Without Subsidies, Is AI Unaffordable?

Authored by Charles Hugh Smith via OfTwoMinds blog,

Let’s pull all this into an undeniable conclusion: AI is based on massively subsidizing users’ costs.

What’s already abundantly clear but verboten to say as it would pop the bubble of AI valuations and triumphalism is that AI is unaffordable once the direct and indirect subsidies are withdrawn. Nothing that consumes this much electricity and requires such an immense scale of costly processing and memory capacity can be low-cost, never mind free.

The major AI platforms and vendors are subsidizing corporate and individual users in the hopes that they can achieve AI sector dominance –and the pricing power that comes with it–via the network effect, the dominance generated by having the majority of users bound by habit or dependence to your platform or tools.

This battle for network effect dominance is playing out in full view:

AI Giants Are Handing Out Tons of Free Computing Power to Grab Startup Share: (wsj.com) Pitched battle for business users comes as AI companies seek lasting streams of revenue.

Hans Ibarra, a founder building an AI-voice startup, has found himself on the receiving end of a big opportunity: Top artificial-intelligence companies such as OpenAI, Anthropic and others desperate to win his business are ramping up discounts.

Across Silicon Valley, startup founders like Ibarra are enjoying a wave of computing credits and fielding competing offers from AI-model makers racing to land new enterprise customers. Cursor, the AI-coding company bought by Elon Musk’s SpaceX, offered a 75% discount through July 5.

“If I’m choosing between a really cheap Chinese model that I actually have to pay for, and a very expensive Anthropic model that I don’t have to pay for, I’m going to pick the Anthropic model,” Acker said. “I’m always going to pick the one for which I have free credits.”

Meanwhile, back in the real world of costs, AI Costs More Than The People It Replaced (forbes.com)(via Tom D.)

It turns out that experienced human workers doing the work right in the first place is cheaper than having AI run a probability distribution process that needs vetting and corrections. And remember, AI isn’t actually “intelligent,” it’s just a probability distribution using natural language.

As management guru Peter Drucker observed, enterprises don’t have profits, they have costs. Purveyors of AI platforms and tools have costs, and so do their customers. Those costs are currently being funded by investors, who are in effect subsidizing the AI companies’ “free” giveaways of horrendously costly “tokens” in a manic, desperate attempt to grab the brass ring of network effect dominance before their cash runs out.

This raises a question: Is this any way to run a railroad? In other words, is this actually a viable business model, burning billions of dollars in cash to lock in network effect dominance in a field that is rapidly obsoleting every iteration of an innately limited mode of computation? Is claiming that a probability distribution is “intelligent” in the same way humans are intelligent a viable business model when there is ample evidence this simply isn’t true?

AI and human intelligence are drastically different–here’s how (scientificamerican.com)

What happens when enterprises have to pay the unsubsidized costs of AI is they immediately curtail their AI spending because the customer-facing / financial benefits of AI are at best elusive and often negative. Peter Drucker was onto something that is currently being lost in the PR-propaganda push of those trying to cash in on the AI euphoria: enterprises don’t have profits, they have costs, and the real-world costs of AI are extraordinarily high while the payoffs are ambiguous.

There are many other hidden subsidies within the AI machinery. There are corporate tax write-off subsidies, energy subsidies, tax credit subsidies for building data centers, and so on. If these were stripped out, what would the real unsubsidized costs of AI be? No one knows, but they would be higher than what’s presented as the cost now.

Then there’s the if it’s legal, it’s moral, and what’s legal is for sale subsidy: AI is built on the systemic theft of copyrighted content. Last month alone, AI scrapers gorged on 246,000 pages from my Of Two Minds server, and hundreds of thousands of pages of my copyrighted works on my mirror site and other sites posting my work.

This is legal, but is it moral? Nobody asks such questions because the important thing is to avoid saddling AI users with the real costs. So if all those content creators get nothing–in effect, subsidizing both AI companies and the users of their AI platforms and tools–well, so what, because if it’s legal, it’s moral, and what’s legal is for sale.

Well that’s just peachy, but let’s do a thought experiment where every creator of copyrighted work got paid for supplying AI with its database, and every user of AI had to pay us content creators. How about a penny a page / image / sound clip? so 246,000 pages per month (again, only a fraction of the total volume of my work that was scraped by AI companies for their “free” use in a single month) would be $2,460 a month paid to me by AI users benefiting directly from my copyrighted work. Wouldn’t that be fair, i.e. moral?

Recall that US copyright law is explicit: all creative content is copyrighted upon completion, period.

How many current users of AI are willing to pay the full unsubsidized costs for their use of AI? We can safely say far fewer than are using the tools for “free” due to subsidies both direct and indirect.

Let’s pull all this into an undeniable conclusion: AI is based on massively subsidizing users’ costs. Once those subsidies end, what’s left are costs, not profits. Play that any way you like, but massive subsidies are not sustainable, though they generate a temporary illusion of viability that can be exploited by those selling a fantasy of future profitability to credulous investors and enterprises.

Left unsaid is a lot of money is being gambled on the illusion that subsidies are sustainable. They’re not. Enterprises don’t have profits, they have costs.