AI 的烂账经济学

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 (updated to version 3.0 last week). 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.

Last Friday, I published the first of a two-part series where I explore the many bubbles that form the basis of the AI bubble — including the tokenomics bubble, and the cult of personality bubbles surrounding Sam Altman and Dario Amodei.  

Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week. 

Soundtrack — Local H — Manifest Destiny (Part 2)

We live in a time of deep uncertainty. On Friday, Anthropic was forced to shut off access to its Mythos and Fable models after the US government imposed an export control ban barring any non-US citizens both inside and outside of the country from accessing them. 

To explain, Fable is basically Anthropic’s supposedly “too dangerous to release” Mythos model with guardrails forbidding you from what appears to be anything biological weapons and cybersecurity, except it was jailbroken within days by Amazon researchers, leading to Amazon CEO Andy Jassy (and other unnamed companies) reporting it to the US commerce department which gave Anthropic 90 minutes to roll back Fable and Mythos due to “national security risks.” Semafor also reports that this all might have happened because China got access to Mythos.

This situation is a complete mess. PCast co-chair and podcaster David Sacks claimed that Anthropic refused to fix the issue, claiming it wasn’t serious, per Business Insider:

During the calls, Amodei tried to clear up what he assumed was a misunderstanding. He pushed back on the administration's concerns, defended the guardrails, and argued that the type of bypass that occurred, which he believed to be specific, did not pose the same risk as a broader "jailbreak" that would allow it to be used without any of the guardrails put in place by Anthropic.

In a blog post after the export controls were put in place, Anthropic said that "no testers have yet been able to find a universal jailbreak — a jailbreak method that can very broadly bypass the model's safeguards, unblocking a wide range of cyber capabilities," and that total avoidance of any jailbreaks isn't now possible for them or any other companies. They defended their systems, which they said "are so strong that many users have complained that they are overly broad."

A White House official told Business Insider that “export controls were a last resort after begging them for hours to work with us”:

Shortly after the call, the Trump administration imposed its export control on the Fable 5 and Mythos 5 models, citing national security authority and banning their use by foreign nationals, according to Anthropic. The company said the "net effect" of the order was to "abruptly disable" the models for all customers "to ensure compliance."

Anthropic claims no begging occurred, and all it got was (as noted above) 90 minutes. According to Axios, the company has dispatched some of its senior technical staff to D.C to negotiate with the Trump Administration, after virtual meetings with White House officials failed to bear fruit. 

In any case, this is a reaping/sowing for the ages. Dario Amodei has spent years selling AI models based on completely fantastical scaremongering about the “rapid advancements” of large language models, cresting the hill in April when he announced Claude Mythos, an LLM that was “too powerful to release” until June 2, when it was released to 150 organizations in 15 countries, and June 9, when it was released with said guardrails under the name “Fable.”

Fable is, of course, just another large language model that’s an indeterminate amount of “better” than the last one. Having talked to multiple people that claim to have used Mythos and deeply enjoyed Davi Ottenheimer’s takedown of its system card, it appears to be much the same model but with security protocols flimsy enough to last only a few days before anonymous researcher Pliny The Liberator broke them. Anthropic has not created recursive self-improvement, nor has it done much more than create a very large language model that gets higher benchmarks in tests built for large language models, wrapped in a veneer of mysticism and panic-hype built to scare organizations in paying them to use it.

The problem with this kind of hype is that you can only use it for so long before somebody believes you. The outright mythology of Mythos existed to scare people and help Anthropic raise at a $965 billion valuation, and because the tech industry has existed fairly divorced from reality, scrutiny, and regulation, Dario Amodei continued to inflate the “Anthropic is too powerful” bubble, believing that all that would happen would that he’d create a new enterprise API business.

Some are attempting to read this story as bullish for Anthropic — that the government will work with it to bring the models back online, creating a proxy marketing campaign for its models — and while I think that’s possible, if not likely, I think there’re many other possibilities.

On Sunday, slopagandist and Microsoft CEO Satya Nadella posted a mealy-mouthed blog on Twitter that didn’t really say very much of anything, but had two interesting comments:

The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see. If all the value is accrued by only a few models, the political economy will simply not tolerate it. There is no societal permission for an AI future that hollows out entire industries.

…

In my view, our priority has to be building a frontier ecosystem, not just a frontier model, so value flows broadly across every company, every industry, and every country. One where every organization can own the learning loop that encodes its institutional knowledge, compounding its human and token capital.

This, combined with Microsoft AI CEO Mustafa Suleyman saying Anthropic’s models were too expensive and Andy Jassy likely being part of the reason that Anthropic got banned makes me think that hyperscalers might be trying to cast doubt on the inevitability of AI labs. While Nadella’s piece has clearly gone through 8 PR people and 16 lawyers, it seems to smell of a company saying that no one model actually matters, and given that it was posted on a Sunday, I’m going to guess it’s about the current Anthropic situation.

It’s hard to see how everything goes back to normal from here. Even if Anthropic gets its models greenlit for availability, it’s clear the government has some animus against it after Q1’s battle with the Department of Defense, and may or may not have been waiting for an opportunity to rattle Dario Amodei’s cage. 

And, according to Axios, there’s a real animus between the US government and Anthropic, caused in part because of its “inability to communicate effectively,” with one source saying that “Anthropic has not done a great job at trying to speak to the administration and appreciate the ideological differences."

Alternatively, the government has taken Anthropic’s (nonsensical) marketing seriously, and thus decided to take the kind of blunt-force authoritarian position you’d expect — shut the whole thing down, as China might use Mythos to uh, do something! 

The other problem is that this is terrible, terrible timing for an AI industry in the throes of a cost crisis. Anthropic and OpenAI’s IPOs depend on myth, hype, and certainty that their growth will never slow. The government’s ability to cut off access at random based on genuine concerns or politicking isn’t a great advertisement at a time when everybody is struggling to find the ROI of AI.

This isn’t a Too Big To Fail or nationalization situation. Amazon and Microsoft are far more scared of the White House than they are of killing their golden goose, and may honestly be relieved to find a reason to bring this era to an end.

You see, Anthropic and OpenAI have much bigger problems than regulation or pissing off Pete Hegseth.

Their business models don’t fucking work.

Can We Wrap This Up Already?

I’ve been saying for years that the underlying economics of AI don’t make sense — that AI labs were intentionally obfuscating the costs of subscriptions and heavily subsidizing users’ compute, and that the moment that that changed, everything would begin to fall apart, and god damn has it finally begun.

As I discussed in last week’s premium newsletter, the AI Tokenomics Bubble is the simplest and most consequential of them all, because it comes back to something I’ve been saying for years: that the majority of users will refuse to pay the actual cost of AI. 

Said bubble inflated through the combined failure of the tech and business media to question AI’s economics and the unprecedented subsidy con perpetuated by Anthropic and OpenAI. Those who dared to suggest that OpenAI burning $5 billion was some sort of problem were dismissed as haters and skeptics that “didn’t care about the future,” with the vast majority of the media completely ignoring the economics until the latter half of 2025. 

The Tokenomics Bubble inflated because everybody aggressively ignored the AI industry’s greatest weakness, choosing instead to repeat tired mythologies about how Uber lost a lot of money (which I’ve refuted here) or Amazon Web Services cost a lot of money (Amazon’s total capex between 2003 and 2017 was $52 billion normalized for inflation) instead of being skeptical of…well, anything.

And now it’s bursting because Anthropic and OpenAI’s customers are in revolt, to the point that they’re planning “drastic” price cuts.

How The Tokenomics Bubble Burst

Alright, let’s do this one last time.

Sometime early in Q1 2026, Anthropic and OpenAI moved all of their enterprise customers to token-based billing, meaning that instead of using subsidized subscriptions with varying (and ridiculous, as I’ll get into) rate limits, big businesses suddenly had to pay for their AI usage based on the actual tokens they used. 

Many hailed this as a masterful gambit, assuming that organizations would have near-infinite budgets for AI services that had yet to prove themselves useful.

It only took a few months for OpenAI and Anthropic’s customers to start sweating. 

In the middle of April, The Information’s Laura Bratton likely burst the AI bubble with a piece about how Uber had burned through its entire annual token budget in a single quarter. 

This kicked off an industry-wide anxiety about the mounting costs of AI, with multiple other companies burning millions of dollars in the space of a few months, including Zillow, which destroyed its annual Cursor budget by the end of May. What really began the downfall was a comment by Uber COO Andrew Macdonald:

"That link is not there yet, right?" he said. "I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25% more useful consumer features."

He said that the trade-off costs from AI are harder to justify because he can't draw a direct link. Earlier this month, CEO Dara Khosrowshahi said in an earnings call that Uber was slowing hiring to counter its investments in AI.

In a single podcast, Andrew Macdonald gave the entire tech industry permission to say the truth: that nobody was actually able to show any ROI despite its massive costs.

This was always going to be a problem. By starting everybody off with subsidized subscriptions, AI labs shielded users from the costs, training them by proxy to use AI models without any concern for efficiency.  

That, and organizations are run by Business Idiots beguiled by a captured tech and business media and a complete disconnection from actual work, meaning that they’d encouraged (or forced) their workers to use AI as much as humanly possible, never once thinking about the costs until they were made to by the AI labs. All it took was a few months of tokenmaxxing to start turning organizations’ stomachs.

This began an increasingly-anxious conversation around AI’s ROI, made worse by the fact that you can’t measure the cost of a task because of the sheer number of models and harnesses, and can’t cleanly translate “AI spend” into “actual financial outcomes.” Toward the end of May, Axios would publish a story about how a company somehow spent $500 million on Anthropic tokens in a single month after failing to set up cost controls.

A few days later, Sam Altman would make a massive fuckup, saying that customers were “totally happy” with their AI spend at the beginning of the year (before token-based billing), and that spend was now a “huge issue,” likely because the costs vastly increased.

Boosters would immediately argue that these massive costs were, in fact, proof that AI was very successful, even if said “success” came from organizations that let their workers burn as many tokens as humanly possible without any consideration of the cost. As I’ve argued previously, the vast majority of Anthropic’s recent surge in revenue comes from experimental revenue from paypigs that it doesn’t deign worthy of clear visibility into their organizational token spend.

In any case, OpenAI and Anthropic need to make a combined $358 billion in annual revenue by 2029 to keep up with their $1.1 trillion in compute commitments. Any slowdown in their growth, as I discussed last week, would be fatal to two companies that have marketed themselves almost entirely by putting the cart before the horse. 

Less Than 3 Months Into Token-Based Billing, Both OpenAI And Anthropic Are Considering Price Cuts

It turns out that Altman wasn’t kidding that costs were a “huge issue” for his customers.

Around a week later, The Wall Street Journal reported that OpenAI was planning “drastic” price cuts to its token prices in response to Anthropic potentially doing the same:

OpenAI is considering drastically lowering the prices it charges users as it seeks to win customers from its rival Anthropic.

The company is weighing significant cuts to what it charges for tokens, the unit of measurement artificial-intelligence firms use to bill for their products, according to people familiar with the matter. The move would be in anticipation of similar cuts the company expects at Anthropic, the people said. 

Business executives have begun to balk at the high prices for AI usage. OpenAI Chief Executive Sam Altman said at a recent event that costs had become “a huge issue.”

If you’re wondering why they might be doing so, earlier in the day, Cisco President and Chief Product Officer Jeetu Patel said exactly what everybody had been thinking but were too scared to admit: that “...the costs of [AI tokens] are far higher than the actual value that these tokens are generating at scale.

I cannot express how deadly these price cuts would be to the AI industry, and how dangerous this conversation has become. The move to token-based billing has created a revolt in the AI industry’s customer base, coming from (as I’ve discussed) a confusion around the actual ROI and utter despair around the costs.

Depending on how “drastic” these discounts are, any (entirely theoretical) gross margin these companies make on inference will be eaten alive…all so that OpenAI and Anthropic can…uh…decrease their revenues? It’s a desperate strategy being deployed, I imagine, because of a massive wall of customer churn as a result of Business Idiots spunking millions of dollars on tokens they’re no longer able to justify. 

Remember: we’re less than three months in to organizations paying the actual costs of LLM-based services, and they’re clearly so outraged at the spiralling costs that both Anthropic and OpenAI are planning to cut the prices of an already-unprofitable service, likely collapsing their revenues while increasing their overall costs. 

I anticipate a few booster quips in response, so let’s address them head-on:

• This will make organizations spend more on AI! 
  • The problem with this idea is that it assumes that organizations are currently burning the amount of tokens they intend to burn forever, when in reality, most organizations have no idea how many tokens they want to burn, just that they’re spending way too much burning them! 
  • This means that there’s every chance this both cuts revenues and ends up with organizations using fewer tokens. Remember, nobody can actually measure the ROI of AI! A 50% price cut doesn’t actually answer the question of “why am I paying so much for this,” and unless the price cuts are to DeepSeek levels (which would also be fatal), it’s hard to see how organizations are going to be won over. 
• They’ll drop the prices then raise them again in the future!
  • Oh you sweet summer child, you really are attached to these companies, aren’t you? What do you think customers will do when the prices go up again? Do you think they’ll say “thank you so much sir for raising the prices”? Or do you think they’ll say “hey man I didn’t like these before and I don’t like them now”?
• They’ll have a haves-and-have-nots system where only some models are discounted but the expensive ones are the only good ones! 
  • …that…that’s what’s happening right now? Even if Anthropic decides it only sells Mythos or Fable or whatever to big enterprises, these are the same big enterprises that are complaining about the price!
• Jevon’s Paradox Jevon’s Paradox Jevon’s-
  • Shut the fuck up!

I Will Fucking Piledrive You If You Mention Jevon’s Paradox Again

Here’s what Jevon’s Paradox means, per Planet Money:

It was within this context that economists rediscovered the Jevons paradox. And they created a modern formulation that's a bit more nuanced. The idea is that making things like cars and appliances more energy efficient creates a "rebound effect." When you make a machine more energy efficient, it effectively lowers the cost of using it. And — hello, the classic law of demand from economics — when stuff gets cheaper, people tend to use or consume more of it.

So, for example, with more-fuel-efficient cars, it gets cheaper to travel every mile, so people drive more miles. Some may decide to stop riding the bus and buy a car. Some families may buy a second car. Others may buy bigger vehicles, like SUVs. With more-efficient light bulbs, people may keep their lights on for longer or build things like the Sphere in Las Vegas.

Newsflash! These price cuts are not happening because Anthropic or OpenAI made their products more efficient! They’re making these price cuts because their customers don’t want to pay their current prices!

In fact, their costs appear to be increasing, which is why they’ve raised (assuming the rounds completely close) over $230 billion in the last six months. You don’t do that unless you think your costs are about to explode or, I dunno, you’re about to massively increase your losses, though the timing and velocity of these price cuts suggests this was a very recent idea.

Oh, right, Jevon’s Paradox! This isn’t that. These companies aren’t getting more efficient. They don’t have any bright ideas to make their businesses lose money, and in fact seem pretty incompetent when it comes to growing their revenues outside of scamming dimwits and selling people $40 for $1.

And that is not hyperbole.

Generative AI Does Not Have A Business Model 

So, you know how I keep going on about “subsidized subscriptions”? And how people online keep saying that they’re not really subsidized?

Well, SemiAnalysis, an extremely pro-AI semiconductor analyst, ran a test made up of random long-horizon coding tasks until they maxed out the limit on OpenAI and Anthropic’s various subscription levels.

Their findings were shocking.

For $200 A Month, You Can Burn $8000 in Anthropic Tokens or $14,000 In OpenAI Tokens

That’s right. Anyone with a $200-a-month Anthropic subscription can burn $8000 in tokens, and with a $200-a-month ChatGPT subscription, you can burn $14,000 in tokens. 

This business fucking stinks! It’s not even a real business! OpenAI and Anthropic have to give away somewhere between 20 and 70 times the cost of their subscription in API tokens, which means that they realize that the vast majority of people value these tokens at a fraction of their real cost. This obscene and wasteful subsidy is what you do when you have little to no confidence in the actual value of your product! 

Sidenote booster quip: But Ed It’s The Gym Model! Newsflash, chuckles! If you’ve got 2000 people who pay $20 a month but barely cost anything it only takes three people spending $14,000 to eat every single dollar of that revenue! And trust me, I’m about to get to the margins.

SemiAnalysis also modeled out — based on the ridiculous assumption that OpenAI and Anthropic have a 75% gross margin on their tokens — what the margin of a user looks like, and I’m sure it’s f-OH MY GOD!

That’s right folks. With the current subsidies, all it takes for a user to have a gross margin of at best negative 25% is for them to use as little as 25% of their rate limit. And this is based on the generous assumption that they have a 75% gross margin on tokens!

I’ll repeat myself: this is not a real business! This is a joke business, a comedy business, a business invented by the Gods as a means of mocking venture capital! For Sam Altman and Dario Amodei to run a business in this fashion is a sign that they have utter contempt for their investors, the tech media, sell-side analysts, and the general public. If you or I ran our lives in this way, we’d be called fiscally irresponsible millennials that believe the world owes us everything.

This isn’t a real business model because generative AI companies are not real businesses. 

Generative AI does not have a business model. It is not a tool with value remotely commensurate with its costs. It isn’t getting cheaper for the providers or the customers. It isn’t becoming “better” in a way that’s measurable using anything other than benchmarks invented specifically for generative AI — an industry-wide coddling of a mediocre technology that only makes money through massive subsidies, FOMO and executive ignorance. It requires endless pre-training, post-training and script-based MacGuffins to do tasks with mathematically-guaranteed hallucinations that burn more tokens, raising costs on customers who are already in mutiny less than a quarter into being forced to pay a cost that is already unprofitable. 

Boosters and the recently-concussed will say that these companies can simply stop training, to which I say if that was possible they’d have already done it, and if they stop training, the models will eventually drift into obscurity. If stopping training was all that it’d take to turn these businesses profitable, they’d have done it already, because inference would be a money-printer rather than a cursed object eating away at Altman and Amodei’s souls. 

The AI Cargo Cult Is Collapsing

I’ve said it once and I’ll say it again: I believe a large majority of AI token spend — and specifically Anthropic’s revenue growth — has come from Business Idiots disconnected from any real work that have become convinced that “lots of AI” would do something other than rack up massive bills. 

And wouldn’t you know I was right!

A little over a month after encouraging its workers to “tokenmaxx,” Meta is now planning to pull back on its AI token spend after realizing it was on track to spend billions on tokens, per The Information: 

Meta Platforms plans to clamp down on skyrocketing AI costs inside the company by imposing limits on employees’ token usage, the company told staff in a memo on Tuesday, just weeks after it pushed them to adopt AI tools in their work.

The company is building an internal platform to track employee AI usage and spending in real time, set budgets and establish limits on employees’ token spend, according to an internal memo reviewed by The Information, which Meta shared with about 6,000 staffers earlier this week. The effort is part of a broader efficiency program aimed at cutting costs.

“We’ve seen an exponential increase in AI usage and [we] are tracking to spend billions on internal use alone in 2026,” the memo said. “At the same time, individuals and teams have limited visibility into and control over how they use AI. In 2027, we expect Meta will move toward managing AI tokens in a more structured way—with budgets, allocation decisions, and supporting tools.”

It didn’t even take two months for Meta to go from encouraging its employees to compete to burn the most tokens to talking like a British MP giving a speech about austerity measures.  

Meanwhile, The Times reports that banks are running up massive nine-to-ten figure bills from “experiments with artificial intelligence tools”:

Ben Faes, chief executive of RWS, said that businesses were becoming increasingly conscious of the costs involved, without a clear outcome on how it should be used.

“It is very exciting, but you know the cost of playing around with all this AI is rising quite dramatically,” he said.

Faes, 54, said he had spoken to two large banks which between them had racked up $1 billion in costs from experimenting with AI without generating a significant return on investment.

“It is a serious point,” he said. “AI isn’t about generating pictures of cats on skateboards. It’s becoming a serious cost centre for businesses.”

These are the kinds of things you say when you’re planning to drastically cut costs, and I think Uber’s COO gave everybody permission to admit the lack of ROI or, well, any measurable benefits of spending millions of dollars on AI tokens. 

Remember: Business Idiots are lemmings! The only reason they wanted to “do AI” was because they read it in the newspaper or heard somebody they thought was intelligent insist that it was the future. These people are extremely sensitive to suggestion (see: Nik Suresh’s Brainwash an Executive Today) and marketing hype, which means they’re also extremely sensitive to peer judgment, meaning that if the worm turns and everybody starts saying “I’m not sure we should spend as much money on AI,” they’ll become anxious to be judged as wasteful for doing something that was considered innovative mere months ago.

Some are suggesting that lower-priced open source models (including some developed in China) for some operations could be the solution, per the Wall Street Journal:

The ecosystem allows autonomous AI systems, or agents, to use cheap models—including those made by Chinese companies like Alibaba and DeepSeek—for many functions. The agents only tap the most capable versions of OpenAI’s ChatGPT and Anthropic’s Claude for more complex tasks. That can reduce costs for some AI-assisted work by as much as 95%, according to executives using the tools. 

“Once we find something that is working well and engineers love, we find ways to make it cost effective,” said Dan Robinson, founder of Detail, a startup that identifies bugs. “There’s really an embarrassment of riches right now coming out of the open source labs.”

Robinson shifted 90% of Detail’s workload from Claude and Google’s Gemini to custom models and GLM, a family of models developed in China.

The problem with this argument is that we’re yet to prove if running these models is profitable (or even sustainable) for any provider, nor do we have tangible proof that they can compete at scale with Anthropic or OpenAI’s more-complex LLMs. 

Citadel Securities argued late last week that they might be:

For the economy at large, simpler models may be the more cost-effective, productivity-augmenting pathway until physical constraints are eased. We hence see growing signs of a bifurcation in frontier vs “everyday” AI usage.

The problem is that the hundreds of billions of dollars of AI data centers full of NVIDIA GPUs are being built with the expectation that there will be incredible demand — over $150 billion a year just to cover what’s under construction — for very large and compute-intensive models. I am still skeptical that this is a real shift away, if only because using open source models requires you to either work with an inference provider or run your own GPUs. 

Nevertheless, even the hint of this migration is enough to start making Business Idiots say “hmmm, what about open source?” even if they don’t know what that means.

But everything comes back to one very simple point: that a lot of AI use (and by extension AI spend) is from the cargo cult mentality of an economy run by the most easily-led dullards in history. They jumped on the AI train because they saw a webinar or read a LinkedIn post or saw a news story about Sam Altman saying his tech was scary or an Atlantic piece saying that Claude Code was ChatGPT 2.0 and thought “fuck, I better throw as much money at this as possible.”

In the end, what is it these organizations are paying for? They’re not replacing anyone, and there isn’t compelling evidence that AI models speed people up. Allowing non-technical people to use LLMs to write code isn’t speeding up the delivery of software in a measurable way, and introduces obvious problems in the sense that, well, you’ve got a bunch of code written by somebody who can’t read or understand it. 

People will argue that AI is “really helpful with research,” despite the fact that any research you receive from AI will absolutely have hallucinations, meaning that if you don’t actually know what the answer is to a particular question (which, I assume, is why you researched it), you’re certain to have some sort of small (or large) fuckup.

In a story that’s a little on the nose, The Financial Times reported last week (covering a study by GPTZero) that a KPMG report (that’s now been taken down) about the benefits of AI had exaggerated the scale of its adoption through multiple AI hallucinations:

The October report, “Redefining excellence in the age of agentic AI”, made numerous false claims about the use of AI by organisations including the Swiss bank UBS, the UK’s National Health Service and the public transit groups Swiss Federal Railways and Transport for London.

The inaccuracies were identified as AI hallucinations by the research group GPTZero and verified by the FT. After being alerted to the issue, UBS said it would ask KPMG to remove the false claims, and the Big Four firm on Thursday pulled the report from some of its websites.

The KPMG report claimed global wealth manager UBS “integrates AI agents across investment advisory, risk management and compliance monitoring”. A spokesperson for [UBS] told the FT the assertions were “factually incorrect”.

The report also included hallucinations about AI agent use by Swiss Federal Railways, Transport for London and NHS Greater Manchester, fabricating entire integrations and product lines in a report that was likely used to justify billions of dollars of spend. 

Per GPTZero, 40 out of 45 of the report’s citations are either fake, make critical mistakes about the contents, or lack enough detail to be used as proof. They also believe that whoever wrote the report let the AI do most of the work:

Our team suspects that the authors of Total Experience used an AI-powered referencing tool to generate the report’s citations because the errors are both mistakes typical of Large Language Models (LLMs) and consistent throughout the reference list. A human would not consistently paraphrase titles, mistake topics for authors (e.g., citation 9), or repeat information across multiple components (e.g., citation 2).

GPTZero also notes that the report is being cited by LLMs as evidence to prove the success of AI agents, poisoning the already-hallucinatory well of information that these models draw upon. 

KPMG has annual revenues of over $39 billion, and sells something called KPMG Workbench which promises to “supercharge your business with [its] multi-agent AI platform, combining advanced, trusted AI agents with insights and deep industry expertise of KPMG professionals.” I assume these are the same professionals that greenlit the report.

It’s likely that this was a mixture of laziness and ignorance, but I also think it might be a situation where the person (or people) writing the report simply couldn’t find any real citations to prove their point, choosing instead to let an LLM crap out some thought-slop in the hopes that nobody would notice.

The fact that Anthropic and OpenAI have any business left after stories like these is proof that the vast majority of companies paying for these services are doing so because they feel pressured to by their peers, investors or the media. 

That’s not a tenable business model! You can only get so far on FOMO, gaslighting, and the vague promise that something good will happen if you hand over your credit card. 

Hell, let’s take it one step further: neither OpenAI nor Anthropic is a real business.

OpenAI and Anthropic Are Not Real Businesses, And Can Only Make Money By Giving It Away

Let’s cut to the chase: these aren’t real companies!

Their businesses only function by subsidizing or swindling their customer base using deceptive media campaigns that say “let people use as much AI as possible,” and it’s becoming clear that token-based billing might genuinely not work as a viable business line. 

The only hope that these companies had was the possibility that they could actually charge something approximating their real costs, though I’d argue that was only the case if there was the option for OpenAI or Anthropic to increase their token costs in the future. 

To make matters worse, it’s abundantly clear that the vast majority of people would never actually pay for the tokens they burn. If OpenAI and Anthropic are allowing their customers to burn such egregious amounts of tokens, it’s because they’ve seen that their customers churn when they don’t get to do so. Anthropic’s aggressive rate limiting in March — which still allowed people to burn far more than their subscription cost! — likely scared the everliving shit out of them, to the point that they signed up to pay Elon Musk $1.25 billion a month for access to his Colossus data centers specifically so that they could give people higher usage limits.

The only way that these companies can make money is by giving it away. Both OpenAI and Anthropic have recently started handing out $1000 in API credits to convince people to move over to Codex or Claude Code. 

Sidenote: Now OpenAI is allowing its users to “bank” their rate limits — meaning that instead of waiting for the weekly (or hourly) reset, you can choose to save them up and, I assume, use them back to back, allowing power users to effectively double-tap OpenAI’s servers once they’ve run through their usage.

Also, for the next two weeks, anyone they refer gets a free trial of Codex and both of them get another banked reset. This is a transparent attempt to juice user numbers at a cost of hundreds (or thousands) of dollars, and will almost exclusively be used by power users gaming the system.

Their services are not valuable enough for people to cover their business expenses, even if you remove the cost of training, which is so severe it drowns out every dollar of revenue on its own. They cannot raise prices — or even bring them in line with their costs — without their users flipping out. Their training costs are necessary to continue making their models an indeterminate level of “better,” which means that they’re a cost of goods sold, and not a capital expenditure. 

As an aside: I’ve been told by somebody that Anthropic has been telling people that they can consider token spend a capital expenditure. I am warning any company in the entire world that if I find out you did this, I will haunt you for the rest of time. I will watch everything you do forever, as this is bullshit accounting that verges on fraudulent, and I can imagine some asshole is going to do it. 

Anthropic and OpenAI want you to believe that their businesses can somehow turn profitable, yet neither of them have any explanation as to how. Anthropic negotiated discounted compute for the first two months of its SpaceX deal as a means of pretending to be profitable for a single quarter, but any price cut — or even customer churn! — will immediately put its finances in a kind of red usually reserved for the deeply embarrassed or steroid-enhanced. 

They do not have a plan. 

You can go on about TPUs, Trainium, Inferentia, and custom silicon for as long as you like — it’s not profitable to run these companies, their costs are too high, and their customers are price-sensitive. Their customers lasted less than three months paying for their actual token burn before crying for mercy. There is no reversing this trend, because if there were, OpenAI and Anthropic would’ve reversed it in any way, shape or form, rather than raising more money than anyone has ever raised before for what appears to be no reason other than to burn it.

OpenAI and Anthropic are unsustainable and recklessly-run companies that do not make sense outside of the broken world of Silicon Valley. The tech industry and venture capital are run by a coterie of has-beens who create no value, and the vague memories of the pre-2015 era, before investors gave up on investing on seed stage companies and decided to joylessly trend-hop for years until they were driven insane by COVID lockdowns and “X: The Everything App.” 

The tech industry is run by people who do not experience real problems or have to run real businesses, because a cluster of fellow grifters will vault them back into the black. Investments are no longer made based on rugged meritocracy or any interest in creating the future — it’s only about the Rot Economy’s mediocre growth-at-all-costs accelerationism and making varying numbers go up, though very rarely ones associated with profits.

I think it’s fair at this point to ask whether you could’ve just hired real people to do the shit that AI has done given its enormous cost. $14,000 could probably get you a great deal out of a real software engineer — hell, you could’ve probably hired an agency to do the work for you and actually have someone manage the risk. 

The completely imaginary assumption about the AI industry is that it’ll magically get cheaper. That is not something that’s happening. More data centers will not make OpenAI or Anthropic profitable. More data centers will not make customers more willing to pay the actual cost of AI. More venture capital funding will not make Anthropic or OpenAI real businesses. 

I agree with anyone saying there should be a pause in the development of generative AI, but I do so based on the belief that this is a doomed grift and science experiment masquerading as an industry that has only gone on this long because it allowed the hardware industry to extract hundreds of billions of dollars from startups, venture capitalists, asset managers, and retirement and insurance funds.

And anyone in Silicon Valley fooling themselves into believing they’re anything other than a corporate stooge is a mark.

Silicon Valley Is A Monoculture

The AI industry is the direct opposite of what made Silicon Valley famous. 

It is a flattening of everything, absorbing the majority of venture capital funding, media attention, talent, and intellectual oxygen, invading whatever space you’re in because investors insist you must have something to do with AI and because everybody has been convinced they have to use it. It is an intellectual black hole, dragging every conversation toward it, demanding the most money, the most focus, endless justifications and defenses from people that must be rejected for questioning whether LLMs are the future. It debases and humiliates its fans by forcing them to constantly face indignities and embarrassments like it deleting entire databases or breaking AWS. It stunts the intellects of those who use it and, in demanding complete devotion to be considered “part of Silicon Valley,” suffocates the kind of meritocratic skepticism that allegedly got these fuckers so rich. 

Silicon Valley was founded on the potentially fictional idea of plucky software developers that rejected the bounds of corporatism. It’s now ingested the worst qualities of corporate America — groupthink, trend-hopping, tribalism, hero-worship, managerial feudalism, and wasteful spending chasing things based on what might make a rich, heavily-coddled oaf smile. There is nothing daring or individualistic about Silicon Valley. At this point, you may as well work at fucking McKinsey. 

Silicon Valley is the establishment. OpenAI and Anthropic are effectively owned by Microsoft, Google and Amazon — they do not have infrastructural or financial dependence, they principally run on their hardware, and if anything happens to them, they will likely be absorbed into multiple arms of the Magnificent Seven. 

Their financial success benefits only the richest people in Silicon Valley and the wealthiest companies on the stock market. They sell themselves as democratizing software as they extract as many dollars as possible from venture capital, all while selling them back a story of spreading “abundance.” 

AI represents the commoditization of startups as a fuel for tech firms with trillion-dollar market caps. AI startups exist only to send money upwards, burning Claude or GPT tokens that run on infrastructure built and owned by the very incumbents that the Valley allegedly takes pride in unseating. 

The groupthink and monoculture of the Valley has gaslit these poor individuals into believing that there’s some sort of happy ending rather than a slow descent into insolvency, duping them into defending expensive, unsustainable tools using mythology that benefits only the richest people on Earth. 

Someone recently said they think Anthropic and OpenAI are “the last startups,” saying that there was no point in building anything else as “everything has been solved or will be shortly.”

I agree, though for totally different reasons.

Anthropic and OpenAI represent what I believe may be the last hypergrowth startups, and their collapse (however it may happen) will represent the end of the dream of founding a little company that turns into the next Google or Meta. 

Neither company was possible without the involvement of Microsoft, Google or Amazon, who provided their earliest funding and, most importantly, their entire physical infrastructure. Anthropic and OpenAI were always entirely dependent on these hyperscalers to shoulder the $100bn+ in infrastructure costs to make training their earliest models or serving inference possible.

The reason there are no other Anthropic or OpenAI-sized startups is that neither of them are actually startups. These are not plucky underdogs who shoulder-barged their way to near-trillion dollar valuations — they’re quite literally subsidiaries of the largest companies in the world, using the mythology of the startup ecosystem to create the mistaken belief that anyone can actually compete with big tech. AI startups are all entirely dependent on big tech, yet sell themselves as rugged individuals.

The fact that Amazon deliberately dobbed Anthropic in to the Commerce Department and neither Microsoft nor Google have shown any interest in defending it suggests that neither really cares if it lives or dies. This would be the exact situation that would prove that Anthropic (or OpenAI) had real leverage over their hyperscale benefactors. Instead, the largest companies in the world have left them to the wolves. 

Anthropic believed it was too big to fail, or at least too big to be stopped. It likely believed it would see a flood of support as it did with its argument with the Department of Defense, but nobody seems particularly interested in defending it. Instead, everybody seems kind of confused and annoyed, and the largest companies in the world are making vague statements about how no one model can be “the best.” 

Silicon Valley, this is your King — a company that grew through conning and scaring and lying to people at scale, overstating both the capabilities and possibilities of its models in the hopes that everybody would be too scared not to pay for them, only to find its business model collapse because you can’t wish your way to a fucking business model. 

While OpenAI is no better, Anthropic is offensive in that it resembles everything that’s ruined the tech industry — a company with a product that costs billions of dollars that can only be sold by talking about what it might do in the future, a masterpiece of grift and hubris that I believe will stumble and crumble in the future.

The next generation of startups will not get built in a system more interested in Twitter clout and trend-chasing than making good software that solves real problems. Braindead, growth-drunk “accelerationists” conflate economic growth with human progress, and as long as they’re in power, the only ones who will build things of note will be the actual outcasts. 

You can’t win as a startup anymore. There is no competing with or scaling without the Magnificent Seven, at least not under the current terms of Silicon Valley. 

And there never will be again without aggressively flushing away the hubris and ignorance of the current generation of venture capitalists that have abandoned building the future in favor of praying at the feet of management consultants and grifters.

—

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 10,000 to 18,000 words, including vast, detailed analyses of the biggest events and companies in the AI bubble.