Hey! Before we go any further — if you want to support my work, please sign up for the premium version of Where’s Your Ed At, it’s a $7-a-month (or $70-a-year) paid product where every week you get a premium newsletter, all while supporting my free work too.
Also, subscribe to my podcast Better Offline, which is free. Go and subscribe then download every single episode.
One last thing: This newsletter is nearly 14,500 words. It’s long. Perhaps consider making a pot of coffee before you start reading.
Good journalism is making sure that history is actively captured and appropriately described and assessed, and it’s accurate to describe things as they currently are as alarming.
And I am alarmed.
Alarm is not a state of weakness, or belligerence, or myopia. My concern does not dull my vision, even though it’s convenient to frame it as somehow alarmist, like I have some hidden agenda or bias toward doom. I profoundly dislike the financial waste, the environmental destruction, and, fundamentally, I dislike the attempt to gaslight people into swearing fealty to a sickly and frail psuedo-industry where everybody but NVIDIA and consultancies lose money.
I also dislike the fact that I, and others like me, are held to a remarkably different standard to those who paint themselves as “optimists,” which typically means “people that agree with what the market wishes were true.” Critics are continually badgered, prodded, poked, mocked, and jeered at for not automatically aligning with the idea that generative AI will be this massive industry, constantly having to prove themselves, as if somehow there’s something malevolent or craven about criticism, that critics “do this for clicks” or “to be a contrarian.”
I don’t do anything for clicks. I don’t have any stocks or short positions. My agenda is simple: I like writing, it comes to me naturally, I have a podcast, and it is, on some level, my job to try and understand what the tech industry is doing on a day-to-day basis. It is easy to try and dismiss what I say as going against the grain because “AI is big,” but I’ve been railing against bullshit bubbles since 2021 — the anti-remote work push (and the people behind it), the Clubhouse and audio social networks bubble, the NFT bubble, the made-up quiet quitting panic, and I even, though not as clearly as I wished, called that something was up with FTX several months before it imploded.
This isn’t “contrarianism.” It’s the kind of skepticism of power and capital that’s necessary to meet these moments, and if it’s necessary to dismiss my work because it makes you feel icky inside, get a therapist or see a priest.
Nevertheless, I am alarmed, and while I have said some of these things separately, based on recent developments, I think it’s necessary to say why.
In short, I believe the AI bubble is deeply unstable, built on vibes and blind faith, and when I say “the AI bubble,” I mean the entirety of the AI trade.
And it’s alarmingly simple, too.
But this isn’t going to be saccharine, or whiny, or simply worrisome. I think at this point it’s become a little ridiculous to not see that we’re in a bubble. We’re in a god damn bubble, it is so obvious we’re in a bubble, it’s been so obvious we’re in a bubble, a bubble that seems strong but is actually very weak, with a central point of failure.
I may not be a contrarian, but I am a hater. I hate the waste, the loss, the destruction, the theft, the damage to our planet and the sheer excitement that some executives and writers have that workers may be replaced by AI — and the bald-faced fucking lie that it’s happening, and that generative AI is capable of doing so.
And so I present to you — the Hater’s Guide to the AI bubble, a comprehensive rundown of arguments I have against the current AI boom’s existence. Send it to your friends, your loved ones, or print it out and eat it.
No, this isn’t gonna be a traditional guide, but something you can look at and say “oh that’s why the AI bubble is so bad.” And at this point, I know I’m tired of being gaslit by guys in gingham shirts who desperately want to curry favour with other guys in gingham shirts but who also have PHDs. I’m tired of reading people talk about how we’re “in the era of agents” that don’t fucking work and will never fucking work. I’m tired of hearing about “powerful AI” that is actually crap, and I’m tired of being told the future is here while having the world’s least-useful most-expensive cloud software shoved down my throat.
Look, the generative AI boom is a mirage, it hasn’t got the revenue or the returns or the product efficacy for it to matter, everything you’re seeing is ridiculous and wasteful, and when it all goes tits up I want you to remember that I wrote this and tried to say something.
The Magnificent 7’s Weakpoint: NVIDIA
As I write this, NVIDIA is currently sitting at $170 a share — a dramatic reversal of fate after the pummelling it took from the DeepSeek situation in January, which sent it tumbling to a brief late-April trip below $100 before things turned around.
The Magnificent 7 stocks — NVIDIA, Microsoft, Alphabet (Google), Apple, Meta, Tesla and Amazon — make up around 35% of the value of the US stock market, and of that, NVIDIA’s market value makes up about 19% of the Magnificent 7. This dominance is also why ordinary people ought to be deeply concerned about the AI bubble. The Magnificent 7 is almost certainly a big part of their retirement plans, even if they’re not directly invested.
Back in May, Yahoo Finance’s Laura Bratton reported that Microsoft (18.9%), Amazon (7.5%), Meta (9.3%), Alphabet (5.6%), and Tesla (0.9%) alone make up 42.4% of NVIDIA’s revenue. The breakdown makes things worse. Meta spends 25% — and Microsoft an alarming 47% — of its capital expenditures on NVIDIA chips, and as Bratton notes, Microsoft also spends money renting servers from CoreWeave, which analyst Gil Luria of D.A.Davidson estimates accounted for $8 billion (more than 6%) of NVIDIA’s revenue in 2024. Luria also estimates that neocloud companies like CoreWeave and Crusoe — that exist only to prove AI compute services — account for as much as 10% of NVIDIA’s revenue.
NVIDIA’s climbing stock value comes from its continued revenue growth. In the last four quarters, NVIDIA has seen year-over-year growth of 101%, 94%, 78% and 69%, and, in the last quarter, a little statistic was carefully brushed under the rug: that NVIDIA missed, though narrowly, on data center revenue. This is exactly what it sounds like — GPUs that are used in servers, rather than gaming consoles and PCs (. Analysts estimated it would make $39.4 billion from this category, and NVIDIA only (lol) brought in $39.1 billion. Then again, it could be attributed to its problems in China, especially as the H20 ban has only just been lifted. In any case, it was a miss!
NVIDIA’s quarter-over-quarter growth has also become aggressively normal — from 69%, to 59%, to 12%, to 12% again each quarter, which, again, isn’t bad (it’s pretty great!), but when 88% of your revenue is based on one particular line in your earnings, it’s a pretty big concern, at least for me. Look, I’m not a stock analyst, nor am I pretending to be one, so I am keeping this simple:
NVIDIA relies not only on selling lots of GPUs each quarter, but **it must always, always sell more GPUs the next quarter.**42% of NVIDIA’s revenue comes from Microsoft, Amazon, Meta, Alphabet and Tesla continuing to buy **more GPUs.**NVIDIA’s value and continued growth is heavily reliant on hyperscaler purchases and **continued interest in generative AI.**The US stock market’s continued health relies, on some level, on five or six companies (it’s unclear how much Apple buys GPU-wise) spending billions of dollars on GPUs from NVIDIA. An analysis from portfolio manager Danke Wang from January found that the Magnificent 7 stocks accounted for 47.87% of the Russell 1000 Index’s returns in 2024 (an index fund of the 1000 highest-ranked stocks on FTSE Russell’s index).
In simpler terms, 35% of the US stock market is held up by five or six companies buying GPUs. If NVIDIA’s growth story stumbles, it will reverberate through the rest of the Magnificent 7, making them rely on their own AI trade stories.
And, as you will shortly find out, there is no AI trade, because generative AI is not making anybody any money.
The Hollow “AI Trade”
I’m so tired of people telling me that companies are “making tons of money on AI.” Nobody is making a profit on generative AI other than NVIDIA. No, really, I’m serious.
The Magnificent 7’s AI Story Is Flawed, With $560 Billion of Capex between 2024 and 2025 Leading to $35 billion of Revenue, And No Profit
If they keep their promises, by the end of 2025, Meta, Amazon, Microsoft, Google and Tesla will have spent over $560 billion in capital expenditures on AI in the last two years, all to make around $35 billion.
This is egregiously fucking stupid.
Microsoft AI Revenue In 2025: $13 billion, with $10 billion from OpenAI, sold “at a heavily discounted rate that essentially only covers costs for operating the servers.”
Capital Expenditures in 2025: $80 billion
As of January 2025, Microsoft’s “annualized” — meaning [best month]x12 — revenue from artificial intelligence was around $13 billion, a number that it chose not to update in its last earnings, likely because it’s either flat or not growing, though it could in its upcoming late-July earnings. Yet the problem with this revenue is that $10 billion of that revenue, according to The Information, comes from OpenAI’s spend on Microsoft’s Azure cloud, and Microsoft offers preferential pricing — “a heavily discounted rental rate that essentially only covers Microsoft’s costs for operating the servers” according to The Information.
In simpler terms, 76.9% of Microsoft’s AI revenue comes from OpenAI, and is sold at just above or at cost, making Microsoft’s “real” AI revenue about $3 billion, or around 3.75% of this year’s capital expenditures, or 16.25% if you count OpenAI’s revenue, which costs Microsoft more money than it earns.
The Information reports that Microsoft made $4.7 billion in “AI revenue” in 2024, of which OpenAI accounted for $2 billion, meaning that for the $135.7 billion that Microsoft has spent in the last two years on AI infrastructure, it has made $17.7 billion, of which OpenAI accounted for $12.7 billion.
Amazon AI Revenue In 2025: $5 billion
Capital Expenditures in 2025: $105 billion
Things do not improve elsewhere. An analyst estimates that Amazon, which plans to spend $105 billion in capital expenditures this year, will make $5 billion on AI in 2025, rising, and I quote, “as much as 80%,” suggesting that Amazon may have made a measly $1 billion in 2024 on AI in a year when it spent $83 billion in capital expenditures.
Last year, Amazon CEO Andy Jassy said that “AI represents for sure the biggest opportunity since cloud and probably the biggest technology shift and opportunity in business since the internet." I think he’s full of shit.
Google AI Revenue: $7.7 Billion (at most)
Capital Expenditures in 2025: $75 Billion
Bank of America analyst Justin Post estimated a few weeks ago that Google’s AI revenue would be in the region of $7.7 billion, though his math is, if I’m honest, a little generous:
Google’s artificial intelligence model is set to drive $4.2 billion in subscription revenue within its Google Cloud segment in 2025, according to an analysis from Bank of America last week.That includes $3.1 billion in revenue from subscribers to Google’s AI plans with its Google One service, Bank of America’s Justin Post estimates.Post also expects that the integration of Google’s Gemini AI features within its Workspace service will drive $1.1 billion of the $7.7 billion in revenue he projects for that segment in 2025.
Google’s “One” subscription includes increased cloud storage across Google Drive, Gmail and Google Photos, and added a $20-a-month “premium” plan in February 2024 that included access to Google’s various AI models. Google has claimed that the “premium AI tier accounts for millions” of the 150 million subscribers to the service, though how many millions is impossible to estimate — but that won’t stop me trying!
Assuming that $3.1 billion in 2025 revenue would work out to $258 million a month, that would mean there were 12.9 million Google One subscribers also paying for the premium AI tier. This isn’t out of the realm of possibility — after all, OpenAI has 15.5 million paying subscribers — but Post is making a generous assumption here. Nevertheless, we’ll accept the numbers as they are.
And the numbers fuckin’ stink! Google’s $1.1 billion in workspace service revenue came from a forced price-hike on those who use Google services to run their businesses, meaning that this is likely not a number that can significantly increase without punishing them further.
$7.7 billion of revenue — not profit! — on $75 billion of capital expenditures. Nasty!
Meta AI Revenue: $2bn to $3bn
Capital Expenditures In 2025: $72 Billion
Someone’s gonna get mad at me for saying this, but I believe that Meta is simply burning cash on generative AI. There is no product that Meta sells that monetizes Large Language Models, but every Meta product now has them shoved into them, such as your Instagram DMs oinking at you to generate artwork based on your conversation.
Nevertheless, we do have some sort of knowledge of what Meta is saying due to the copyright infringement case Kadrey v. Meta. Unsealed judgment briefs revealed in April that Meta is claiming that “GenAI-driven revenue will be more than $2 billion,” with estimates as high as $3 billion. The same document also claims that Meta expects to make $460 billion to $1.4 trillion in total revenue through 2035, the kind of thing that should get you fired in an iron ball into the sun.
Meta makes 99% of its revenue from advertising, and the unsealed documents state that it “[generates] revenue from [its] Llama models and will continue earning revenue from each iteration,” and “share a percentage of the revenue that [it generates] from users of the Llama models…hosted by those companies,” with the companies in question redacted. Max Zeff of TechCrunch adds that Meta lists host partners like AWS, NVIDIA, Databricks, Groq, Dell, Microsoft Azure, Google Cloud, and Snowflake, so it’s possible that Meta makes money from licensing to those companies. Sadly, the exhibits further discussing these numbers are filed under seal.
Either way, we are now at $332 billion of capital expenditures in 2025 for $28.7 billion of revenue, of which $10 billion is OpenAI’s “at-cost or just above cost” revenue. Not great.
Tesla Does Not Appear To Make Money From Generative AI
Capital Expenditures In 2025: $11 billion
Despite its prominence in the magnificent 7, Tesla is one of the least-exposed of the magnificent 7 to the AI trade, as Elon Musk has turned it into a meme stock company. That doesn’t mean, of course, that Musk isn’t touching AI. xAI, the company that develops racist Large Language Model “Grok” and owns what remains of Twitter, apparently burns $1 billion a month, and The Information reports that it makes a whopping $100 million in annualized revenue — so, about $8.33 million a month. There is a shareholder vote for Tesla to potentially invest in xAI, which will probably happen, allowing Musk to continue to pull leverage from his Tesla stock until the company’s decaying sales and brand eventually swallow him whole.
But we’re not talking about Elon Musk today.
Apple’s AI Story Is Weird
Capital Expenditures In 2025: around $11 billion
Apple Intelligence radicalized millions of people against AI, mostly because it fucking stank. Apple clearly got into AI reluctantly, and now faces stories about how they “fell behind in the AI race,” which mostly means that Apple aggressively introduced people to the features of generative AI by force, and it turns out that people don’t really want to summarize documents, write emails, or make “custom emoji,” and anyone who thinks they would is a fucking alien.
In any case, Apple hasn’t bet the farm on AI, insomuch as it hasn’t spent two hundred billion dollars on infrastructure for a product with a limited market that only loses money.
The Fragile Five — Amazon, Google, Microsoft, Meta and Tesla — Are Holding Up The US Stock Market By Funding NVIDIA’s Future Growth Story
To be clear, I am not saying that any of the Magnificent 7 are going to die — just that five companies’ spend on NVIDIA GPUs largely dictate how stable the US stock market will be. If any of these companies (but especially NVIDIA) sneeze, your 401k or your kid’s college fund will catch a cold.
I realize this sounds a little simplistic, but by my calculations, NVIDIA’s value underpins around 8% of the value of the US stock market. At the time of writing, it accounts for roughly 7.5% of the S&P 500 — an index of the 500 largest US publicly-traded companies. A disturbing 88% of Nvidia’s revenue comes from enterprise-scale GPUs primarily used for generative AI, of which five companies’ spend makes up 42% of its revenue. In the event that any one of these companies makes significant changes to their investments in NVIDIA chips, it will eventually have a direct and meaningful negative impact on the wider economy.
NVIDIA’s earnings are, effectively, the US stock market’s confidence, and everything rides on five companies — and if we’re honest, really four companies — buying GPUs for generative AI services or to train generative AI models. Worse still, these services, while losing these companies massive amounts of money, don’t produce much revenue, meaning that the AI trade is not driving any real, meaningful revenue growth.
But Ed, They Said Something About Points of Growth-
Silence!
Any of these companies talking about “growth from AI” or “the jobs that AI will replace” or “how AI has changed their organization” are hand-waving to avoid telling you how much money these services are actually making them. If they were making good money and experiencing real growth as a result of these services, they wouldn’t shut the fuck up about it! They’d be in your ear and up your ass hooting about how much cash they were rolling in!
And they’re not, because they aren’t rolling in cash, and are in fact blowing nearly a hundred billion dollars each to build massive, power-hungry, costly data centers for no real reason.
Don’t watch the mouth — watch the hands. These companies are going to say they’re seeing growth from AI, but unless they actually show you the growth and enumerate it, they are hand-waving.
Ed! Amazon Web Services Took Years To Become Profitable! People Said Amazon Would Fail!
So, one of the most annoying and consistent responses to my work is to say that either Amazon or Amazon Web Services “ran at a loss,” and that Amazon Web Services — the invention of modern cloud computing infrastructure — “lost money and then didn’t.”
The thing is, this statement is one of those things that people say because it sounds rational. Amazon did lose money, and Amazon Web Services was expensive, that’s obvious, right?
The thing is, I’ve never really had anyone explain this point to me, so I am finally going to sit down and deal with this criticism, because every single person who mentions it thinks they just pulled Excalibur from the stone and can now decapitate me. They claim that because people in the past doubted Amazon — because, or in addition to the burn rate of Amazon Web Services as the company built out its infrastructure — that I too am wrong, because they were wrong about that.
This isn’t Camelot, you rube! You are not King Arthur!
I will address both the argument itself and the “they” part of it too — because if the argument is that the people that got AWS wrong should not be trusted, then we should no longer trust them, the people actively propagandizing our supposed generative AI future.
Right?
So, I’m honestly not sure where this argument came from, because there is, to my knowledge, no story about Amazon Web Services where somebody suggested its burnrate would kill Amazon.
But let’s go back in time to the May 31 1999 piece that some might be thinking of, called “Amazon.bomb,” and how writer Jacqueline Doherty was mocked soundly for “being wrong” about Amazon, which has now become quite profitable.
I also want to be clear that Amazon Web Services didn’t launch until 2006, and Amazon itself would become reliably profitable in 2003. Technically Amazon had opened up Amazon.com’s web services for developers to incorporate its content into their applications in 2002, but what we consider AWS today — cloud storage and compute — launched in 2006.
But okay, what did she actually say?
Unfortunately for Bezos, Amazon is now entering a stage in which investors will be less willing to rely on his charisma and more demanding of answers to tough questions like, when will this company actually turn a profit? And how will Amazon triumph over a slew of new competitors who have deep pockets and new technologies?We tried to ask Bezos, but he declined to make himself or any other executives of the company available. He can ignore Barron’s, but he can’t ignore the questions.Amazon last year posted a loss of $125 million [$242.6m in today’s money) on revenues of $610 million [$1.183 billion in today’s money]. And in this year’s first quarter it got even worse, as the company posted a loss of $61.7 million [$119.75 million in today’s money] on revenues of $293.6 million [$569.82 million in today’s money].
Her argument, for the most part, is that Amazon was burning cash, had a ton of competition from other people doing similar things, and that analysts backed her up:
“The first mover does not always win. The importance of being first is a mantra in the Internet world, but it’s wrong. The ones that are the most efficient will be successful,” says one retail analyst. “In retailing, anyone can build a great-looking store. The hard part is building a great-looking store that makes money.”
Fair arguments for the time, though perhaps a little narrow-minded. The assumption wasn’t that what Amazon was building was a bad idea, but that Amazon wouldn’t be the ones to build it, with one saying:
“Once Wal-Mart decides to go after Amazon, there’s no contest,” declares Kurt Barnard, president of Barnard’s Retail Trend Report. “Wal-Mart has resources Amazon can’t even dream about.”
In simpler terms: Amazon’s business model wasn’t in question. People were buying shit online. In fact, this was just before the dot com bubble burst, and when optimism about the web was at a high point. Yet the comparison stops there — people obviously liked buying shit online, it was the business models of many of these companies — like WebVan — that sucked.
But Let’s Talk About Amazon Web Services
Amazon Web Services was an outgrowth of Amazon’s own infrastructure, which had to expand rapidly to deal with the influx of web traffic for Amazon.com, which had become one of the world’s most popular websites and was becoming increasingly more-complex as it sold things other than books. Other companies had their own infrastructure, but if a smaller company wanted to scale, they’d basically need to build their own thing.
It’s actually pretty cool what Amazon did! Remember, this was the early 2000s, before Facebook, Twitter, and a lot of the modern internet we know that runs on services like Amazon Web Services, Microsoft Azure and Google Cloud. It invented the modern concept of compute!
But we’re here to talk about Amazon Web Services being dangerous for Amazon and people hating on it.
A November 2006 story from Bloomberg talked about Jeff Bezos’ Risky Bet to “run your business with the technology behind his web site,” saying that “Wall Street [wanted] him to mind the store.” Bezos, referred to as a “one-time internet poster boy” that became “a post-dot-com piñata.” Nevertheless, this article has what my haters crave:
But if techies are wowed by Bezos’ grand plan, it’s not likely to win many converts on Wall Street. To many observers, it conjures up the ghost of Amazon past. During the dot-com boom, Bezos spent hundreds of millions of dollars to build distribution centers and computer systems in the promise that they eventually would pay off with outsize returns. That helped set the stage for the world’s biggest Web retail operation, with expected sales of $10.5 billion this year…All that has investors restless and many analysts throwing up their hands wondering if Bezos is merely flailing around for an alternative to his retail operation. Eleven of 27 analysts who follow the company have underperform or sell ratings on the stock–a stunning vote of no confidence. That number of sell recommendations is matched among large companies only by Qwest Communications International Inc. (Q ), according to investment consultant StarMine Corp. It’s more than even the eight sell opinions on struggling Ford Motor Co. (F )
Pretty bad, right? My goose is cooked? All those analysts seem pretty mad!
Except it’s not, my goose is raw! Yours, however, has been in the oven for over a year!
Emphasis mine:
By all accounts, Amazon’s new businesses bring in a minuscule amount of revenue. Although its direct cost of providing them appears relatively low because the hardware and software are in place, Stifel Nicolaus & Co. (SF ) analyst Scott W. Devitt notes: “There’s not going to be any economic return from any of these projects for the foreseeable future.” Bezos himself admits as much. But with several years of heavy spending already, he’s making this a priority for the long haul. “We think it’s going to be a very meaningful business for us one day,” he says. “What we’ve historically seen is that the seeds we plant can take anywhere from three, five, seven years.”
That’s right — the ongoing costs aren’t the problem.
Hey wait a second, that’s a name! I can look up a name! Scott W. Devitt now works at Wedbush as its managing director of equity research, and has said AI companies would enter a new stage in 2025…god, just read this:
The second stage is “the application phase of the cycle, which should benefit software companies as well as the cloud providers. And then, phase three of this will ultimately be the consumer-facing companies figuring out how to use the technology in ways that actually can drive increased interactions with consumers.”
The analyst says the market will enter phase two in 2025, with software companies and cloud provider stocks expected to see gains. He adds that cybersecurity companies could also benefit as the technology evolves.
Dewitt specifically calls out Palantir, Snowflake, and Salesforce as those who would “gain.” In none of these cases am I able to see the actual revenue from AI, but Salesforce itself said that it will see no revenue growth from AI this year. Palantir also, as discovered by the Autonomy Institute’s recent study, recently added to the following to its public disclosures:
There are significant risks involved in deploying AI and there can be no assurance that using AI in our platforms and products will enhance or be beneficial to our business, including our profitability.
What I’m saying is that analysts can be wrong! And they can be wrong at scale! There is no analyst consensus that agrees with me. In fact, most analysts appear to be bullish on AI, despite the significantly-worse costs and total lack of growth!
Yet even in this Hater’s Parade, the unnamed journalist makes a case for Amazon Web Services:
Sooner than that, those initiatives may provide a boost for Amazon’s retail side. For one, they potentially make a profit center out of idle computing capacity needed for that retail operation. Like most computer networks, Amazon’s uses as little as 10% of its capacity at any one time just to leave room for occasional spikes. It’s the same story in the company’s distribution centers. Keeping them humming at higher capacity means they operate more efficiently, besides giving customers a much broader selection of products. And the more stuff Amazon ships, both its own inventory or others’, the better deals it can cut with shippers.
But Amazon Web Services Cost Money Ed, Now You Shall Meet Your End!
Nice try, chuckles!
In 2015, the year that Amazon Web Services became profitable, Morgan Stanley analyst Katy Huberty believed that it was running at a “material loss,” suggesting that $5.5 billion of Amazon’s “technology and content expenses” was actually AWS expenses, with a “negative contribution of $1.3 billion.”
Here is Katy Huberty, the analyst in question, declaring six months ago that “2025 [will] be the year of Agentic AI, robust enterprise adoption, and broadening AI winners.”
So, yes, analysts really got AWS wrong. But putting that aside, there might actually be a comparison here! Amazon Web Services absolutely created a capital expenditures drain on Amazon. From Forbes’s Chuck Jones:
In 2014 Amazon had $4.9 billion in capital expenditures, up 42% from 2013’s $3.4 billion. The company has a wide range of items that it buys to support and grow its business ranging from warehouses, robots and computer systems for its core retail business and AWS. While I don’t expect Amazon to detail how much goes to AWS I suspect it is a decent percentage, which means AWS needs to generate appropriate returns on the capital deployed.
In today’s money, this means that Amazon spent $6.76 billion in capital expenditures on AWS in 2014. Assuming it was this much every year — it wasn’t, but I want to make an example of every person claiming that this is a gotcha — it took $67.6 billion and ten years (though one could argue it was nine) of pure capital expenditures to turn Amazon Web Services into [a business that now makes billions of dollars a quarter in profit](https://www.geekwire.com/2025/amazons-quarterly-profits-soar-to-a-record-20-billion-but-cloud-growth-comes-up-short/?ref=wheresyoured.at#%3A~%3Atext=Although+AWS+sales+(%2428.8%2C%2421.2+billion+for+the+quarter.).
That’s $15.4 billion less than Amazon’s capital expenditures for 2024, and less than one-fifteenth its projected capex spend for 2025. And to be clear, the actual capital expenditure numbers are likely much lower, but I want to make it clear that even when factoring in inflation, Amazon Web Services was A) a bargain and B) a fraction of the cost of what Amazon has spent in 2024 or 2025.
A fun aside: On March 30 2015, Kevin Roose wrote a piece for New York Magazine about the cloud compute wars, in which he claimed that, and I quote, “there’s no reason to suspect that Amazon would ever need to raise prices on AWS, or turn the fabled “profit switch” that pundits have been speculating about for years.” Less than a month later Amazon revealed Amazon Web Services was profitable. They don’t call him “the most right man in tech journalism” for nothing!
Generative AI and Large Language Models Do Not Resemble Amazon Web Services or The Greater Cloud Compute Boom, As Generative AI Is Not Infrastructure
Some people compare Large Language Models and their associated services to Amazon Web Services, or services like Microsoft Azure or Google Cloud, and they are wrong to do so.
Amazon Web Services, when it launched, comprised of things like (and forgive how much I’m diluting this) Amazon’s Elastic Compute Cloud (EC2), where you rent space on Amazon’s servers to run applications in the cloud, or Amazon Simple Storage (S3), which is enterprise-level storage for applications. In simpler terms, if you were providing a cloud-based service, you used Amazon to both store the stuff that the service needed and the actual cloud-based processing (compute, so like your computer loads and runs applications but delivered to thousands or millions of people).
This is a huge industry. Amazon Web Services alone brought in revenues of over $100 billion in 2024, and while Microsoft and Google don’t break out their cloud revenues, they’re similarly large parts of their revenue, and Microsoft has used Azure in the past to patch over shoddy growth.
These services are also selling infrastructure. You aren’t just paying for the compute, but the ability to access storage and deliver services with low latency — so users have a snappy experience — wherever they are in the world. The subtle magic of the internet is that it works at all, and a large part of that is the cloud compute infrastructure and oligopoly of the main providers having such vast data centers. This is much cheaper than doing it yourself, until a certain point. Dropbox moved away from Amazon Web Services as it scaled. It also allows someone else to take care of maintenance of the hardware and make sure it actually gets to your customers. You also don’t have to worry about spikes in usage, because these things are usage-based, and you can always add more compute to meet demand.
There is, of course, nuance — security-specific features, content-specific delivery services, database services — behind these clouds. You are buying into the infrastructure of the infrastructure provider, and the reason these products are so profitable is, in part, because you are handing off the problems and responsibility to somebody else. And based on that idea, there are multiple product categories you can build on top of it, because ultimately cloud services are about Amazon, Microsoft and Google running your infrastructure for you.
Large Language Models and their associated services are completely different, despite these companies attempting to prove otherwise, and it starts with a very simple problem: why did any of these companies build these giant data centers and fill them full of GPUs?
Amazon Web Services was created out of necessity — Amazon’s infrastructure needs were so great that it effectively had to build both the software and hardware necessary to deliver a store that sold theoretically everything to theoretically anywhere, handling both the traffic from customers, delivering the software that runs Amazon.com quickly and reliably, and, well, making sure things ran in a stable way. It didn’t need to come up with a reason for people to run web applications — they were already doing so themselves, but in ways that cost a lot, were inflexible, and required specialist skills. AWS took something that people already did, and what there was a proven demand for, and made it better. Eventually, Google and Microsoft would join the fray.
And that appears to be the only similarity with generative AI — that due to the ridiculous costs of both the data centers and GPUs necessary to provide these services, it’s largely impossible for others to even enter the market.
Yet after that, generative AI feels more like a feature of cloud infrastructure rather than infrastructure itself. AWS and similar megaclouds are versatile, flexible and multifaceted. Generative AI does what generative AI does, and that’s about it.
You can run lots of different things on AWS. What are the different things you can run using Large Language Models? What are the different use cases, and, indeed, user requirements that make this the supposed “next big thing”?
Perhaps the argument is that generative AI is the next AWS or similar cloud service because you can build the next great companies on the infrastructure of others — the models of, say, OpenAI and Anthropic, and the servers of Microsoft.
So, okay, let’s humour this point too. You can build the next great AI startup, and you have to build it on one of the megclouds because they’re the only ones that can afford to build the infrastructure.
One small problem.
Companies Built On Top Of Large Language Models Don’t Make Much Money (In Fact, They’re Likely All Deeply Unprofitable)
Let’s start by establishing a few facts:
Outside of one exception — Midjourney, which claimed it was profitable in 2022, which may not still be the case, I’ve reached out to ask— every single Large Language Model company is unprofitable, often wildly so.Outside of OpenAI, Anthropic and Anysphere (which makes AI coding app Cursor), there are no Large Language Model companies — either building models or services on top of others’ models — that make more than $500 million in annualized revenue (meaning month x 12), and outside of Midjourney ($200m ARR) and Ironclad ($150m ARR), according to The Information’s Generative AI database, and Perplexity (which just announced it’s at $150m ARR), there are only twelve generative AI-powered companies making $100 million annualized (or $8.3 million a month) in revenue. Though the database doesn’t have Replit (which recently announced it hit $100 million in annualized revenue), I’ve included it in my calculations for the sake of fairness. Of these companies, two have been acquired — Moveworks (acquired by ServiceNow in March 2025) and Windsurf (acquired by Cognition in July 2025).For the sake of simplicity, I’ve left out companies like Surge, Scale, Turing and Together, all of whom run consultancies selling services for training models.Otherwise, there are seven companies that make $50 million or more ARR ($4.16 million a month).
None of this is to say that one hundred million dollars isn’t a lot of money to you and me, but in the world of Software-as-Service or enterprise software, this is chump change. Hubspot had revenues of $2.63 billion in its 2024 financial year.
[Content truncated due to length…]
You must log in or # to comment.