'This shit's so expensive': a note on generative models and software margins
So far, AI hasn’t been profitable for Big Tech | Ars Technica
The cost to Microsoft exceeds $20 a month per user on average, according to a person familiar with the matter. In some cases, individual power users have cost the company as much as $80 a month.
Software usually has extraordinarily high margins. Tech’s nonsense management tactics are contingent on being able to float shit away on a tide of money. You only get away with bullshit software dev, shipping abysmal apps, and constant security lapses if you are making an obscene amount of money.
This is how Microsoft and Adobe survive. Monopoly rents on margins that are only possible because your core good is non-rival and non-exclusive. The fundamental problem with generative models is that they are 10x too expensive to work with the industry’s default business models and structure.
Either these companies who are going all-in on “AI” need to fundamentally change everything about how they work – laying off a bunch of people won’t make ML compute 10x cheaper so they’d need to change the org to survive on razor-thin margins – or they need to discover some undefined magical way of lowering compute costs 10x. So far they’re opting for magic.
With the caveat that I’m not an accountant, but the last time I checked many tech companies like Microsoft categorise software dev costs on cutting edge products like ML as capital investment. It’s R&D on building an entirely new capability. So, those cost numbers are likely almost entirely compute costs, the cost of the work itself has probably been shifted to another box somewhere.
Productivity gains won’t make a meaningful difference here. The only path forward for them is either an exponential drop in cost (Moore’s law on steroids) or raising prices.
You only get away with raising prices if the product is doing something of substantial genuine business value, which most generative models do not. Even the best-case scenarios coming from studies outright funded by the vendors themselves (which means they are bullshit, you should never take vendor-sponsored research seriously as anything other than marketing) top out at 20-30% gains in productivity. The reality is likely to be a fraction of that simply because the problems they are trying to automate (writing memos and emails) aren’t that valuable of a problem and because the software user interfaces that mediate between the end-user and the model are abysmally bad: chatbots and insecure prose-oriented text boxes. The $10-20 USD a month they’re currently charging is probably too high for broad non-bubble use.
The problem with betting on Moore’s law to solve the cost problem is, as I pointed out, current models and the software that wraps them, are just not delivering that much business value and they are generally extraordinarily slo. Process improvement gains will be soaked up for the foreseeable future just from fixing the product: lowering latency, improving training speeds to make models more dynamic, responsive, and “fresh”, specialised work to make the models more useful and less abysmally unsafe at valuable business problems.
These companies are also all-in on chasing AGI, which is a never-ending sci-fi boondoggle that will soak up endless amounts of money and processing power and leave absolutely nothing behind except billions over-invested in a product category that has no hope of ever paying any of it back.
I don’t see how any of it is going to meaningfully lower costs over the next 5 years.
That does not bode well for a safe resolution of the current funding bubble.