AI Is A Money Trap

In the last week, we’ve had no less than three different pieces asking whether the massive proliferation of data centers is a massive bubble, and though they, at times, seem to take the default position of AI’s inevitable value, they’ve begun to sour on the idea that it’s going to happen soon.

Meanwhile, quirked-up threehundricorn OpenAI has either raised or is about to raise another $8.3 billion in cash, less than two months since it raised $10 billion from SoftBank and a selection of venture capital firms. 

I hate to be too crude, but where the fuck is this money going? Is OpenAI just incinerating capital? Is it compute? Is it salaries? Is it compute? Is it to build data centers, because SoftBank isn’t actually building anything for Stargate? 

The Information suggested OpenAI is using the money to build data centers — possibly the only worse investment it can make other than generative AI, and it’s one that it can’t avoid because OpenAI also is somehow running out of compute. And now they're in "early-stage discussions" about an employee share sale that would value the company at $500 billion, a ludicrous number that shows we're leaving the realm of reality. To give you some context, Shopify's market cap is $197 billion, Salesforce's is $248 billion, and Netflix's is $499 billion. Do you really think that OpenAI is worth more than these companies? Do you think they're worth more than AMD at a $264 billion market cap? Do you?

AHhhhhhhh-

Amongst this already-ridiculous situation sits the issue of OpenAI and Anthropic’s actual revenues, which I wrote about last week, and have roughly estimated to be $5.26 billion and $1.5 billion respectively (as of July). In any case, these estimates were made based on both companies’ predilection for leaking their “annualized revenues,” or monthx12. 

This extremely annoying term is one that I keep bringing up because it’s become the de-facto way for generative AI companies to express their revenue, and both OpenAI and Anthropic are leaking them intentionally, and doing so in a way that suggests they’re not using even the traditional ways of calculating them. OpenAI leaked on July 30 2025  that it was at $12 billion annualized revenue — so around $833 million in a 30-day period — yet two days later on August 1 2025 the New York Times reported they were at $13 billion annualized revenue, or $1.08 billion of monthly revenue.

It’s very clear OpenAI is not talking in actual calendar months, at which point we can assume something like a trailing 30 day window (as in the “month” is just 30 days rather than a calendar month). We can, however, declaratively say that it’s not doing “the month of June” or “the month of July” because if it was, OpenAI wouldn’t have given two vastly different god damn numbers in the same two day period. That doesn’t make any sense. There are standard ways to handle annualized revenue, and it's clear they're not following them.

And to be even clearer, while I can’t say for certain, I believe these leaks are deliberate. OpenAI’s timing matches exactly with fundraising. 

On Anthropic’s side, these revenues are beginning to get really weird. Anthropic went from making $72 million ($875 million annualized) in January to $433 million in July — or at least, it leaked on July 1, 2025 that it was at $4 billion annualized to The Information ($333 million a month) and claimed it had reached $5 billion annualized revenue ($416 million) to Bloomberg on July 29 2025 . 

How’d it get there? I’m guessing it was from cranking up prices on Cursor, and we’ve had the confirmation that’s the case thanks to The Information reporting that $1.4 billion of its annualized revenue is from its top two customers (so around $116 million a month), the biggest of which is Cursor. Confusingly, The Information also says that Anthropic’s Claude Code is “generating nearly $400 million in annualized revenue, roughly doubling from just a few weeks ago,” meaning about $33 million of monthly revenue. 

In any case, I think Cursor is a huge indicator of the current fragility of the bubble — and the fact that for most AI startups, there’s simply no way out, because being acquired or going public does not appear to be a viable route. 

Cursor Is A Systemic Risk To The AI Industry

I know it sounds a little insane, but I believe that Cursor is the weak point of the entire AI bubble, and I’ll explain why, and how this could go. This is, by no means, inevitable, but I cannot work out what Cursor does other than this.

1. Cursor makes — before, at least, their massive changes to their service — $500 million in annualized revenue, so around $42 million a month. This makes it the single-highest earning generative AI company that isn’t called OpenAI or Anthropic, and the highest-earning company built on top of (primarily) Anthopic’s technology. Its success is symbolic to the greater movement, and just as it hit its peak, Anthropic (and OpenAI, to a lesser extent) decided to add priority processing and priority service tiers, demanding more money up front and causing Cursor to have to massively degrade its service. I explain in detail in my premium piece from a few weeks ago.
   1. To explain in short, Cursor’s AI-powered coding editor used to have fairly unrestrained access to the various models provided by these companies. In mid-June — a few weeks after Anthropic introduced “priority tiers” that required companies to pay up-front and guarantee a certain throughput of tokens and increased costs on using prompt caching, a big part of AI coding — Cursor massively changed the amount its users could use the product, and introduced a $200-a-month subscription. 
      1. As an aside to this, Anthropic also competes with Cursor’s AI coding product with their own service, Claude Code. 
2. Cursor, as Anthropic’s largest client (the second largest being Github Copilot), represents a material part of its revenue, and its surging popularity meant that they were sending more and more revenue Anthropic’s way.
3. Anthropic used this opportunity to raise prices on accessing its models to continue providing service at an acceptable level to Cursor’s customers by introducing “Priority Tier” access on May 30 2025.
4. This has allowed Anthropic to juice its revenues, and due to the upfront nature of these contracts, Cursor is locked-in regardless of how well it does. The net result of these cost increases means that Cursor’s product is less attractive to its customers, and will thus make it less money.
5. At this point, one has to ask — how does Cursor survive? Their product isn’t profitable, and the means it used to make their company so successful have become untenable. It has guaranteed a certain throughput of tokens-per-second to the major model developers, chief of them Anthropic, but Cursor itself said it’s signed multi-year deals with multiple cloud providers like OpenAI, xAI and Google.
6. Cursor’s product is now worse. People are going to cancel their subscriptions. Its annualized revenue will drop, and its ability to raise capital will suffer as a direct result. It will, regardless of this drop in revenue, have to pay the cloud companies what it owes them, as if it had the business it used to. I have spoken to a few different people, including a company with an enterprise contract, that are either planning to cancel or trying to find a way out of their agreements with Cursor.
7. If Cursor is allowed to die, it will be unable to pay a chunk of Anthropic’s revenue — and yes, the revenue of other providers too. It will also bring into question whether it’s possible to build — putting aside any questions of profitability — a business of any kind offering services built on top of generative AI, and in turn bring into doubt the veracity of investing in this sector.
8. It will also call into question whether any other generative AI company is a real business. 
9. This will naturally lead to the question of why we’re building all these god damn data centers!

Cursor, at this point, faces two options: die, or get acquired. This is not an attack on anyone who works at the company, nor anything personal. The unit economics of this business do not make sense and yet, on some level, its existence is deeply important to the valley’s future. 

So, who could acquire Cursor? 

OpenAI? OpenAI couldn’t acquire Windsurf because it was too worried Microsoft would get the somehow-essential IP of one of what feels like a hundred different AI-powered coding environments. It also already tried and failed to buy Cursor, and if I’m honest, I bet Cursor would sell now. Honestly, Cursor fucked up bad not selling then. It could have got $10 billion and Sam Altman would’ve had to accelerate the funding clause. It would’ve been so god-damn sick, but now the only “sick” thing here is Cursor’s fragile, plagued business model. 

How about Anthropic? Eh! It already has their own extremely-expensive coding environment, Claude Code, which I estimated loses the company 100% to 10,000% of a subscription per-customer a few weeks ago, and now Anthropic is adding weekly limits on accounts, which will, I believe, create some of the most gnarly churn in SaaS history. Also, does Anthropic really want to acquire its largest customer? Also, with what money? It’s not raising $5 billion to bail out Cursor. Anthropic needs that to feed directly into Andy Jassy’s pocket to keep offering increasingly-more-complex models that never quite seem to be good enough. 

Google? It just sort-of-bought Windsurf! It can’t do that again. It’s already given out the participation trophy multiple billions of dollars to investors and founders so nobody has to get embarrassed about this, and then allowed Cognition to pick up the scraps of a business that made $6.83 million a month after burning $143 million of investor capital (TechCrunch reports Windsurf was left with $100 million in cash post-acquisition). TechCrunch also reports that Cognition paid $250 million for what remained, and that this deal didn’t actually pay out the majority of Windsurf’s employees,

Meta? If I’m Cursor’s CEO, I am calling Mark Zuckerberg and pretending that I think the only person in the world who can usher in the era of Superintelligence is the guy who burned more than $45 billion on the metaverse and believes that not wearing AI glasses in the future will be a disadvantage. I would be saying all manner of shit about the future, and that the only way to do this was to buy my AI-powered coding startup that literally can’t afford to exist.

And that really is the problem. These companies are all going through the same motions that every company before them did — raise as much money as possible, get as big as possible, and eventually scale to the point you’re fat with enterprise cash. 

Except the real problem is that, just like big tech’s new gluttony of physical real estate it's taken on, generative AI companies are burdened with a constant and aggressive form of cloud debt — the endless punishment of the costs of accessing the API for generative AI models that always seem to get a little better, but never in such a way that anything really changes other than how much Anthropic and OpenAI are going to need at the end of the month or they break your startup’s legs.

I’m not even trying to be funny! Anthropic raised its prices on Cursor so severely it broke its already-unprofitable business model. These products — while also, for the most part, not producing that much revenue — need to be sold with users being aware of (and sensitive to) the cost of providing them, and Cursor’s original product was $20-a-month for 500 “fast requests” of different models, in the same way that accessing Claude Code on any subscription is either $20, $100, or $200 a month rather than paying per API call, because these companies all sell products that shield the customer from the actual costs of running the services.

The irony is that, despite being willing to kill these companies by fundamentally changing the terms upon which they access these models, Anthropic is also, in some way, dependent on Cursor, Replit, and other similar firms continuing to buy tokens at the same rate as before, as that consumption is baked into its ARR figures, as well as the forward-looking revenue projections. 

It is, in some sense, a Kobayashi Maru. Anthropic has an existential need to screw over its customers by hiking rates and imposing long-term commitments, but its existence is also, in some way, predicated on these companies continuing to exist. If Cursor and Replit both die, that’s a significant chunk of Anterior’s API business gone in a flash — and, may I remind you, that significantly overshadows its subscription business (making it almost like an inverse of OpenAI, where subscriptions drive the bulk of revenue). 

Anthropic’s future is wedded to Cursor, and I just don’t see how Cursor survives, let alone exits, or gets subsumed by another company in a way that mirrors how acquisitions have worked since…ever.  

If Cursor does not sell for a healthy amount — I’m talking $10 billion plus, and I mean actually sell, not “the founders are hired in a strange contractual agreement that pays out investors and its assets are sold to Rick from Pawn Stars” — it will prove that no generative AI company, to this date, has actually been successful. In reality, I expect a Chumlee-esque deal that helps CEO Michael Truell buy a porsche while his staff makes nothing.

Is Cursor worth $10 billion? Nope! No matter how good its product may or may not be, it is not good enough to be sold at a price that doesn’t require Cursor to incinerate hundreds of millions of dollars with no end in sight.

And this ultimately gives us the real conundrum — why aren’t generative AI startups selling?  

No, Really, Why Are There So Few Generative AI Acquisitions?

Before we go any further, there have been some acquisitions, but they are sparse, and seem almost entirely centered around bizarre acqui-hires and confusing fire sales.

AMD bought Silo AI, “the largest private AI lab in Europe,” in August 2024 for $665 million, which appears to be the only real acquisition in generative AI history, and appears to be partially based on Silo’s use of AMD’s GPUs. 

Elsewhere, NVIDIA bought OctoAI for an estimated $250 million in September 2024, after buying Brev.dev in July 2024 for an undisclosed sum, and then Gretel in March 2025. Yet in all three cases these are products to deploy generative AI, and not products built on top of generative AI or AI models. Canva bought “generative AI content and research company” Leonardo.AI in July 2024 for an undisclosed sum.

Really, the only significant one I’ve seen was on July 29 2025 — publicly-traded customer service platform NICE buying AI-powered customer service company Cognigy in a $955 million deal. According to Cxtoday, Cognigy expects about $85 million in revenue this year, though nobody appears to be talking about costs. However, Cognigy, according to some sources, charges tens or hundreds of thousands per contract for its “AI voice agents” that can “understand and respond to user input in a natural way.”

Great! We’ve got one real-deal “company built on models” acquisition, and it’s a company that most people haven’t heard of making around $7 million a month. 

Let’s take a look at the others. 

• Inflection AI to Microsoft, which was not an acquisition but a “$650 million licensing deal” that, according to FastCompany, may be more like $1 billion when you include things like how much it paid Inflection CEO and former Deepmind co-founder Mustafa Suleyman.
  • According to FastCompany, the deal involves a license to sell Inflection’s models, a waiver against any employee claims against Inflection or Microsoft, paying off investors, and some sort of unnamed compensation for employees. 
  • Oof! That’s a stinky deal.
• Windsurf to Google (and Cognition), which was also not an acquisition. Windsurf’s c-suite went to Google for $2.4 billion, which paid them off along with its investors, and then the rest of the staff and the product got acquired by Cognition for $250 million.
  • According to TechCrunch, investors made $1.2 billion on the deal, with Windsurf co-founders Varun Mohan and Douglas Chen making another $1.2 billion, and its staff getting to start a new job at a different company building something else, with, according to TechCrunch “a large portion of Windsurf’s approximately 250 employees” not benefiting from the deal.
  • As an aside, Mohan and Chen fucking suck. Couldn’t afford to break off some of those billions for your people, huh? What a pair of fucking assholes.
• Io Products to OpenAI, an all-stock acquisition. This deal is a farce and it’s unclear if OpenAI actually bought anything. $6.4 billion in stock? For what? Jony Ive’s weird face staring at you lovingly as he says stuff like “I Think We Should Make It Look Like A Circle” while taking a $5 million dollar salary? Get outta here. Not real money. 
• Character.ai to Google. This too was not an acquisition. Google, to quote the Wall Street Journal, “paid $2.7 billion to bring back an AI Genius who quit in frustration.” While I don’t like the term “genius,” Shazeer was one of the authors of the original “Attention Is All You Need” paper that began the Large Language Model era. Nevertheless, much like Inflection, Google paid a licensing fee to Character.ai for its models and hired, according to The Information, “its cofounders and many of its engineers,” creating a fund that would pay out any vesting shares (as in the shares you are given when you join a company that you accrue over the time you work there) until July 2026. 

Outside of one very industry-specific acquisition, there just doesn’t seem to be the investor hunger to buy a company valued at $9.9 billion.

And you have to ask why. If AI is, as promised, the thing that’ll radically change our economy, and these companies are building the tools that’ll bring about that change, why does nobody want to buy them? 

And, in the broader term, what does it mean when these companies — those with $10bn, or in the case of OpenAI, $300bn valuations — can’t be bought, and can’t go public? Where does this go? What happens next? What’s the gameplan here? How will the venture firms that ploughed billions of capital into these businesses bring a return for their LPs if there are no IPOs or buyouts? 

The economic implications of these questions are, quite frankly, terrifying — especially when you consider the importance that VC has historically held in building the US tech ecosystem, and they raise further questions about the impact of an AI bubble on companies that are promising, and do have a viable business model, and a product with actual fit, but won’t be able to actually raise any cash. 

“But Ed, what if Cursor turns profitable now?”

Great! I would believe it was possible if it had ever, ever happened, which it has not. 

I’m not even being sarcastic or rude. It has just not happened. No company that actually stakes their entire product on generative AI appears to be able to make money. Glean, a company that makes at best $8.3 million a month ($100 million annualized revenue) said it had $550 million in cash December of last year, and then had to raise $150 million in June of this year. Where did that money go? Why does a generative search engine product with revenues that are less than a third of the Cincinnati Reds baseball team need half a billion dollars to make $8.3 million a month? 

I’m not saying these companies are unnecessary, so much as they may very well be impossible to run as real businesses. This isn’t even a qualitative judgment of any one generative AI company. I’m just saying, if any of these were good businesses, they would be either profitable or being acquired in actual deals, and there would be good businesses by now. 

The amount of cash they are burning does not suggest they’re rapidly approaching any kind of sane burn rate, or we would have heard. Putting aside any kind of skepticism I have, anything you may hold against me for what I say or the way I say it, where are the profitable companies? Why isn’t there one, outside of the companies creating data to train the AI models, or Nvidia? We’re three years in, and we haven’t had one.

We also have had no exits and no IPOs. There has been no cause for celebration, no validation of a business model through another company deciding that it was necessary to continue its dominance by raising funds on the public market, or allowing actual investors — flawed though they may be — act as the determiner of their value. 

It is unclear what the addition of Windsurf’s intellectual property adds to Cognition, much like it’s a little unclear what differentiates Cognition’s so-called AI-powered software engineer “Devin” from anything else on the market. I hear Goldman is paying for it, and said the stupidest shit I’ve ever heard to CNBC that nevertheless shows how little it’s actually paying for:

“We’re going to start augmenting our workforce with Devin, which is going to be like our new employee who’s going to start doing stuff on the behalf of our developers,” Argenti told CNBC. “Initially, we will have hundreds of Devins [and] that might go into the thousands, depending on the use cases.” 

Hundreds of Devins = hundreds of seats. At a very optimistic 500 users at the highest-end pricing of $500-a-month (if it’s $20-a-month, Cognition is making a whole, at most, less than $20,000 a month) — and let’s assume that it does a discount at enterprise scale, because that always happens — that’s $250,000 a month! Wow! $3 million in revenue? On a trial basis? Amazing!

Sidenote: I'm so impressed! To be clear, it’s probably far fewer seats and far fewer dollars a month. 

In fact, I can’t find a shred of evidence that Cognition otherwise makes much money. Despite currently raising $300 million at a $10 billion valuation, I can find no information about Cognition’s revenues beyond one comment from The Information from July 2024, when Cognition raised at a $2 billion valuation:

Cognition’s fundraise is the latest example of AI startups raising capital at sky-high valuations despite having little or no revenue.” 

In a further move per The Information that is both a pale horse and a deeply scummy thing to do, Cognition has now laid off 30 people from the Windsurf team, and is now offering the remaining 200 buyouts equal to 9 months of salary and, I assume, the end of any chance to accrue further stock in Cognition. CEO Scott Wu said the following in the email telling Windsurf employees about the layoffs and buyouts:

“We don’t believe in work-life balance—building the future of software engineering is a mission we all care so deeply about that we couldn’t possibly separate the two,” he said. “We know that not everyone who joined Windsurf had signed up to join Cognition where we spend 6 days at the office and clock 80+ hour weeks.”

All that piss, vinegar, and burning of the midnight oil does not appear to have created a product that actually matters. I realize this is a little cold, but if you’re braying and smacking your chest about your hard-charging, 6-days-a-week office culture, you should be able to do better than “we have one publicly-known customer and nobody knows our revenue.” Maybe it’s a little simpler: Cognition paid $250 million to acquire Windsurf so that it could, after the transaction, say they have $82 million in annualized revenue.

If that’s the case, this is one of the dodgiest, weirdest acquisitions I’ve seen in my life — two founders getting a few hundred million dollars between them and their investors, and a few of their colleagues moving with them to Google, leaving the rest of the staff effectively jobless or in Hell with little payoff for their time working at Windsurf.

I can only imagine how it must have felt to go from being supposedly acquired by OpenAI to this farcical “rich get richer” bullshit. It also suggests that the actual underlying value of Windsurf’s IP was $250 million.

So, I ask, why, exactly, is Cognition worth $10 billion? And why did it have to raise $300 million after raising “hundreds of millions” according to Bloomberg in March? Where is the money going? It doesn’t seem to have great revenue, Carl Brown of the Internet of Bugs revealed it faked the demo of “Devin the AI powered software developer” last year, and Devin doesn’t even rank on SWE-benchmark, the industry standard for model efficacy at coding tasks. 

At best, it’s now acquired their own unprofitable coding environment and the smidgen of revenue associated. How would Cognition go public? What is the actual exit path for Cognition, or any other generative AI startup? 

Get Acquired, Go Public, Or Die

And that, right there, is Silicon Valley’s own housing crisis, except instead of condos houses they can’t afford with sub-prime adjustable rate mortgages, venture capitalists have invested in unprofitable, low-revenue startups with valuations that they can never sell at. And, like homeowners in the dismal years of 2008 and 2009, they’re almost certainly underwater — they just haven’t realized it yet.

Where consumers were unable to refinance their mortgages to bring their monthly payments down, generative AI startups face pressure to continually raise at higher and higher valuations to keep up with their costs, with each one making it less likely their company will survive. 

The other difference is that, in the case of the housing crisis, those who were able to hold onto their properties eventually saw their equity recover to their pre-crash levels, in part because housing is essential and because its price is influenced just as much by supply and demand, as it is the ability for people to finance the purchase of properties, and when the population increases, so too does the demand for housing. None of that is true with AI. There’s a finite number of investors, a finite number of companies, and a finite amount of capital — and those companies are only as valuable as the expectations that investors have for them, and as the broader sentiment towards AI. 

Who is going to buy Cognition? Because the only other opportunity for the investors who put the money into this company to make money here — let alone to recoup their initial investment —  is for Cognition to go public. Do you think Cognition will go public? How about Cursor? It’s worth $9.9 billion, and there was a rumour that it was raising at a valuation of $18 billion to $20 billion back in June.

Do you see Perplexity, at a valuation of $18 billion, selling to another company? The alternative, as discussed, is that Perplexity, a company with 15 million users and, at $150 million annualized revenue, is still making less than half of the revenue of the Cincinnati Reds baseball team ($325 million in annual revenue, and that’s real money, not “annualized revenue”), must go public. Perplexity has, at this point, raised over a billion dollars to lose $68 million in 2024 on $34 million of revenue. 

By comparison, the Cincinnati Reds is a great business, with a net monthly income of $29 million, all to provide a service that upsets and humiliates millions of people from Ohio every year for the pleasure of America. 

Putting aside the Reds, what exactly is it that Perplexity could offer to the public markets as a stock, or to an acquirer? Apple considered acquiring it in June, but Apple tends to acquire the companies it wants to integrate into the core business (as was the case with Siri), which makes me think that Perplexity leaked information about a deal that was never really serious. Hell, Meta talked about acquiring it too. Isn’t it weird that two different companies talked about buying Perplexity but neither of them did it? CEO Aravind Srivinas said in July that he wanted to “remain independent,” which is a weird thing to say after talking to two giant multi-trillion-dollar market cap tech firms about selling to them. 

It’s almost as if nobody actually wants to buy Perplexity, or any of these sham companies, which I know sounds mean, but if you are worth billions or tens of billions of dollars and you can’t make more than a bottom-tier baseball team in fucking Ohio, you are neither innovative nor deserving of said valuation. 

But really, my pissiness and baseball comparisons aside, what exactly is the plan for these companies? They don’t make enough money to survive without a continuous flow of venture capital, and they don’t seem to make impressive sums of money even when allowed to burn as much as they’d like. These companies are not being forced to live frugally, or at least have yet to be made to, perhaps because they’re all actively engaged at spending as much money as possible in pursuit of finding an idea that makes more money than it loses. This is not a rational or reasonable way to proceed.

Yes, there are startups that can justify burning capital. Yes, there are companies that have burned hundreds of millions of dollars to find their business models, or billions in the case of Uber, but none of these companies are like those companies in the generative AI space. GenAI businesses don’t have the same economics, nor do they have the same total addressable markets. If you’re going to say “Amazon Web Services,” I already explained why you’re wrong a few weeks ago.

These startups are their VC firms’ subprime mortgages, overstuffed valuations with no exit route, and no clear example of how to sell them or who to sell them to.

The closest they’ve got is using generative AI startups as beauty pageants for guys wearing Patagonia, finding ways to pretend that the guy who runs an AI startup — sorry, AI lab — is some sort of mysterious genius versus just another founder in just another bubble with just another overstuffed valuation. 

The literal only liquidity mechanism (outside of Cognigy) that generative AI has had so far is “selling AI talent to big tech at a premium.” Nobody has gone or is going public, and if they are not going public, the only route for these companies is to either become profitable — which they haven’t — or sell to somebody, which they do not.

But I’ve been dancing around the real reason they won’t sell: because, fundamentally, generative AI does not let companies build something new. Anyone that builds a generative AI product is ultimately just prompting the model, albeit in increasingly more-complex ways at the scale of something like Claude Code — though Anthropic has the advantage of being one of the main veins of infrastructure. This means that a generative AI company owns very few unique things beyond their talent, and will forever be at the mercy of any and all decisions that their model provider makes, such as increasing prices or creating competing products. 

I know it sounds ludicrous, but this is the reality of these companies. While there are some companies that have some unique training and models, none of them seem to be building interesting or unique products as a result. 

If your argument is that these things take some time — how long do they have? 

No, really! So many of you have said that “this is what happens, they burn a bunch of money, they grow, and then…” and then you stop short because the next thing you say is “turn profitable by getting enterprise customers.” Nobody can do the first part and few can do the second part in anything approaching a consistent fashion. 

But really, how long should we give them? Three years? 

Perplexity’s had three years and a billion dollars, it doesn’t seem to be close to profitable. How long does Perplexity deserve, exactly? An eternity? 

Every single example of a company that has “burned a lot of money and then not done so in the end” has been a company with a physical thing or connections to the real world, with the exception of Facebook, which was never the kind of cash-burning monstrosity that generative AI is. 

There has never been a software company that has just chewed through hundreds of millions — or billions — of dollars and then suddenly became profitable, mostly because the magical valuations of software have been in their ability to transcend infrastructure. One’s unit economics in the sales of software like Microsoft Office or providing access to Instagram do not require the most powerful graphics processing units run at full tilt at all times, and those are products that people like and want to use every day. 

I get people saying “they’re in the growth stage!” about a few companies, but when all of them are unprofitable, and even the unprofitable ones outside of OpenAI and Anthropic aren’t really making impressive amounts of money anyway? C’mon! This isn’t anything like any boom that leads to something, and it’s because the economics do not make sense.

And that’s before we get to OpenAI and Anthropic!

OpenAI and Anthropic, And The Impossible Road Ahead

So, as a reminder, OpenAI appears to have burned at least ten billion dollars in the last two months. It is has just raised another $8.3 billion dollars (after raising $10 billion in June according to the New York Times), and intends to receive around $22.5 billion by the end of year from SoftBank, and that is assuming it becomes a for-profit entity by the end of the year, and if that doesn’t happen, the round gets cut to $20 billion total, meaning that SoftBank would only be on the hook for a further $1.7 billion.

I am repeating myself, but I need you to really get this: OpenAI just got $10 billion in June 2025, and had to raise another $8.3 billion in August 2025. That is an unbelievable cash burn, one dwarfing any startup in history, rivalled only by xAI, makers of “Grok, the racist LLM,” losing it over $1 billion a month.

I should be clear that if OpenAI does not convert to a for-profit, there is no path forward. To continue raising capital, OpenAI must have the promise of an IPO. It must go public, because at a valuation of $300 billion, OpenAI can no longer be acquired, because nobody has that much money and, if let’s be real, nobody actually believes OpenAI is worth that much. The only way to prove that anybody does is to take OpenAI public, and that will be impossible if it cannot convert.

And, ironically, Softbank’s large and late-stage participation makes any exit harder, as early investors will see their holdings diluted as a percentage of total equity — or whatever the hell we’re calling it. While a normal company could just issue equity, and deal with the dilution that way, OpenAI’s structure necessitates a negotiation where companies can obstruct the entire process if they see fit. 

Speaking of companies that might obstruct that transition, let’s talk about Microsoft.  As I asked in my premium newsletter a few weeks ago, what if Microsoft doesn’t want OpenAI to convert? It owns all the IP, it owns access to all OpenAI’s research, and already runs most of its infrastructure. While — assuming a best-case scenario — that it would end up owning a massive chunk of the biggest tech startup of all time (I’m talking about equity, not OpenAI’s current profit-sharing units), Microsoft might also believe that it stands more to gain by letting AI die and assuming its role in the AI ecosystem. 

Embrace. Extend. Extinguish. 

But let’s assume it converts, and OpenAI now…has to continue raising money at a rate that will require it, allegedly, to only need to raise $17 billion in 2027. 

That number doesn’t make sense, considering it already had to bring forward its $8.3 billion fundraise by at least three months, but let’s stick with that idea. OpenAI believes it will be profitable, somehow, by 2030, and even if we assume that, that means it intends to burn over a hundred billion dollars to get there.

Is the plan to take OpenAI public, dumping a toxic asset onto the public markets, only to let it flounder and convulse and die for all to see? Can you imagine OpenAI’s S-1? How well do you think this company would handle a true financial audit from a major accounting firm? 

If you want to know what that looks like, google “WeWork,” which went from tech industry darling to joke in a matter of days, in part because it was forced to disclose how bad things actually were on its S-1. No, really, read this article.

With that in mind, I feel similarly about Anthropic. Nobody is buying this company at $170 billion, and thus the only way to access liquidity would be to take it public, and show the world how a company that made $72 million in January 2025 and then more than $400 million in July 2025 also loses $3 billion or more after revenue, and then let the market decide on its fair price. 

No, Really, What’s The Plan?

The arguments against my work always come down to “costs will go down” and “these products will become essential.” Outside of ChatGPT, there’s really no proof that these products are anything remotely essential, and I argue there’s very little about ChatGPT that Microsoft couldn’t provide with rate limits via Copilot. 

I’d also argue that “essential” is a very subjective term. Essential — in the sense that some people use it as search — doesn’t mean that it’s useful for enterprises, or the majority of people. 

And, I guess, ChatGPT somehow makes $1 billion a month in revenue selling access to premium versions of ChatGPT — though I’m not 100% sure how. Assuming it has 20 million customers paying $20 a month, that’s $400 million a month, then 5 million business customers paid an average of $100 each, that’s $900 million…and is that average really that good? Are that many people paying $35 a month, or $50, or $200? OpenAI doesn’t break out the actual revenues behind these numbers for a reason, and I believe that reason is “they don’t look as good.”

What’s OpenAI’s churn like? And does it really, as I wrote last week, end the year making more than Spotify at $1.5 billion a month? 

We don’t know, and OpenAI (much like Anthropic) has never shared actual revenues, choosing instead to leak to the media and hope to obfuscate the actual amounts of money being spent on its services. 

Anyway, long story short, these companies are unprofitable with no end in sight, don’t even make that much money in most cases, are valued more than anybody would ever buy them for, do not have much in the way of valuable intellectual property, and the two biggest players burn billions of dollars more than they make.

But Ed! The Government Will Give Them Money Forever!

Even if this were going to happen — it will not! — who would they give the money to and for how long? Would they give it to all the startups? Is every startup going to get a Paycheck Protection Program but for generative AI? How would that play out in rural red districts (where big tech has never been popular), which are being hit with both massive cuts to welfare, as well as the shockwaves of a trade war that has made American agricultural exports (like feedstocks, which previously went to China by the shipload) less appealing worldwide? 

So they bail out OpenAI, then stuff it full of government contracts to the tune of $15 billion a year, right? Sorry, just to be clear, that’s the low end of what this would take to do, and they’ll have to keep doing it forever, until Sam Altman can build enough data centers to…keep burning billions, because there’s no actual plan to make this profitable. 

Say this happens. Now what? America has a bullshit generative AI company attached to the state that doesn’t really innovate and doesn’t really matter in any meaningful way, except that it owns a bunch of data centers? 

I don’t think this happens! I think this is a silly idea, and the most likely situation would be that Microsoft would unhinge its jaw and swallow OpenAI and its customers whole. Hey, did you know that Microsoft’s data center construction is down year-over-year, and it’s basically signed no new data center leases? I wonder why it isn’t building these new data centers for OpenAI? Who knows. 

Stargate isn’t saving it, either. As I wrote previously, Stargate doesn’t actually exist beyond the media hype it generated.

And yes, OpenAI is offering ChatGPT at $1 for a year to US government workers - and I cannot express how little this means other than that they are horribly desperate. This product doesn't do enough to make it essential, and this fire sale doesn't change anything.

Anyway, does the government do this for everybody? Because everyone else is gonna need it as none of these companies can go public as they all suffer from the burden of generative AI. And, if the government does it, will it also subsidize the compute of for-profit companies like Cursor? To what end? Where is the limit?

What If Generative AI Just Isn’t Profitable? 

I think this is a question that we have to seriously consider at this point, because its ramifications are significant.

If I’m honest, I think the future of LLMs will be client-side on egregiously-expensive personal setups for enthusiasts, and in a handful of niche enterprise roles. Large Language Models do not scale profitably, and their functionality is not significant enough to justify the costs of running them. By immediately applying old economics — the idea that you would pay a monthly fee to have relatively-unlimited access — companies like OpenAI and Anthropic immediately trained users to use their products in a way that was antithetical to their costs. 

Then again, had these models been served in a way that was mindful of their costs, there would likely have been no way to even get this far. If OpenAI is making a billion dollars a month, it is possibly losing that much (or more) after revenue, and that’s the money it can get selling the product in a form that can never turn profitable. If OpenAI charged in line with its actual costs, would it even be able to justify a freely-available version of ChatGPT, outside of a few free requests? 

The revenue you see today is what people are willing to pay for a product that loses money, and I cannot imagine they would pay as much if the companies in question charged their costs. If I’m wrong, Cursor will be just fine, and that’s assuming that Cursor’s current hobbled form is even profitable, which it has not said it is.

So, you’ve got an entire industry of companies that struggle to do anything other than lose a lot of money. Great. 

And now we have a massive expansive data centre buildout, the likes of which we’ve never seen, all to capture demand for a product that nobody makes much money selling.

This, naturally, leads to an important question: how do these people building data centers actually make money?

Data Center Developers Aren’t Making Money On AI Either, And Big Tech’s Capex Is The Majority Of “AI Capex”

Last week, the Wall Street Journal published one of the more worrying facts I’ve seen in the last two years:

Investor and tech pundit Paul Kedrosky says that, as a percentage of gross domestic product, spending on AI infrastructure has already exceeded spending on telecom and internet infrastructure from the dot-com boom—and it’s still growing. He also argues that one explanation for the U.S. economy’s ongoing strength, despite tariffs, is that spending on IT infrastructure is so big that it’s acting as a sort of private-sector stimulus program.…Capex spending for AI contributed more to growth in the U.S. economy in the past two quarters than all of consumer spending, says Neil Dutta, head of economic research at Renaissance Macro Research, citing data from the Bureau of Economic Analysis.

A global accounting of this infrastructure spending would be even bigger, as it would include capex from these companies’ most important partners. Foxconn has recently spent big building out factories for Apple in India, which just supplanted China as the source of the majority of U.S.-destined iPhones, according to Canalys. And the world’s largest chip manufacturer, TSMC, spent about $10 billion on capex in its most recent quarter.

The massive buildout of data centers — and the associated physical gear like chips, servers, and raw materials for building them — has become a massive, dominant economic force…building capacity for an industry that is yet to prove it can make real revenues. 

And no, Microsoft talking about its Azure revenue in its last quarterly earnings for the first time is not the same thing, as it stopped explicitly stating their AI revenue in January (when it was $13 billion annualized).

Anyway, AI capex allegedly — though I have some questions about this figure! — accounts for 1.2% of the US GDP in the first half of the year, and accounted for more than half of the (to quote the Wall Street Journal) “already-sluggish” 1.2% growth rate of the US economy. 

Another Wall Street Journal piece published a few days later discussed how data center development is souring the free cash flow for big tech, turning them from the kind of “asset-light” businesses that the markets love into entities burdened by physical real estate and their associated costs:

For years, investors loved those companies because they were “asset-light.” They earned their profits on intangible assets such as intellectual property, software, and digital platforms with “network effects.” Users flocked to Facebook, Google, the iPhone, and Windows because other users did. Adding revenue required little in the way of more buildings and equipment, making them cash-generating machines.

This can be seen in a metric called free cash flow, roughly defined as cash flow from operations minus capital expenditures. It excludes things such as noncash impairment charges that can distort net income. This is arguably the purest measure of a business’s underlying cash-generating potential. Amazon, for example, tells investors: “Our financial focus is on long-term, sustainable growth in free cash flow.”

…

From 2016 through 2023, free cash flow and net earnings of Alphabet, Amazon, Meta and Microsoft grew roughly in tandem. But since 2023, the two have diverged. The four companies’ combined net income is up 73%, to $91 billion, in the second quarter from two years earlier, while free cash flow is down 30% to $40 billion, according to FactSet data. Apple, a relative piker on capital spending, has also seen free cash flow lag behind.

These numbers are all very scary, and I mean that sincerely, but they also fail to express why. How much was actually spent on AI capex in the US? One would think two different articles on this subject would include that number versus a single quarter’s worth, but from my estimates, I expect capital expenditures from the Magnificent Seven alone to crest $200 billion in the first half of 2025, with Axios estimating they’d spend around $400 billion this year. 

Most articles are drafting off of a blog from Paul Kedrosky, who estimates total AI capex would be somewhere in the region of $520 billion in total for the year, which felt conservative to me, so I did the smart thing and asked him. Kedrosky noted that these numbers focus entirely on the four big spenders — Microsoft, Google, Meta and Amazon, and his own estimated $312 billion capex, and the 1.2% number came from the assumption that the US GDP in 2025 will be around $28 trillion (which, I add, is significantly lower than other forecasts, which puts it closer to $30 trillion). 

Kedrosky, in his own words, was trying to be conservative, using public data and then building his analysis from there. I, personally, believe his estimate is too conservative — because it doesn’t factor in the capital expenditures from Oracle, which (along with Crusoe) is building the vast Abilene Texas data center for OpenAI, or any private data center developers sinking cash into AI capex. 

When I asked him to elaborate, he estimated that “...AI spend, all-in, was around half of 3.0%  Q2 real GDP growth, so 2-3x the lower bound, given multipliers, debt, etc. it could be half of US GDP full-year GDP growth.”

That’s so cool! Half of the US economy’s growth came from building data centers for generative AI, which has the combined revenue of a little more than the fucking smart watch industry in 2024.  

Another troubling point is that big tech doesn’t just buy data centers and then use them, but in many cases pays a construction company to build them, fills them with GPUs and then leases them from a company that runs them, meaning that they don’t have to personally staff up and maintain them. This creates an economic boom for construction companies in the short term, as well as lucrative contracts for ongoing support…as long as the company in question still wants them. While Microsoft or Amazon might use a data center and, indeed, act as if it owns it, ultimately somebody else is holding the bag and the ultimate responsibility for the data centers.

One such company is QTS, a data center developer that leases to both Amazon and Meta according to the New York Times, which was acquired by Blackstone in 2021 for $10 billion. Since then, Blackstone has used commercial mortgage-backed securities — I know! — to raise over $8.7 billion since then to sink into QTS’ expansion, and as of mid-July said it’d be investing $25 billion in AI data centers and energy.

Blackstone, according to the New York Times, sees “strong demand from tech companies,” who are apparently “willing to sign what they describe as airtight leases for 15 to 20 years to rent out data center space.” 

Yet the Times also names another problem — the “unanswered question” of how these private equity firms actually exit these situations. Blackstone, KKR and other asset management firms do not buy companies with the intention of syphoning off revenue, but to pump them up and sell them to another company. Much like AI startups, it isn’t obvious who would buy QTS at what I imagine would be a $25 billion or $30 billion valuation, meaning that Blackstone would have to take them public. Similarly, KKR’s supposed $50 billion partnership with investment firm Energy Capital partners to build data centers and their associated utilities does not appear to have much of an exit plan either.

And let’s not forget Oracle, OpenAI, and Crusoe’s abominable mess in Abilene Texas, where Oracle is paying for the $40 billion of GPUs and Crusoe is spending $15 billion raised from Blue Owl Capital and Primary Digital infrastructure to build data centers for OpenAI, a company that loses billions of dollars a year. Why? So that OpenAI can, allegedly starting in 2028, pay Oracle $30 billion a year for compute, and yes, I am being fully serious. 

To be clear, OpenAI, by my estimates, has only made around $5.26 billion this year (and will have trouble hitting its $12.7 billion revenue projection for 2025), and will likely lose more than $10 billion to do so.

Oracle will, according to The Information, owe Crusoe $1 billion in payments across the 15 year span of its lease. How does Crusoe afford to pay back its $15 billion in loans? Beats me! The Information says it’s raising $1 billion to “take on cloud giants” by “earning construction management fees and rent, and it can sell its stake in the project upon reaching certain completion milestones,” while also building its own AI compute, making the assumption that the demand is there outside of hyperscalers.

Then there’s CoreWeave, my least-favourite company in the world. As I discussed a few months ago, CoreWeave is burdened by obscene debt and a horrifying cash burn, and has seen its stock spike to a high of $183 on June 20, 2025 to around $111 as of writing this sentence, which has led to its all-stock attempt to acquire developer Core Scientific for $9 billion to start to fall apart as shareholders balk at the worrisome drop in CoreWeave’s stock price. CoreWeave has, since going public, had to borrow billions of dollars to fund its obscene capital expenditures to handle the upcoming October 2025 start date for OpenAI’s $11.9 billion, 5-year-long deal for compute, which is also when CoreWeave must start paying off its largest loan. CoreWeave lost $314 million in its last earnings, and I see no path to profitability or, honestly, its ability to keep doing business if the market sours.

Coreweave, I add, is pretty much reliant on Microsoft as its primary customer. While this relationship has been fairly smooth (so far, and as far as we know), this dependence also presents an existential threat to Coreweave, and is part of the reason why I’m so pessimistic about its survival. Microsoft has its own infrastructure, and has every incentive to cut out middlemen when it's able to meet supply with the demand it itself owns (or leases, rather than subcontracts out), simply because middlemen add costs and shrink margins. If Microsoft walks, what’s left? How does it service its ongoing obligations, and its mountain of debt? 

In all of these cases, data center developers seem to have very few options as to making actual money. We have companies spending billions of dollars to vastly expand their data center footprint, but very little evidence that doing so results in revenue let alone some sort of payoff, and similarly, the actual capital expenditures they’re making are…much smaller than those of big tech. 

Digital Realty Trust — one of the largest developers with over 300 data centers worldwide and $5.55 billion in revenue in 2024 — only spent $3.5 billion in capex last quarter, and Equinix ($8.7 billion revenue in 2024), which has 270 of them, put capex at $3.5 billion too. NTT Global Data Centers, which has over 160 data centers, has dedicated $10 billion in capital expenditures “through 2027” to build out data centers. 

Yet in many of these cases, it’s because these companies are — to quote a source of mine — “functionally obsolete for this cycle,” because legacy data centers are not plug-and-play ready for GPUs to slot into. Any investment in capex by these companies would have to be for both GPUs and either building or retrofitting (basically ripping the insides out of old) data centers.

This means that the money flowing into AI data centers is predominantly going to neoclouds like CoreWeave and Crusoe, and all seems to flow back to private equity firms that never thought about where the cashout might be. Blackstone led CoreWeave’s $7.5 billion loan with Magnetar Capital, and Crusoe signed a deal a week ago with infrastructure firm Blackstone-owned Tallgrass to build a data center in Wyoming, all of which seems very good for Blackstone unless you think “how does it actually make money here,” as private equity firms do not generally like to hold assets longer than five years. 

Even if it did, its capital expenditures are a drop in the bucket in the grand scheme of things. Assuming Crusoe burns, as The Information suggests it will, as much as $4 billion in 2025, CoreWeave spends as much as $20 billion, Digital Realty Trust spends $14 billion, Global Data Centers spends $3.33 billion (that’s $10bn over 3 years), and Equinix spends $14 billion. That’s $55.33 billion in AI capex spent in 2025 from the largest developers of data centers in the world.

For some context, as discussed above, $102 billion was spent by Meta, Alphabet, Microsoft and Amazon in the last quarter. 

Private equity may ultimately face the same problem as many AI startups: there is no clear exit strategy for these investments. In the absence of real liquidity, firms will likely resort to all manner of financial engineering (read: bullshit) —marking up portfolio companies using internally generated valuations, charging fees on those inflated marks, and using those marks to entice new commitments from limited partners.

Compounding this is their ability to lend increasing amounts of capital to their own portfolio companies via affiliated private credit vehicles—effectively recycling capital and pushing valuation risk further down the line. This kind of self-reinforcing leverage loop is particularly opaque in private credit, which now underpins much of the AI infrastructure buildout. The complexity of these arrangements makes it hard to anticipate the full economic fallout if the cycle breaks down, but the systemic risk is building.

In any case, the supposed “AI capex boom” that is driving the US economy is not, as reported, driven by the massive interest in building out AI infrastructure for a variety of customers. 

The reality is simple: the majority of all AI capex is from big tech, which is a massive systemic weakness in our economy.

Big Tech Has Become A Systemic Weakness In The US Economy (and Markets)

While some might say that “AI capex” has swallowed the US economy, I think it’s more appropriate to say that Big Tech Capex Has Swallowed The US Economy.

I also want to be clear that the economy — which is the overall state of the country’s production and consumption of stuff, and the flow of money between participants in said economy — and the markets (as in the stock market) are very different things, but the calculations from Kedrosky and others have now allowed us to see where one might hit the other.

You see, the markets do not actually represent reality. While Microsoft, Amazon, Google, and Meta might want you to think there’s a ton of money in AI, their growth is mostly from selling further iterations and contracts for their existing stuff, or in the case of Meta further increasing its ad revenue. The economy is where things are actually bought and sold, representing the economic effects of both the things happening to build out AI and selling access to services and the AI models themselves. I recognize this is simplistic, but I am laying it out for a reason.

As I discussed at length in the Hater’s Guide to the AI Bubble, NVIDIA is the weak point in the stock market, representing roughly 19% of the value of the Magnificent 7, which in turn makes up about 35% of the value of the US stock market. The associated Magnificent Seven stocks have seen a huge boom through their own growth, which has been mistakenly and incorrectly attributed to revenue from AI, which as I laid out previously, is about $35 to $40 billion in the last two years. Nevertheless, the markets can continue to be irrational because all they care about is “number going up,” as the “value” of a stock is oftentimes disconnected from the value of the company itself, instead associated with its propensity for growth.

GDP and other measurements of the economy aren’t really something you can fudge quite as easily (at least, in transparent, democratic societies), nor can you say a bunch of fancy words to make people feel better in the event that growth stalls or declines. 

This leads me to my principle worry: that “AI capex” is actually a term for the expenditures of four companies, namely Microsoft, Amazon, Google and Meta, with NVIDIA’s GPU sales being part of that capex too.

While we can include others like Oracle, Musk’s xAI, and various Neoclouds like CoreWeave and Crusoe — who, according to D.A. Davidson’s Gil Luria, will account for about 10% of NVIDIA’s GPU sales in 2025 — the reality is that whatever economic force is being driven by “AI investment” is really just four companies building and leasing data centers to burn on generative AI, a product that makes a relatively small amount of money before losing a great deal more. 

42% of NVIDIA’s revenue comes from the Magnificent Seven (per Laura Bratton at Yahoo Finance), which naturally means that big tech is the lynchpin of investment in data centers. 

I’ll put it far more simply: if AI capex represents such a large part of our GDP and economic growth, our economy does, on some level, rest on the back of Microsoft, Google, Meta and Amazon and their continued investment in AI. What should worry everybody is that Microsoft — which makes up 18.9% of NVIDIA’s revenue — has signed basically no leases in the last 12 months, and its committed datacenter construction and land purchases are down year-over-year. 

While its capex may not have dipped yet (in part because the chip-heavy nature of generative AI means that capex isn’t exclusively dominated by property), it’s now obvious that if it does there will be direct effects on both the US economy and stock market, as Microsoft is part of what amounts to a stimulus package propping up America’s economic growth. 

And not to repeat the point too much, but big tech has yet to actually turn anything resembling a profit on these data centers, and isn’t making much revenue at all out of generative AI.

How, exactly, does this end? What is the plan here? Is big tech going to spend hundreds of billions a year in capital expenditures on generative AI in perpetuity? Will they continue to buy more and more NVIDIA chips as they do so? 

At some point, surely these companies have built enough data centers? Surely, at some point, they’ll run out of space to put these GPUs in? Is the plan to, by then, make so much money from AI that it won’t matter? What does NVIDIA do at that point? And how does the US economy rebound from the loss of activity that follows?

As I’ve said again and again, the generative AI bubble is, and always has been, fundamentally irrational, and inherently gothic, playing in the ruins, patterns and pathways of previous tech booms despite this one having little or no resemblance to them. Though the tech industry loves to talk about building a glorious future, its present is one steeped in rituals of decay and death, where the virtues of value creation and productivity take a backseat to burning billions and lying to the public again and again and again. The way in which the media has participated in these lies is disgusting.

Venture capital, still drunk off the fumes of 2021, keeps running the old playbook: shove as much money into a company as possible in the hopes you can dump it onto an acquirer or the public markets, only to get high on their own supply, pushing valuations to the point that there is no possible liquidity event for the majority of big private AI companies as a result of their overstuffed valuations, burdensome business models and lack of any real intellectual property.

And, like the rest of the AI bubble, Silicon Valley’s only liquidity path out of the bubble is big tech itself. Without Google, Character.ai and Windsurf’s founders would likely have been left for dead, and the same goes for Inflection, and I’d even argue Scale AI, whose $14.3 billion “investment” from Meta effectively decapitated the company, removing its CEO Alexandr Wang, leaving the rest of the company to die, laying off 14% of its staff and 500 contractors mere weeks after its CEO and investors cashed in. 

In fact, generative AI is turning out to be a fever dream entirely made up by big tech. OpenAI would be dead if it wasn’t for the massive infrastructure provided by Microsoft at-cost in return for rights to its IP, research, and the ability to sell its models on top of the tens of billions of dollars of venture capital thrown into its billion-dollar cash incinerator. Anthropic would be dead if both Google and Amazon — the latter of which provides much of its infrastructure — hadn’t invested billions in keeping it alive so that it can burn $3 billion or more in 2025 while fucking over its enterprise customers and rate limiting the rest. 

The generative AI industry is, at its core, unnatural. It does not make significant revenue compared to its unbelievable costs, nor does it have much revenue potential. It requires, unlike just about every software revolution, an unbelievable amount of physical infrastructure to run, and because nobody but big tech can afford to build the infrastructure necessary, creates very little opportunity for competition or efficiency. As the markets are in the throes of the growth-at-all-costs Rot Economy, they have failed to keep big tech in line, conflating big tech’s ability to grow with growth driven as a result of their capital expenditures. Sensible, reasonable markets would notice the decay of free cash flow or the ridiculousness of big tech’s capex bonanza, but instead they clap and squeal every time Satya Nadella jingles his keys.

What is missing is any real value generation. Again, I tell you, put aside any feelings you may have about generative AI itself, and focus on the actual economic results of this bubble. How much revenue is there? Why is there no profit? Why are there no exits? Why does big tech, which has sunk hundreds of billions of dollars into generative AI, not talk about the revenues they’re making? Why, for three years straight, have we been asked to “just wait and see,” and for how long are we going to have to wait to see it?

What’s incredible is that the inherently compute-intensive nature of generative AI basically requires the construction of these facilities, without actually representing whether they are contributing to the revenues of the companies that operate the models (like Anthropic or OpenAI, or any other business that builds upon them). As the models get more complex and hungry, more data centers get built — which hyperscalers book as long-term revenue, even though it’s either subsidised by said hyperscalers, or funded by VC money. This, in turn, stimulates even more capex spending. And without having to answer any basic questions about longevity or market fit. 

Yet the worst part of this financial farce is that we’ve now got a built-in economic breaking point in the capex from AI. At some point capex has to slow — if not because of the lack of revenues or massive costs associated, but because we live in a world with finite space, and when said capex slow happens, so will purchases of NVIDIA GPUs, which will in turn, as proven by Kedrosky and others, slow America’s economic growth.

And that growth is pretty much based on the whims of four companies, which is an incredibly risky and scary proposition. I haven’t even dug into the wealth of private credit deals that underpin buildouts for private AI “neoclouds” like CoreWeave, Crusoe, Nebius, and Lambda, in part because their economic significance is so much smaller than big tech’s ugly, meaningless sprawl. 

To quote Kedrosky: 

We are in a historically anomalous moment. Regardless of what one thinks about the merits of AI or explosive datacenter expansion, the scale and pace of capital deployment into a rapidly depreciating technology is remarkable. These are not railroads—we aren’t building century-long infrastructure. AI datacenters are short-lived, asset-intensive facilities riding declining-cost technology curves, requiring frequent hardware replacement to preserve margins.

You can’t bail this out, because there is nothing to bail out. Microsoft, Meta, Amazon and Google have plenty of money and have proven they can spend it. NVIDIA is already doing everything it can to justify people spending more on its GPUs. There’s little more it can do here other than soak up the growth before the party ends. 

That capex reduction will bring with it a reduction in expenditures on NVIDIA GPUs, which will take a chunk out of the US stock market. Although the stock market isn’t the economy, the two things are inherently linked, and the popping of the AI bubble will have downstream ramifications, just like the dot com bubble did on the wider economy.

Expect to see an acceleration in layoffs and offshoring, in part driven by a need for tech companies to show — for the first time in living memory — fiscal restraint. For cities where tech is a major sector of the economy — think Seattle and San Francisco — there’ll be knock-on effects to those companies and individuals that support the tech sector (like restaurants, construction companies building apartments, Uber drivers, and so on). We’ll see a drying-up of VC funding. Pension funds will take a hit — which will affect how much people have to spend in retirement. It’ll be grim. 

Worse than that is the fact that these data centers will be, by definition, non-performing assets, and one that inflicted an opportunity cost that’ll be almost impossible to calculate. While a house, once built and sold, technically falls into that category (it doesn’t add to any economic productivity), people at least need somewhere to live. Shelter is an essential component of life. You can live without a data center the size of Manhattan. 

What would have happened if companies like Microsoft and Meta instead spent the money on things that actually drove productivity, or created a valuable competitive business that drove economic activity? Hell, even if they just gave everyone a 10% raise, it would have likely been better for the economy than this, if we’re factoring in things like consumer spending. 

It’s just waste. Profligate, pointless waste. 

In summary, we’re already facing the prospect of a recession, and though I am not an economist, I can imagine it being a particularly nasty one given that the Magnificent Seven accounted for 47.87% of the Russell 1000 Index’s returns in 2024. Even if big tech somehow makes this crap profitable, it’s hard to imagine that they’ll counterbalance any capex reduction with revenue, because there doesn’t seem to be that much revenue in generative AI to begin with.

This is what happens when you allow the Rot Economy to run wild, building the stock market and tech industry on growth over everything else. This is what happens when the tech media repeatedly fails to hold the powerful to account, catering to their narratives and making excuses for their abominable, billion-dollar losses and mediocre, questionably-useful products. 

Waffle on all you want about the so-called “agentic era” or “annualized revenues” that make you hot under the collar — I see no reason for celebration about an industry with no exit plans and needless capital expenditures that appear to be one of the few things keeping the American economy growing. 

I have been writing about the tech industry’s obsession with generative AI for two years, and never have I felt more grim. Before this was an economic uncertainty — a way that our markets might contract, that big tech might take a big haircut, that a bunch of money might be wasted but otherwise the world would keep turning. 

It feels as if everything is aligning for disaster, and I fear there’s nothing that can be done to avert it.