Soundtrack: Queens of the Stone Age - First It Giveth
Before we go any further: This is, for the third time this year, the longest newsletter I’ve ever written, weighing in somewhere around 18,500 words. I’ve written it specifically to be read at your leisure — dip in and out where you’d like — but also in one go.
This is my comprehensive case that yes, we’re in a bubble, one that will inevitably (and violently) collapse in the near future.
I’ll also be cutting this into a four-part episode starting tomorrow on my podcast Better Offline.
I deeply appreciate your time. If you like this newsletter, please think about subscribing to the premium, which I write weekly.
Thanks for reading.
"No of course there isn’t enough capital for all of this. Having said that, there is enough capital to do this for at least a little while longer."Analyst Gil Luria of D.A. Davidson, on being asked if there was enough available capital to build OpenAI’s promised 17 Gigawatts of data centers.
Alright, let’s do this one last time.
In 2022, a (kind-of) company called OpenAI surprised the world with a website called ChatGPT that could generate text that sort-of sounded like a person using a technology called Large Language Models (LLMs), which can also be used to generate images, video and computer code.
Large Language Models require entire clusters of servers connected with high-speed networking, all containing this thing called a GPU — graphics processing units. These are different to the GPUs in your Xbox, or laptop, or gaming PC. They cost much, much more, and they’re good at doing the processes of inference (the creation of the output of any LLM) and training (feeding masses of training data to models, or feeding them information about what a good output might look like, so they can later identify a thing or replicate it).
These models showed some immediate promise in their ability to articulate concepts or generate video, visuals, audio, text and code. They also immediately had one glaring, obvious problem: because they’re probabilistic, these models can’t actually be relied upon to do the same thing every single time.
So, if you generated a picture of a person that you wanted to, for example, use in a story book, every time you created a new page, using the same prompt to describe the protagonist, that person would look different — and that difference could be minor (something that a reader should shrug off), or it could make that character look like a completely different person.
Moreover, the probabilistic nature of generative AI meant that whenever you asked it a question, it would guess as to the answer, not because it knew the answer, but rather because it was guessing on the right word to add in a sentence based on previous training data. As a result, these models would frequently make mistakes — something which we later referred to as “hallucinations.”
And that’s not even mentioning the cost of training these models, the cost of running them, the vast amounts of computational power they required, the fact that the legality of using material scraped from books and the web without the owner’s permission was (and remains) legally dubious, or the fact that nobody seemed to know how to use these models to actually create profitable businesses.
These problems were overshadowed by something flashy, and new, and something that investors — and the tech media — believed would eventually automate the single thing that’s proven most resistant to automation: namely, knowledge work and the creative economy.
This newness and hype and these expectations sent the market into a frenzy, with every hyperscaler immediately creating the most aggressive market for one supplier I’ve ever seen. NVIDIA has sold over $200 billion of GPUs since the beginning of 2023, becoming the largest company on the stock market and trading at over $170 as of writing this sentence only a few years after being worth $19.52 a share.
Sidenote: those figures reflect the fact that Nvidia’s stock split 10-to-1 in 2024 — or, said plainly, if you held one share before the split, you’d hold ten shares afterwards, changing the unit price of the company’s equity (making it cheaper to buy a share, and thus, more accessible to retail investors) without changing the absolute value of the company. This bit isn’t necessarily important to what I’ve written, but given the subject of this newsletter, I think it’s important to lean towards being as explicit as possible about the numbers I share.
While I’ve talked about some of the propelling factors behind the AI wave — automation and novelty — that’s not a complete picture. A huge reason why everybody decided to “do AI” was because the software industry’s growth was slowing, with SaaS (Software As A Service) company valuations stalling or dropping, resulting in the terrifying prospect of companies having to “under promise and over deliver” and “be efficient.”
Things that normal companies — those whose valuations aren’t contingent on ever-increasing, ever-constant growth — don’t have to worry about, because they’re normal companies.
Suddenly, there was the promise of a new technology — Large Language Models — that were getting exponentially more powerful, which was mostly a lie but hard to disprove because “powerful” can mean basically anything, and the definition of “powerful” depended entirely on whoever you asked at any given time, and what that person’s motivations were.
The media also immediately started tripping on its own feet, mistakenly claiming OpenAI’s GPT-4 model tricked a Taskrabbit into solving a CAPTCHA (it didn’t — this never happened), or saying that “people who don’t know how to code already [used] bots to produce full-fledged games,” and if you’re wondering what “full-fledged” means, it means “pong” and a cobbled-together rolling demo of SkyRoads, a game from 1993.
The media (and investors) helped peddle the narrative that AI was always getting better, could do basically anything, and that any problems you saw today would be inevitably solved in a few short months, or years, or, well, at some point I guess.
LLMs were touted as a digital panacea, and the companies building them offered traditional software companies the chance to plug these models into their software using an API, thus allowing them to ride the same generative AI wave that every other company was riding.
The model companies similarly started going after individual and business customers, offering software and subscriptions that promised the world, though this mostly boiled down to chatbots that could generate stuff, and then doubled down with the promise of “agents” — a marketing term that’s meant to make you think “autonomous digital worker” but really means “broken digital product.”
Throughout this era, investors and the media spoke with a sense of inevitability that they never really backed up with data. It was an era based on confidently-asserted “vibes.” Everything was always getting better and more powerful, even though there was never much proof that this was truly disruptive technology, other than in its ability to disrupt apps you were using with AI — making them worse by, for example, suggesting questions on every Facebook post that you could ask Meta AI, but which Meta AI couldn’t answer.
“AI” was omnipresent, and it eventually grew to mean everything and nothing. OpenAI would see its every move lorded over like a gifted child, its CEO Sam Altman called the “Oppenheimer of Our Age,” even if it wasn’t really obvious why everyone was impressed. GPT-4 felt like something a bit different, but was it actually meaningful?
The thing is, Artificial Intelligence is built and sold on not just faith, but a series of myths that the AI boosters expect us to believe with the same certainty that we treat things like gravity, or the boiling point of water.
Can large language models actually replace coders? Not really, no, and I’ll get into why later in the piece.
Can Sora — OpenAI’s video creation tool — replace actors or animators? No, not at all, but it still fills the air full of tension because you can immediately see who is pre-registered to replace everyone that works for them.
AI is apparently replacing workers, but nobody appears to be able to prove it! But every few weeks a story runs where everybody tries to pretend that AI is replacing workers with some poorly-sourced and incomprehensible study, never actually saying “someone’s job got replaced by AI” because it isn’t happening at scale, and because if you provide real-world examples, people can actually check.
To be clear, some people have lost jobs to AI, just not the white collar workers, software engineers, or really any of the career paths that the mainstream media and AI investors would have you believe.
Brian Merchant has done excellent work covering how LLMs have devoured the work of translators, using cheap, “almost good” automation to lower already-stagnant wages in a field that was already hurting before the advent of generative AI, with some having to abandon the field, and others pushed into bankruptcy. I’ve heard the same for art directors, SEO experts, and copy editors, and Christopher Mims of the Wall Street Journal covered these last year.
These are all fields with something in common: shitty bosses with little regard for their customers who have been eagerly waiting for the opportunity to slash contract labor. To quote Merchant, “the drumbeat, marketing, and pop culture of ‘powerful AI’ encourages and permits management to replace or degrade jobs they might not otherwise have.”
Across the board, the people being “replaced” by AI are the victims of lazy, incompetent cost-cutters who don’t care if they ship poorly-translated text. To quote Merchant again, “[AI hype] has created the cover necessary to justify slashing rates and accepting “good enough” automation output for video games and media products.”
Yet the jobs crisis facing translators speaks to the larger flaws of the Large Language Model era, and why other careers aren’t seeing this kind of disruption.
Generative AI creates outputs, and by extension defines all labor as some kind of output created from a request. In the case of translation, it’s possible for a company to get by with a shitty version, because many customers see translation as “what do these words say,” even though (as one worker told Merchant) translation is about conveying meaning. Nevertheless, “translation” work had already started to condense to a world where humans would at times clean up machine-generated text, and the same worker warned that the same might come for other industries.
Yet the problem is that translation is a heavily output-driven industry, one where (idiot) bosses can say “oh yeah that’s fine” because they ran an output back through Google Translate and it seemed fine in their native tongue. The problems of a poor translation are obvious, but the customers of translation are, it seems, often capable of getting by with a shitty product.
The problem is that most jobs are not output-driven at all, and what we’re buying from a human being is a person’s ability to think.
Every CEO talking about AI replacing workers is an example of the real problem: that most companies are run by people who don’t understand or experience the problems they’re solving, don’t do any real work, don’t face any real problems, and thus can never be trusted to solve them. The Era of the Business Idiot is the result of letting management consultants and neoliberal “free market” sociopaths take over everything, leaving us with companies run by people who don’t know how the companies make money, just that they must always make more.
When you’re a big, stupid asshole, every job that you see is condensed to its outputs, and not the stuff that leads up to the output, or the small nuances and conscious decisions that make an output good as opposed to simply acceptable, or even bad.
What does a software engineer do? They write code! What does a writer do? They write words! What does a hairdresser do? They cut hair!
Yet that’s not actually the case.
As I’ll get into later, a software engineer does far more than just code, and when they write code they’re not just saying “what would solve this problem?” with a big smile on their face — they’re taking into account their years of experience, what code does, what code could do, all the things that might break as a result, and all of the things that you can’t really tell from just looking at code, like whether there’s a reason things are made in a particular way.
A good coder doesn’t just hammer at the keyboard with the aim of doing a particular task. They factor in questions like: How does this functionality fit into the code that’s already here? Or, if someone has to update this code in the future, how do I make it easy for them to understand what I’ve written and to make changes without breaking a bunch of other stuff?
A writer doesn’t just “write words.” They jostle ideas and ideals and emotions and thoughts and facts and feelings into a condensed piece of text, explaining both what’s happening and why it’s happening from their perspective, finding nuanced ways to convey large topics, none of which is the result of a single (or many) prompts but the ever-shifting sand of a writer’s brain.
Good writing is a fight between a bunch of different factors: structure, style, intent, audience, and prioritizing the things that you (or your client) care about in the text. It’s often emotive — or at the very least, driven or inspired by a given emotion — which is something that an AI simply can’t replicate in a way that’s authentic and believable.
And a hairdresser doesn’t just cut hair, but cuts your hair, which may be wiry, dry, oily, long, short, healthy, unhealthy, on a scalp with particular issues, at a time of year when perhaps you want to change length, at a time that fits you, in “the way you like” which may be impossible to actually write down but they get it just right. And they make conversation, making you feel at ease while they snip and clip away at your tresses, with you having to trust that they’ll get it right.
This is the true nature of labor that executives fail to comprehend at scale: that the things we do are not units of work, but extrapolations of experience, emotion, and context that cannot be condensed in written meaning. Business Idiots see our labor as the result of a smart manager saying “do this,” rather than human ingenuity interpreting both a request and the shit the manager didn’t say.
What does a CEO do? Uhhh, um, well, a Harvard study says they spend 25% of their time on “people and relationships,” 25% on “functional and business unit reviews,” 16% on “organization and culture,” and 21% on “strategy,” with a few percent here and there for things like “professional development.”
That’s who runs the vast majority of companies: people that describe their work predominantly as “looking at stuff,” “talking to people” and “thinking about what we do next.” The most highly-paid jobs in the world are impossible to describe, their labor described in a mish-mash of LinkedInspiraton, yet everybody else’s labor is an output that can be automated.
As a result, Large Language Models seem like magic. When you see everything as an outcome — an outcome you may or may not understand, and definitely don’t understand the process behind, let alone care about — you kind of already see your workers as LLMs.
You create a stratification of the workforce that goes beyond the normal organizational chart, with senior executives — those closer to the class level of CEO — acting as those who have risen above the doldrums of doing things to the level of “decisionmaking,” a fuzzy term that can mean everything from “making nuanced decisions with input from multiple different subject-matter experts” to, as ServiceNow Bill McDermott did in 2022, “[make] it clear to everybody [in a boardroom of other executives], everything you do: AI, AI, AI, AI, AI.”
The same extends to some members of the business and tech media that have, for the most part, gotten by without having to think too hard about the actual things the companies are saying.
I realize this sounds a little mean, and I must be clear it doesn’t mean that these people know nothing, just that it’s been possible to scoot through the world without thinking too hard about whether or not something is true. When Salesforce said back in 2024 that its “Einstein Trust Layer” and AI would be “transformational for jobs,” the media dutifully wrote it down and published it without a second thought. It fully trusted Marc Benioff when he said that Agentforce agents would replace human workers, and then again when he said that AI agents are doing “30% to 50% of all the work in Salesforce itself,” even though that’s an unproven and nakedly ridiculous statement.
Salesforce’s CFO said earlier this year that AI wouldn’t boost sales growth in 2025. One would think this would change how they’re covered, or how seriously one takes Marc Benioff.
It hasn’t, because nobody is paying attention. In fact, nobody seems to be doing their job.
This is how the core myths of generative AI were built: by executives saying stuff and the media publishing it without thinking too hard.
AI is replacing workers! AI is writing entire computer programs! AI is getting exponentially more-powerful! What does “powerful” mean? That the models are getting better on benchmarks that are rigged in their favor, but because nobody fucking explains it, regular people are regularly told that AI is “powerful.”
The only thing “powerful” about generative AI is its mythology. The world’s executives, entirely disconnected from labor and actual production, are doing the only thing they know how to — spend a bunch of money and say vague stuff about “AI being the future.” There are people — journalists, investors, and analysts — that have built entire careers on filling in the gaps for the powerful as they splurge billions of dollars and repeat with increasing desperation that “the future is here” as absolutely nothing happens.
You’ve likely seen a few ridiculous headlines recently. One of the most recent, and most absurd, is that that OpenAI will pay Oracle $300 billion over four years, closely followed with the claim that NVIDIA will “invest” “$100 billion” in OpenAI to build 10GW of AI data centers, though the deal is structured in a way that means that OpenAI is paid “progressively as each gigawatt is deployed,” and OpenAI will be leasing the chips (rather than buying them outright). I must be clear that these deals are intentionally made to continue the myth of generative AI, to pump NVIDIA, and to make sure OpenAI insiders can sell $10.3 billion of shares.
OpenAI cannot afford the $300 billion, NVIDIA hasn’t sent OpenAI a cent and won’t do so if it can’t build the data centers, which OpenAI most assuredly can’t afford to do.
NVIDIA needs this myth to continue, because in truth, all of these data centers are being built for demand that doesn’t exist, or that — if it exists — doesn’t necessarily translate into business customers paying huge amounts for access to OpenAI’s generative AI services.
NVIDIA, OpenAI, CoreWeave and other AI-related companies hope that by announcing theoretical billions of dollars (or hundreds of billions of dollars) of these strange, vague and impossible-seeming deals, they can keep pretending that demand is there, because why else would they build all of these data centers, right?
That, and the entire stock market rests on NVIDIA’s back. It accounts for 7% to 8% of the value of the S&P 500, and Jensen Huang needs to keep selling GPUs. I intend to explain later on how all of this works, and how brittle it really is.
The intention of these deals is simple: to make you think “this much money can’t be wrong.”
It can. These people need you to believe this is inevitable, but they are being proven wrong, again and again, and today I’m going to continue doing so.
Underpinning these stories about huge amounts of money and endless opportunity lies a dark secret — that none of this is working, and all of this money has been invested in a technology that doesn’t make much revenue and loves to burn millions or billions or hundreds of billions of dollars.
Over half a trillion dollars has gone into an entire industry without a single profitable company developing models or products built on top of models. By my estimates, there is around $44 billion of revenue in generative AI this year (when you add in Anthropic and OpenAI’s revenues to the pot, along with the other stragglers) and most of that number has been gathered through reporting from outlets like The Information, because none of these companies share their revenues, all of them lose shit tons of money, and their actual revenues are really, really small.
Only one member of the Magnificent Seven (outside of NVIDIA) has ever disclosed its AI revenue — Microsoft, which stopped reporting in January 2025, when it reported “$13 billion in annualized revenue,” so around $1.083 billion a month.
Microsoft is a sales MACHINE. It is built specifically to create or exploit software markets, suffocating competitors by using its scale to drive down prices, and to leverage the ecosystem that it’s created over the past few decades. $1 billion a month in revenue is chump change for an organization that makes over $27 billion a quarter in PROFITS.
Don’t worry Satya, I’ll come back to you later.
“But Ed, the early days!” Worry not — I’ve got that covered.
This is nothing like any other era of tech. There has never been this kind of cash-rush, even in the fiber boom. Over a decade, Amazon spent about one-tenth of the capex that the Magnificent Seven spent in two years on generative AI building AWS — something that now powers a vast chunk of the web, and has long been Amazon’s most profitable business unit.
Generative AI is nothing like Uber, with OpenAI and Anthropic’s true costs coming in at about $159 billion in the past two years, approaching five times Uber’s $30 billion all-time burn. And that’s before the bullshit with NVIDIA and Oracle.
Microsoft last reported AI revenue in January. It’s October this week. Why did it stop reporting this number, you think? Is it because the numbers are so good it couldn’t possibly let people know?
As a general rule, publicly traded companies — especially those where the leadership are compensated primarily in equity — tend to brag about their successes, in part because said bragging boosts the value of the thing that the leadership gets paid in. There’s no benefit to being shy. Oracle literally made a regulatory filing to boast it had a $30 billion customer, which turned out to be OpenAI, who eventually agreed (publicly) to spend $300 billion in compute over five years.
Which is to say that Microsoft clearly doesn’t have any good news to share, and as I’ll reveal later, they can’t even get 3% of their 440 million Microsoft 365 subscribers to pay for Microsoft 365 Copilot.
If Microsoft can’t sell this shit, nobody can.
Anyway, I’m nearly done, sorry, you see, I’m writing this whole thing as if you’re brand new and walking up to this relatively unprepared, so I need to introduce another company.
In 2020, a splinter group jumped off of OpenAI, funded by Amazon and Google to do much the same thing as OpenAI but pretend to be nicer about it until they have to raise from the Middle East. Anthropic has always been better at coding for some reason, and people really like its Claude models.
Both OpenAI and Anthropic have become the only two companies in generative AI to make any real progress, either in terms of recognition or in sheer commercial terms, accounting for the majority of the revenue in the AI industry.
In a very real sense, the AI industry’s revenue is OpenAI and Anthropic. In the year where Microsoft recorded $13bn in AI revenues, $10 billion came from OpenAI’s spending on Microsoft Azure. Anthropic burned $5.3 billion last year — with the vast majority of that going towards compute. Outside of these two companies, there’s barely enough revenue to justify a single data center.
Where we sit today is a time of immense tension. Mark Zuckerberg says we’re in a bubble, Sam Altman says we’re in a bubble, Alibaba Chairman and billionaire Joe Tsai says we’re in a bubble, Apollo says we’re in a bubble, nobody is making money and nobody knows why they’re actually doing this anymore, just that they must do it immediately.
And they have yet to make the case that generative AI warranted any of these expenditures.
That was undoubtedly the longest introduction to a newsletter I’ve ever written, and the reason why I took my time was because this post demands a level of foreshadowing and exposition, and because I want to make it make sense to anyone who reads it — whether they’ve read my newsletter for years, or whether they’re only just now investigating their suspicions that generative AI may not be all it’s cracked up to be.
Today I will make the case that generative AI’s fundamental growth story is flawed, and explain why we’re in the midst of an egregious bubble.
This industry is sold by keeping things vague, and knowing that most people don’t dig much deeper than a headline, a problem I simply do not have.
This industry is effectively in service of two companies — OpenAI and NVIDIA — who pump headlines out through endless contracts between them and subsidiaries or investments to give the illusion of activity.
OpenAI is now, at this point, on the hook for over a trillion dollars, an egregious sum for a company that already forecast billions in losses, with no clear explanation as to how it’ll afford any of this beyond “we need more money” and the vague hope that there’s another Softbank or Microsoft waiting in the wings to swoop in and save the day.
I’m going to walk you through where I see this industry today, and why I see no future for it beyond a fiery apocalypse.
While everybody (reasonably!) harps on about hallucinations — which, to remind you, is when a model authoritatively states something that isn’t true — but the truth is far more complex, and far worse than it seems.
You cannot rely on a large language model to do what you want. Even the most highly-tuned models on the most expensive and intricate platform can’t actually be relied upon to do exactly what you want.
A “hallucination” isn’t just when these models say something that isn’t true. It’s when they decide to do something wrong because it seems the most likely thing to do, or when a coding model decides to go on a wild goose chase, failing the user and burning a ton of money in the process.
The advent of “reasoning” models — those engineered to ‘think’ through problems in a way reminiscent of a human — and the expansion of what people are (trying) to use LLMs for demands that the definition of an AI hallucination be widened, not merely referring to factual errors, but fundamental errors in understanding the user’s request or intent, or what constitutes a task, in part because these models cannot think and do not know anything.
However successful a model might be in generating something good *once*, it will also often generate something bad, or it’ll generate the right thing but in an inefficient and over-verbose fashion. You do not know what you’re going to get each time, and hallucinations multiply with the complexity of the thing you’re asking for, or whether a task contains multiple steps (which is a fatal blow to the idea of “agents.”
You can add as many levels of intrigue and “reasoning” as you want, but Large Language Models cannot be trusted to do something correctly, or even consistently, every time. Model companies have successfully convinced everybody that the issue is that users are prompting the models wrong, and that people need to be “trained to use AI,” but what they’re doing is training people to explain away the inconsistencies of Large Language Models, and to assume individual responsibility for what is an innate flaw in how large language models work.
Large Language Models are also uniquely expensive. Many mistakenly try and claim this is like the dot com boom or Uber, but the basic unit economics of generative AI are insane. Providers must purchase tens or hundreds of thousands of GPUs each costing $50,000 a piece, and hundreds of millions or billions of dollars of infrastructure for large clusters. And that’s without mentioning things like staffing, construction, power, or water.
Then you turn them on and start losing money. Despite hundreds of billions of GPUs sold, nobody seems to make any money, other than NVIDIA, the company that makes them, and resellers like Dell and Supermicro who buy the GPUs, put them in servers, and sell them to other people.
This arrangement works out great for Jensen Huang, and terribly for everybody else.
I am going to explain the insanity of the situation we find ourselves in, and why I continue to do this work undeterred. The bubble has entered its most pornographic, aggressive and destructive stage, where the more obvious it becomes that they’re cooked, the more ridiculous the generative AI industry will act — a dark juxtaposition against every new study that says “generative AI does not work” or new story about ChatGPT’s uncanny ability to activate mental illness in people.
The Markets Have Become Dependent On NVIDIA, and NVIDIA Is Taking Extraordinary, Dangerous Measures To Sustain Growth, Investing In Companies Specifically To Raise Debt To Buy Their Own GPUs
So, let’s start simple: NVIDIA is a hardware company that sells GPUs, including the consumer GPUs that you’d see in a modern gaming PC, but when you read someone say “GPU” within the context of AI, they mean enterprise-focused GPUs like the A100, H100, H200, and more modern GPUs like the Blackwell-series B200 and GB200 (which combines two GPUs with an NVIDIA CPU).
These GPUs cost anywhere from $30,000 to $50,000 (or as high as $70,000 for the newer Blackwell GPUs), and require tens of thousands of dollars more of infrastructure — networking to “cluster” server racks of GPUs together to provide compute, and massive cooling systems to deal with the massive amounts of heat they produce, as well as the servers themselves that they run on, which typically use top-of-the-line data center CPUs, and contain vast quantities of high-speed memory and storage. While the GPU itself is likely the most expensive single item within an AI server, the other costs — and I’m not even factoring in the actual physical building that the server lives in, or the water or electricity that it uses — add up.
I’ve mentioned NVIDIA because it has a virtual monopoly in this space. Generative AI effectively requires NVIDIA GPUs, in part because it’s the only company really making the kinds of high-powered cards that generative AI demands, and because NVIDIA created something called CUDA — a collection of software tools that lets programmers write software that runs on GPUs, which were traditionally used primarily for rendering graphics in games.
While there are open-source alternatives, as well as alternatives from Intel (with its ARC GPUs) and AMD (Nvidia’s main rival in the consumer space), these aren’t nearly as mature or feature-rich.
Due to the complexities of AI models, one cannot just stand up a few of these things either — you need clusters of thousands, tens of thousands, or hundreds of thousands of them for it to be worthwhile, making any investment in GPUs in the hundreds of millions or billions of dollars, especially considering they require completely different data center architecture to make them run.
A common request — like asking a generative AI model to parse through thousands of lines of code and make a change or an addition — may use multiple of these $50,000 GPUs at the same time, and so if you aspire to serve thousands, or millions of concurrent users, you need to spend big. Really big.
It’s these factors — the vendor lock-in, the ecosystem, and the fact that generative AI only works when you’re buying GPUs at scale — that underpin the rise of Nvidia. But beyond the economic and technical factors, there are human ones, too.
To understand the AI bubble is to understand why CEOs do the things they do. Because an executive’s job is so vague, they can telegraph the value of their “labor” by spending money on initiatives and making partnerships.
AI gave hyperscalers the excuse to spend hundreds of billions of dollars on data centers and buy a bunch of GPUs to go in them, because that, to the markets, looks like they’re doing something. By virtue of spending a lot of money in a frighteningly short amount of time, Satya Nadella received multiple glossy profiles, all without having to prove that AI can really do anything, be it a job or make Microsoft money.
Nevertheless, AI allowed CEOs to look busy, and once the markets and journalists had agreed on the consensus opinion that “AI would be big,” all that these executives had to do was buy GPUs and “do AI.”
We are in the midst of one of the darkest forms of software in history, described by many as an unwanted guest invading their products, their social media feeds, their bosses’ empty minds, and resting in the hands of monsters. Every story of its success feels bereft of any real triumph, with every literal description of its abilities involving multiple caveats about the mistakes it makes or the incredible costs of running it.
Generative AI exists for two reasons: to cost money, and to make executives look busy. It was meant to be the new enterprise software and the new iPhone and the new Netflix all at once, a panacea where software guys pay one hardware guy for GPUs to unlock the incredible value creation of the future.
Generative AI was always set up to fail, because it was meant to be everything, was talked about like it was everything, is still sold like it’s everything, yet for all the fucking hype, it all comes down to two companies: OpenAI, and, of course, NVIDIA.
NVIDIA was, for a while, living high on the hog. All CEO Jensen Huang had to do every three months was saying “check out these numbers” and the markets and business journalists would squeal with glee, even as he said stuff like “the more you buy the more you save,” in part tipping his head to the (very real and sensible) idea of accelerated computing, but framed within the context of the cash inferno that’s generative AI, seems ludicrous.
Huang’s showmanship worked really well for NVIDIA for a while, because for a while the growth was easy. Everybody was buying GPUs. Meta, Microsoft, Amazon, Google (and to a lesser extent Apple and Tesla) make up 42% of NVIDIA’s revenue, creating, at least for the first four, a degree of shared mania where everybody justified buying tens of billions of dollars of GPUs a year by saying “the other guy is doing it!”
This is one of the major reasons the AI bubble is happening, because people conflated NVIDIA’s incredible sales with “interest in AI,” rather than everybody buying GPUs. Don’t worry, I’ll explain the revenue side a little bit later. We’re here for the long haul.
Anyway, NVIDIA is facing a problem — that the only thing that grows forever is cancer.
On September 9 2025, the Wall Street Journal said that NVIDIA’s “wow” factor was fading, going from beating analyst estimates in by nearly 21% in its Fiscal Year Q2 2024 earnings to scraping by with a mere 1.52% beat in its most-recent earnings — something that for any other company, would be a good thing, but framed against the delusional expectations that generative AI has inspired, is a figure that looks nothing short of ominous.
Per the Wall Street Journal:
Already, Nvidia’s 56% annual revenue growth rate in its latest quarter was its slowest in more than two years. If analyst projections hold, growth will slow further in the current quarter.
[Content truncated due to length…]
You must log in or # to comment.