AI Doesn't Have ROI

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large. My Hater's Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle. Over the last three weeks, I’ve published an exhaustive three-part guide to how the AI bubble might collapse, the events that might trigger it, and the consequences. Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week. Something changed in the last week.Shortly after Uber COO Andrew Macdonald said that it was “getting harder to justify” spending money on AI as it was “very hard to draw a line” from that spend to useful consumer features (after its CTO said Uber burned its entire annual token budget in four months), Axios’ Madison Mills reported that one company had accidentally spent $500 million in the space of a month on Anthropic’s models after failing to set spend limits. A few days later, Mills would report that other companies were now looking for ways to reduce their AI spend.That’s because, as I’ve said before, nobody can actually measure the ROI of AI, or even create a standard measurement of the cost of a task thanks to the inevitable hallucination-prone nature of LLMs and the ever-growing list of different harnesses and “agentic” (sigh) interfaces. Every different prompt and project and interaction can go wrong in a way that is hard to predict or plan for other than having an eternal vigilance that the supposed “intelligence” doesn’t do something catastrophically stupid, because LLMs have no thoughts, consciousness or ability to learn outside of pre and post-training. If you can’t measure how good something is, how much it might cost, or what your return on investment might be, it’s fair to ask why you’re even paying for it in the first place.

People are (reasonably!) harping on about the ROI problem, but I think the “can’t really measure the cost” part is an even bigger problem. Yesterday, Microsoft’s GitHub Copilot moved all customers to token-based billing from a premium request model (as I reported a week before everyone) as users had been allowed to burn thousands of dollars of tokens on a $39-a-month subscription. Customers are irate. One burned through 50% of their monthly credits in a single prompt, another burned 60% in the space of a few hours, another 31% in a single prompt, another estimated that they’d burn their monthly credits in the space of a single five hour session, another burned nearly half of their credits in eight prompts, another around 14% of their credits in two prompts, and another lamented that GitHub Copilot had gone from their favorite subscription to their most-stressful overnight after burning 33% of their monthly balance in a few hours.And, to be clear, this is during a promotional period where you get $11 or $21 in free monthly credits:These users — much like the users of effectively every subsidized AI subscription — never really knew how much anything they did cost, because Microsoft intentionally hid the actual cost of prompts and allowed users to spend obscene amounts as a way of boosting growth for GitHub Copilot. This problem is industry-wide.Every single user of every single AI subscription service is having their tokens subsidized and the actual cost of AI obfuscated. As a result, every frothy, fluffy hype-piece about Claude Code or AI in general is a kalopsia — the belief that something is more beautiful than it really is. Educational Sidebar! While many of you may know this, for those just joining me, let me break down how the average AI subscription works. You pay a monthly subscription to, say, Anthropic or OpenAI’s services, and get to use these services as much as you’d like subject to both daily and weekly “rate limits.” None of these companies ever really explain what that rate limit might be, giving users instead a vague percentage gauge and leaving them to work it out on their own.

When you use an AI model, you feed information into it via input tokens (a token is about ¾ of a word) and receive outputs via output tokens, and companies bill on a per-million token basis. While models can “cache” information as a means of avoiding having to read or write it again, every single interaction costs money, regardless of its success or efficacy. This is why every AI startup is inherently unprofitable — they’re literally sending every penny of their venture capital money directly to Anthropic and OpenAI to power their unprofitable services. AI labs may be able to run their own infrastructure and save some costs, but we have no evidence that this makes anything “profitable.”For example, Anthropic lets you burn anywhere from $8 to $13.50 in tokens for every dollar of subscription cost, and while AI boosters will say that Anthropic is “profitable on inference,” nobody actually has proof outside of theoretical scenarios posed by CEO Dario Amodei.Think of it like this: if you’re using an AI subscription with rate limits but no actual costs, any mistakes a model makes — such as getting stuck in a loop or just doing the wrong thing — can be dismissed as the troubled nature of early-stage technology, because the “cost” was $20, $100, or $200 for the entire month. Anthropic, OpenAI and every other AI company deliberately obfuscated these costs because they knew that the second a user actually had to pay for the fuckups of an AI model they’d scream like they were being stung to death by bees.Despite Promises To The Contrary, AI Is Getting More ExpensiveThis issue bubbled to the surface in the last few months because Anthropic and OpenAI both quietly moved all of their enterprise customers to token-based billing in Q1 2026, and because these enterprise customers are run by Business Idiots with no connection to actual work, CEOs encouraged (or actively incentivized) their workers to use AI as much as possible, in some cases even making one’s AI use a KPI that could cost them their job.

These same workers were conditioned — through their use of AI subscription products that hide the true costs — to use them as if they cost nothing, all while being screamed at by useless middle managers to “make sure to adopt AI at scale,” all while never, ever having any awareness of what a particular unit of work cost.This was always a recipe for destruction. The overwhelming majority of AI users are completely divorced from and actively trained to ignore the true cost of AI tokens, which means they naturally use these services in a way that’s actively uneconomical. Every frothy hype-piece you’ve read has been written by somebody who has been conned into ignoring the true cost of AI, all in service of spreading a technology that’s unreliable, inconsistent and expensive at its core, and never, ever seems to get cheaper. Sidenote: Even with the “cost of intelligence” (the per-million token cost) of models coming down, models are using far, far more tokens for the same task, ultimately raising the cost of inference. Put another way, imagine if the cost of gas got cheaper but the distance between you and your destination kept getting longer.OpenAI, Anthropic and other AI companies have actively conspired to mislead the world about the true costs of AI, and it was working great right up until they decided to try charging what it actually cost. Less than a quarter into the shift to token-based billing, enterprises are freaking the fuck out, with Walmart setting token limits on its internal “Code Puppy” AI coding tool, with a spokesperson saying that it “wanted employees to apply AI in ways that create value” mere days after Amazon SVP Dave Treadwell told employees to “not use AI just for the sake of using AI.”The last few years of AI hype have been built on lies. Every company has conspired to make you think that AI is affordable and sustainable, that profitability was possible, that hallucinations were fixable, and that any problems you faced today were a result of being in “the early innings.” In reality, the AI industry has absorbed over a trillion dollars, effectively all tech talent, the majority of startup funding, the majority of media coverage, the art and work of millions of people, and been given chance after chance after chance to fix the obvious, glaring issues.

Every time a skeptic dared to stand out and say that none of this made sense, they were told that it was just like Uber (it’s not) or that Amazon Web Services cost a lot of money (it cost $52 billion over the course of 14 years and was cash-flow positive in nine), that “costs always come down,” and that everything would magically be alright as long as they were patient for an indeterminate amount of time.Four years and a trillion dollars in, AI is more expensive, its companies more cash-intensive, its products just as unreliable, and its boosters more desperate than ever to make you ignore reality as a means of empowering one of a few ultra-rich oafs. Products from OpenAI and Anthropic are built to ingratiate and coddle losers while creating work-shaped outputs that are good enough to impress braindead executives, imbeciles and middle management hall monitors that don’t do any real work, and the reason it’s worked this long is that both companies intentionally misled everybody about how much the real costs were.I must repeat myself: AI is more expensive today than it was three years ago, and it is not getting cheaper. Sam Altman’s comments about “intelligence too cheap to meter” were lies. NVIDIA’s Blackwell GPUs didn’t make it cheaper, and its Vera Rubin GPUs won’t either. Google’s TPUs won’t do it, Amazon’s Trainium or Inferentia chips won’t do it, Vera Rubin CPUs won’t do it, OpenAI’s chips won’t do it, and no, DeepSeek won’t do it either. People chose — and still choose — to believe that AI would get cheaper because they think things got cheaper over time in the past, which is sort of true but not remotely similar in any way, because the cost of running and training AI models comes from using the hardware as well as its upfront cost. Large Language Models require expensive GPUs thanks to their reliance on power-intensive parallel processing, and larger, more-complex models in turn require more GPUs to both train and run inference with.And three generations in, NVIDIA GPUs don’t appear to be bringing the cost down at all, which heavily-suggests that the inherent business model of generative AI is broken.