Last year I wrote one of my favourite pieces ever — The Hater’s Guide To The AI Bubble — and followed it up with The Hater’s Guide To The AI Bubble Volume 2 several months later. Sadly, I’ve realized “volume” is a terrible way to structure something like this, because each volume is more of an update, which is why today’s newsletter will move to a versioning system.

The AI bubble is a psyop, a melodrama, a financial crisis, and a mask-off moment for the Business Idiots that run the vast majority of our economy. It is the largest-scale exploitation of ignorance in history, gnawing at the intellectual weaknesses of society by presenting just enough information or just enough proof to substantiate a trillion-plus dollars of investment and manufactured consent for a technology that, based on how many discuss it, doesn’t actually exist.

And it’s revealed how many rich and powerful people are either (or both) credulous and woefully ignorant.

To be clear, LLMs are real and do some things, but they don’t do any of the things that Dario Amodei is talking about when he says that AI will wipe out 50% of white collar jobs.

We’re four years into this joyless slog and people are still talking about AI’s “potential” and what it “will” do and that we’re in the early innings of a technology that, for the most part, is still doing exactly what it was doing at the beginning with refinements that never come close to reaching the vacuous heights of boosters’ promises. 

Markets are moved by poorly-written fan fiction by outright scam artists and deceptive hedge fund gargoyles because those selling AI services have entirely disconnected the minds of the markets and the media from reality. This is because con artists like Amodei and Altman constantly discuss what AI will or might or theoretically could do rather than what it actually does, because if they had to do that they’d have to say it constantly loses money and doesn’t have a measurable return on investment.

As I said on Bloomberg this week, the markets and the media have conflated capital expenditures for data centers with a thriving AI industry. In reality, 89%+ of all AI revenues and 90%+ of all compute demand comes from two companies — OpenAI and Anthropic — largely based on money-losing subsidized AI subscriptions and unrestrained token burn at organizations run by imbeciles that will go away now that executives are having trouble justifying it because there’s no ROI, in part because AI is too inconsistent and unreliable, and in part because you can’t really measure how much a task will cost. 

Now enterprises are already capping their AI spend, with many more to follow after multiple companies blew through their annual token budgets in a few months. The sheer volume of the “AI ROI” conversation is remarkable considering that Anthropic and OpenAI only moved enterprises to token-based billing — paying the actual costs of AI — in Q1 of this year. 

Remember: the total, actual revenue of the entire AI industry — including OpenAI, Google, Microsoft, Amazon, and Anthropic — has barely reached $100 billion in 2026. That includes every ounce of compute spend, every penny of the $500 million that a single customer accidentally spent on Anthropic’s API, and every cent of NVIDIA’s backstop deal with CoreWeave. More importantly, absolutely nobody is making a profit outside of those selling the bits that go inside a data center. 

Both OpenAI and Anthropic lose billions of dollars a year, with no end in sight, though Anthropic did a great job swindling the media by having a single “profitable” quarter thanks to Elon Musk discounting two months of compute. Anthropic has already filed to go public, with OpenAI  allegedly not far behind. Neither of these companies are fit for public investors.

Their products are inconsistent, unreliable and only ever seem to get “better” in a kind of wobbly way that can only be measured by increasingly-less-useful benchmarks that they specifically train to beat. Despite many people (and some companies like Spotify) claiming that AI is writing “most” code, nobody can seem to explain what that means. It isn’t saving money, it isn’t saving time, it isn’t making companies ship better or more-functional products, and the only tangible examples of its effects are that it broke AWS several times and deleted a company’s database.

It’s unclear where AI exists outside of coding and the various places companies have shoved it. 

I’ve spent years trying to catalogue other, non-coding use cases, and most of what I’ve found are vague descriptions of companies like Goldman Sachs maybe launching agents “soon” at some point to do something maybe and this weird story with Novo Nordisk claiming that it was “integrating ChatGPT’s models to analyze complex data sets” despite them claiming to have done this for years.

That’s because generative AI is, no matter how many hats or harnesses or deterministic processes you add, limited by its mathematically-certain hallucinations. These models are probabilistic, guessing at what the ideal output may be, which means that every bit of information they produce is suspicious and every decision they make is brainless, thoughtless and arbitrary. They do not “know” things, they do not have “thoughts,” and no amount of API connections will fix that problem.

As a result, nobody has really got a clear answer as to what everybody is doing with AI. Code? Image generation? Using it as a shitty search engine? Using it as a companion? You can’t really rely on it to do anything. When a model hallucinates an incorrect answer to something you know is true it’s a problem that can be fixed — when it hallucinates an incorrect decision with your codebase, that’s fucked everything up to a near-permanent end.

This is the ultimate problem with AI. You can try and dress it up with billions of investment and supposed ways to mitigate hallucinations, but it still makes — and will continue to make — mistakes that it has no idea are mistakes. 

Well, okay, the other problem is that generative AI just isn’t built to do most jobs. It can generate stuff and summarize stuff at varying degrees of complexity, but the more complex the generation, the more likely it is to hallucinate. The only way to reduce hallucinations is pre-training (shoving stuff into the model at the beginning) and post-training (training it on what “good” looks like), and neither of these actually solve the problem. It is clumsy, inaccurate, unreliable, expensive and cumbersome. 

AI cannot do the vast majority of jobs, and the only reason that anybody thinks that it can is that the vast majority of CEOs have no actual connection to the work that enriches them, and because AI can do an impression of something that looks like work, they choose to believe it can do anything. It can burp out a half-functional prototype with the company’s name on it or legitimate-looking legal or financial document, and that’s all it takes for a fuckwit with a high salary and a low IQ to think it’s capable of replacing everybody.

If I were wrong, it would actually be replacing people. You’d be able to point to both the data and the proof. You’d have single-person software companies making billions of dollars, hyperscalers would have their companies destroyed by people copying and bettering their software, accountants and lawyers and writers and every other knowledge work career would be dead, not threatened with constant layoffs that are mostly connected to improving profits, but actually dead, untenable, impossible to work in thanks to the “power of AI.”

In reality, AI is dramatic only in its mediocrity and the ferocity with which it’s proven how ignorant most authority figures and executives have become. Every boss demands you use it, every app screams at you to try its integration, every news story tells you it will replace you imminently, but in the end it doesn’t appear to do very much beyond generating and summarizing at varying levels of complexity. 

The media categorically failed to scrutinize an industry built to exploit it, as I said earlier in the week:

This hype was unsustainable without buckets of lies, misinformation and a captured tech and business media. The value of AI has been inflated by the vagueness of how it’s discussed. For example, major media outlets will gladly write that “AI can build software,” but said sentence suggests that you can just type “build me Slack 2” into Claude and have it fart out a fully-functional, production-ready piece of software, rather than a quasi-functional mound of code-slop that can do enough to trick a business idiot or lazy journalist, but little else. 

The consent has been manufactured and the markets are engorged with semiconductor stocks running because people keep mistaking the availability of debt for actual, real demand for AI compute. The geniuses in private credit and the greater markets saw the amounts that hyperscalers were spending on data centers and the ascent of OpenAI and thought “fuck me up, grandpa,” leading to $178.5 billion in data center debt deals in the US in 2025 and $50 billion in data center construction in April 2026 alone. 

Yet it turns out that data centers take anywhere from 18 to 36 months to build, with Microsoft finishing a grand total of zero of the data centers it broke ground on in 2023, and JP Morgan saying a month ago that 60% of capacity planned for completion in 2027 hasn’t even started construction, with another 7% delayed, per the Wall Street Journal. 

And despite the supposed 100GW+ of data center capacity being planned, AI compute demand doesn’t really exist outside of Anthropic and OpenAI, two companies that rely on perpetual flows of venture capital and debt to survive. Between them, they’ve raised over $200 billion in the last six months, and their revenue streams are inherently based on either unprofitable AI startups subsidizing their subscriptions, their own unprofitable subsidized subscriptions, or experimental token spend borne of companies allowing their employees to burn as much as they’d like, which is already coming to an end.

At the top of the pile lies NVIDIA, the largest company on the stock market, which sells GPUs that are so expensive that once cash-rich hyperscalers are now having to take on mountains of debt or, in Google and Oracle’s case, dump tens of billions of dollars of new stock into the markets. NVIDIA’s continued growth relies on a dwindling subset of clients, with 54% of its last quarter’s revenue and 64% of its accounts receivable coming from three customers in its last quarterly earnings. 

Demand is somehow both incredibly high for data center components but so low for AI compute that NVIDIA has agreed to spend $30 billion over the next six years to rent GPU capacity. 

That’s because the AI buildout is being driven by people who haven’t bothered to check whether the demand is real, much like AI is being adopted by people that don’t bother to do any real work, much like AI is sold based on things that it can’t actually do. 

Midwits and the incurious will say this is just like the Dot Com Bubble (it isn’t and won’t leave behind any useful infrastructure), or Uber (it isn’t) or Amazon Web Services (it isn’t) because they want to rationalize the waste. In reality, the people running the tech industry are listless Business Idiots throwing as much cash at the problem as possible rather than facing the fact that they’ve backed a dead-end technology because they’ve run out of hypergrowth ideas.

Today’s piece is an attempt at a little fun — a raucous, aggressive rundown of the major players and stories of the AI Bubble, both as a refresher for those who already know and a guide for those that don’t.

Welcome to the Hater’s Guide To The AI Bubble 3.0.

Coming Up On This Week’s Where’s Your Ed At Premium:

• The Rot-Com Bubble — A Guide To How The AI Bubble Got Inflated
• Why You Keep Being Told AI Is Powerful
• How The AI Industry Is Almost Entirely Wrappers For OpenAI and Anthropic’s Models
• How NVIDIA’s Findom Operation Conned Every Hyperscaler
• How Microsoft’s AI Strategy Has Fallen Off The Rails
• How Google Is Using AI To Destroy Its Legacy
• How Amazon Lost The Plot And Became Anthropic’s Paypig 
• How Mark Zuckerberg Burned $158 Billion To Buy GPUs For Effectively No Reason
• How SpaceX Became Musk’s Last Gasp Attempt For Exit Liquidity
• How Anthropic Is The Greatest Exploitation of the Media and Economy In Tech History To Prop Up An Unsustainable Company Run By The Most Annoying People Imaginable
• How OpenAI Became A Miserable Failson With Too Many Ideas, Unsustainable Economics, and No Plan For The Future
• How The ROI Conversation Could Burst The AI Bubble