investors. The paradox continues, as they need to increase the inference share to increase revenue number, that increases valuation. Increase inference share also lowers their own valuation multiple, that lowers their valuation. A perfect balance of this paradox must be achieve, as the ability to rapidly increase compute is super limited.

IPO'ing with inference at 50% or lower means the market can still expect smarter models in the future from the spend of their own compute. That justifies 30x P/S multiples, above any other big tech company. At $100/$150B in annual revenue, that puts an Anthropic IPO at $3 to $4 trillion. If they IPO with inference above the 75% benchmark, the multiple could compress to 10x, or 15x with a premium, meaning the same annual revenue could represent a $1 to $2 trillion IPO. Up to 75% valuation discount, it's a big thing.

Thanks,

Joao