99% of Model Revenue Comes from Code Generation
I made that number up. The real figure may be 95%. The point being, now that META will no longer be spending $500-plus million in one month tokenmaxxing, now that AWS has been burned, now that UBER is putting the breaks on token spend, what will happen to the valuation of Anthropic and OpenAI as exponential growth slows?
A slowdown in per capita token usage does not mean that AI does not provide an ROI on code automation. What it means is that companies can’t do stupid things like META’s tokenmaxxing effort, which was to award large variable comp packages to engineers simply for burning tokens without measuring how much of that slop was shippable.
Zuckerberg is a terrible operator. How many failures are investors going to give him a pass on? Zuckerberg killed FB with Ads, he killed Instagram with Ads. The Metaverse is dead, and Zuckerberg will never have a quality LLM.
If you put a gun to my head, I would say “Yes, Mark Zuckerberg is committing fraud as I do not believe all of his users are real (I would guess as much as 50% are fake, not the 5% META reported in 2018), and I believe that many of those bot users are owned by META and purchase Ads on META’s network using META’s cash.” I wonder if as META spends more capital building data centers, will it have less capital with which to buy fake Ads? In other words, how much dry powder does META have with which to buy fake Ads?
AWS needs a new CEO. AWS fired too many engineers too quickly. No way AI is picking up the slack. I would not touch AWS with a 10-foot pole. All of our work sits on Google Cloud for the time being. When Google starts firing engineers at scale, we will have to evaluate whether it makes sense to rewire Kilby, CEORater and our other products to Azure or on-prem.
Back to Anthropic and OpenAI. The exponential growth is nearing the end. Will the models be able to train themselves? Maybe. If that happens, we could see a step function up in improved performance, but we will still need far more context than a 1 million token context window (perhaps a 1 billion context window is what is required to get the models to where I am not screaming at them every third day).
Will agentic AI catch hold? Sure, simple tasks and workflows can be automated via the combination of AI plus software automations such as cron jobs, RPA, etc. But large scale automation will require that operational silos be broken down, that some real-world processes be migrated to blockchains which will facilitate automation.
If your sellside analyst starts touting Anthropic, OpenAI, etc., ask them if they have used the models to write code. Have they used Opus 4.8 and Codex 5.5 and Gemini’s inferior models? How often? How recently? If it has been two months since you last used the models, you don’t know what you are talking about. If it has been one month, you will not have used Opus 4.8, and therefore will not know the first thing about it. If you logged in, asked it to open your calendar, and logged out, you have not used the models. I could go on an on. There is so much froth and fraud and greed out there it is sickening.
Oh, before I forget, SpaceX is a joke.