None of what I write in this newsletter is about sowing doubt or "hating," but a sober evaluation of where we are today and where we may end up on the current path. I believe that the artificial intelligence boom — which would be better described as a generative AI boom — is (as I've said before) unsustainable, and will ultimately collapse. I also fear that said collapse could be ruinous to big tech, deeply damaging to the startup ecosystem, and will further sour public support for the tech industry.
The reason I'm writing this today is that it feels like the tides are rapidly turning, and multiple pale horses of the AI apocalypse have emerged: “a big, stupid magic trick” in the form of OpenAI's (rushed) launch of its "o1 (codenamed: strawberry") model, rumored price increases for future OpenAI models (and elsewhere), layoffs at Scale AI, and leaders fleeing OpenAI. These are all signs that things are beginning to collapse.
As a result, I think it’s important to explain how precarious things are, and why we are in a trough of magical thinking. I want to express my concerns about the fragility of this movement and the obsessiveness and directionlessness that brought us here, and I want some of us to do better.
I also — and this, perhaps, is something I haven’t focused on as much as I should have — want to highlight the potential human cost of an AI bubble implosion. Whether Microsoft and Google (and the other big generative AI backers) slowly wind down their positions, or cannibalize their companies to keep OpenAI and Anthropic (as well as their own generative AI efforts) alive, I’m convinced that the end result will be the same. I fear tens of thousands of people will lose their jobs, and much of the tech industry will suffer as they realize that the only thing that can grow forever is cancer.
There won’t be much levity in this piece. I’m going to paint you a bleak picture — not just for the big AI players, but for tech more widely, and for the people who work at tech companies — and tell you why I think the conclusion to this sordid saga, as brutal and damaging as it will be, is coming sooner than you think.
Let's begin.
How Does Generative AI Survive?
As we speak, OpenAI — the ostensible non-profit that may soon become a for-profit — is currently raising a funding round at a valuation of at least $150 billion, and is expected to raise at least $6.5 billion — but potentially as much as $7 billion — led by Josh Kushner's Thrive Capital, with rumored participation from both NVIDIA and Apple. As I've explained in agonizing detail in the past, OpenAI will have to continue to raise more money than any startup has ever raised in history, in perpetuity, to survive.
Worse still, OpenAI also is trying to raise $5 billion in debt from banks “in the form of a revolving credit facility” according to Bloomberg, and the terms on revolving credit facilities tend to have higher rates of interest.
The Information also reports that OpenAI is in talks with MGX, a $100 billion investment fund backed by the United Arab Emirates to invest primarily in AI and semiconductor companies, as well as potentially raising from the Abu Dhabi Investment Authority. This should be the biggest warning sign that things are going poorly, because absolutely nobody raises from the UAE or Saudis because they want to. They’re the place you go if you need a lot of money and you’re not confident anybody else will give it to you.
Sidenote: as CNBC points out, one of the foundational partners of MGX, Mubadala, owns around $500m in Anthropic equity, which it acquired from the bankruptcy of FTX’s assets. I’m sure Amazon and Google are thrilled about that conflict of interest!
As I discussed at the end of July, OpenAI needs to raise at least $3 billion — but more like $10 billion to survive, as it is on course to lose $5 billion in 2024, a number that's likely to increase as more complex models demand more compute and more training data, with Anthropic CEO Dario Amodei predicting that future models may cost as much as $100 billion to train.
As an aside, a "$150 billion valuation" in this case refers to how OpenAI is pricing shares in the company for investors — though "shares" is even a fuzzy term in this case. For example, an investment of $1.5 billion at a $150 billion valuation in a normal company would get you "1%" of the company, though, as I'll get to, in OpenAI's case, things are a lot weirder.
OpenAI has already tried to raise at a $100 billion valuation earlier in the year, with some investors balking at the price tag, in part because (and I paraphrase The Information's Kate Clark and Natasha Mascarenhas) of a growing concern over the overvaluation of generative AI companies.
To get the round done, OpenAI may also convert from a non-profit to a for-profit entity, but the most confusing part of this story appears to be what investors are actually getting. Kate Clark of The Information reports that investors in the round are being told, and I quote, that they "don't get traditional equity for their cash...instead, they receive units that promise a slice of the company's profits — once it starts to generate them."
It's unclear whether a conversion to a for-profit entity would fix this problem, as OpenAI's bizarre non-profit-with-a-for-profit-arm corporate structure means that Microsoft has the rights to 75% of OpenAI's profits as part of its 2023 investment — though a shift to a for-profit structure could, theoretically, include equity. Nevertheless, investing in OpenAI gives you "profit participation units" (PPUs) in lieu of equity, and as Jack Raines wrote at Sherwood, "if you own OpenAI PPUs, and the company never turns a profit and you don’t sell them to someone who thinks OpenAI eventually will turn a profit, your PPUs are worthless."
Over the weekend, Reuters published a report that said any $150 billion valuation would be “contingent” on whether it’s able to rework its entire corporate structure — and, in the process, remove the cap on profits for investors, which are limited to 100x the original stake. This capped profit structure was introduced in 2019, with OpenAI saying any profits that exceed that ceiling would “be returned to the Nonprofit for the benefit of humanity.” The company has, in recent years, altered that rule to allow for a 20 percent yearly increase in the cap, starting from 2025.
Given OpenAI’s existing profit-sharing deals with Microsoft — not to mention its deep, deep unprofitability — any such returns are, at the best, theoretical. At the risk of sounding glib, 500% of nothing is still nothing.
Reuters also added that any transition to a for-profit structure (and thus, a higher valuation than its most recent $80bn) would also force OpenAI to renegotiate with existing investors, who would see their stakes diluted.
Separately, the Financial Times reports that investors have to "sign up to an operating agreement that states: "It would be wise to view any investment in [OpenAI's for-profit subsidiary] in the spirit of a donation," and that"OpenAI "may never make a profit," a truly insane thing to sign that makes any investor in OpenAI fully deserving of any horrible fate that follows such a ridiculous investment.
In effect, investors aren't getting a piece of OpenAI, or any kind of control over OpenAI, but rather shares in the future profits of a company that loses over five billion dollars a year, and will likely lose more in 2025, if it makes it that far.
OpenAI’s models and products — and we'll get into their utility in a bit — are deeply unprofitable to operate, with the Information reporting that OpenAI is paying Microsoft an estimated $4 billion in 2024 to power ChatGPT and its underlying models, and that's with Microsoft giving it a discounted $1.30-per-GPU-an-hour cost, as opposed to the regular $3.40 to $4 that other customers pay. This means that OpenAI would likely be burning more like $6 billion a year on server costs if it wasn’t so deeply wedded to Microsoft — and that's before you get into costs like staffing ($1.5 billion a year), and, as I've discussed, training costs that are currently $3 billion for the year and will almost certainly increase.
While The Information reported that OpenAI's revenue is $3.5 to $4.5 billion a year in July, The New York Times reported last week that OpenAI's annual revenues have "now topped $2 billion," which would mean that the end-of-year numbers will likely trend toward the lower end of the estimate.
In short, OpenAI is burning money, will only burn more money, and to continue burning more money it will have to raise money from investors that are signing a document that says "we may never make a profit."
As I've written about previously, OpenAI's other problem is that generative AI (and by extension the model GPT and the product ChatGPT) doesn't solve the complex problems that would justify its massive costs. It has massive, intractable problems as a result of these models being probabilistic, meaning that they don't know anything, they're just generating an answer (or an image, or a translation, or a summary) based on training data, something that model developers are running out of at an incredible pace.
Hallucinations — which occur when models authoritatively states something that isn't true (or in the case of an image or a video makes something that looks...wrong) — are impossible to resolve without new branches of mathematics, and while you might be able to reduce or mitigate them, their existence makes it hard for business-critical applications to truly rely on generative AI.
And even if they did, it isn't clear whether generative AI actually provides much business value at all. The Information reported last week that customers of Microsoft's 365 suite (which includes things like Word, Excel, PowerPoint and Outlook, and more importantly a number of business-focused software packages which in turn feed into consultancy services from Microsoft) are barely adopting its AI-powered "Copilot" products, with somewhere between 0.1% and 1% of its 440 million seats (that’s $30 to $50 per person) paying for the features. One firm testing the AI features is quoted as saying that "most people don't find it that valuable right now," and others saying that "many businesses haven't seen [breakthroughs] in productivity and other benefits" and they're "not sure when they will."
And how much is Microsoft charging for these inessential features? A whopping $30-a-seat on top of what you’re already paying, or as much as $50-a-month extra for “Copilots for Sales.” They’re effectively asking customers to double their spend - and by the way, that’s an annual commitment! - for products that don’t seem to be that helpful.
As a note: Microsoft is such a clusterfuck that it will likely require its own newsletter in the future.
This is the current state of generative AI - the literal leader in productivity and business software cannot seem to find a product that people will pay for, in part because the results are so mediocre, and in part because the costs are so burdensome that it’s hard to justify them. If Microsoft needs to charge this much, it’s either because Satya Nadella wants to hit $500 billion in revenue by 2030 (as revealed as part of a memo included in a public hearing over Microsoft’s acquisition of Activision Blizzard), the costs are too high to charge much less, or some combination of the two.
Yet the argument is almost always that the future of AI will blow us away - that the next generation of Large Language Models are just around the corner, and they’re going to be incredible.
Last week, we got our first real, definitive glimpse of what’s around that corner that future. And boy, was it underwhelming.
A Big, Stupid Magic Trick
OpenAI launched o1 — codenamed Strawberry — on Thursday night, with all the excitement of a dentist’s appointment. Across a series of tweets, Sam Altman described o1 as OpenAi’s “most capable and aligned models yet.” Though he conceded that o1 was “still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it,” he promised it would deliver more accurate results when performing the kinds of activities where there is a definitive right answer — like coding, math problems, or answering science questions.
That, by itself, is incredibly revealing — but we’ll expand on that later. First, let’s talk about how it actually works. I’m going to introduce a bunch of new concepts here, but I promise I won’t delve too deeply into the weeds. And if you actually want to read OpenAI’s explanation, you can find it here.
When presented with a problem, o1 breaks it down into individual steps that — hopefully — would lead to a correct answer in a process called Chain of Thought. It’s also a little easier if you consider o1 as two parts of the same model.
On each step, one part of the model applies reinforcement learning, with the other one (the model outputting stuff) “rewarded” or “punished” based on the perceived correctness of their progress (the steps in its “reasoning”), and altering its strategies when punished. This is different to how other Large Language Models work in the sense that the model is generating outputs then looking back at them, then ignoring or approving “good” steps to get to an answer, rather than just generating one and saying “here ya go.”
While this sounds like a seismic breakthrough, or even another step towards the highly-feted artificial general intelligence (AGI) — it isn’t — and you can tell by the fact that OpenAI opted to release o1 as its own standalone product, rather than a number update to GPT. It’s telling that the examples demonstrated by OpenAI — like math and science problems — are ones where the answer can be known ahead of time, and a solution is either correct or false, thus allowing the model to guide the Chain of Thought through each step.
You’ll notice OpenAI didn’t show the o1 model trying to tackle complex problems, whether mathematical or otherwise, where the solution isn’t known in advance. By its own admission, OpenAI has heard reports that o1 is more prone to hallucinations than GPT-4o, and is less inclined to admit when it doesn’t have the answer to a question when compared to other previous models. This is because, despite there being a model that checks its work, the work-checking part of the model is still capable of hallucinations.
According to OpenAI, it’s also — thanks to its Chain of Thought — more convincing to human users. Because o1 provides more detailed answers, people are more inclined to trust the outputs, even when they’re completely wrong.
If you think I’m being overtly hard on OpenAI, consider the ways in which the company has marketed o1. It has described the reinforcement training process as “thinking” and “reasoning,” when, in fact, it’s making guesses, and then guessing on the correctness of those guesses at each step, where the end destination is often something that can be known in advance.
It's an insult to people — actual human beings — who, when they think, are acting based on a complex tapestry of factors: from their experience, to the knowledge they’ve accumulated across their lifetimes, to their brain chemistry. While we may too “guess” about the correctness of each when reasoning through a complex problem, our guesses are based on something concrete, rather than an inelegant mathematical flail, as with o1.
And by God, it’s expensive.
Pricing for o1-preview is $15 per million input tokens and $60 per million output tokens. In essence, it’s three times as expensive as GPT-4o for input and four times as expensive for output. There is, however, a hidden cost. Data scientist Max Woolf reports that OpenAI’s “reasoning tokens” — the outputs it uses to get you your final answer — are not visible in the API, meaning that not only are o1’s prices higher, the nature of the product requires it to charge you more often. All of the things it generates to “consider” an answer (to be clear, this model is not “thinking”) are also charged for, making complex answers for things like coding likely incredibly expensive.
So, let’s talk about accuracy. On Hacker News — the Reddit-style site owned by Sam Altman’s former alum, Y Combinator — one person complained about o1 hallucinating libraries and functions when presented with a programming task, and making mistakes when asked questions where the answer isn’t readily available on the Internet.
On Twitter, Henrik Kniberg, a startup founder and former game developer, asked o1 to write a python program that multiplied two numbers and then calculated the expected output of said program. While it wrote the code correctly (although, said code could have been more succinct, using one line instead of two), the actual result was wildly incorrect. Karthik Kannan, himself a founder of an AI company, tried a programming task on o1, where it also hallucinated a non-existent command for the API he was using.
Another person, Sasha Yanshin, tried to play a game of chess with o1 and it hallucinated an entire piece onto the board. And then lost.
Because I’m a little shit, I also tried asking o1 to list the number of states with “A” in the name. After contemplating for eighteen seconds, it provided the names of 37 states, including Mississippi. The correct number, by the way, is 36.
When asked to list the states with the letter “W” in the name, it pondered for eleven seconds and included North Carolina and North Dakota.
I also asked o1 to count the number of times the letter “R” appears in the word strawberry — its pre-release codename. It said two.
OpenAI claims that o1 “performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology.” Just not in geography, it seems. Or basic elementary-level English language tests. Or math. Or programming.
This is, I should note, the big stupid magic trick I predicted in a previous newsletter. OpenAI is shoving Strawberry out the door as a means of proving to investors (and the greater public) that the AI revolution is still here, and what they have is a clunky, unexciting and expensive model.
Worse still, it's kind of hard to explain why anybody should give a shit about o1. While Sam Altman will likely try and trump up its "reasoning abilities," what people (such as those with the money to keep bankrolling him) will see is the 10-20 second wait time for answers, which still have issues with basic factual accuracy, and the complete lack of any exciting new features.
Nobody gives a shit about "better" answers anymore — they want it to do something new, and I don't think that OpenAI has any idea how to make that happen. Altman’s limp attempts to anthropomorphize o1 by making it “think” and use “reasoning” are obvious attempts to suggest this is somehow part of the path to AGI, but even the most staunch AI advocates are having trouble getting excited.
In fact, I'd argue that o1 shows that OpenAI is both desperate and out of ideas.
The prices are not decreasing, the software is not becoming more useful, and the "next generation" model that we’ve been hearing about since last November has turned out to be a dud. These models are also desperate for training data, to the point that almost every Large Language Model has ingested some sort of copyrighted material. This desperation led Runway, one of the largest generative video players, to create a “company-wide effort” to collect thousands of YouTube videos and pirated content to train their models, and a federal lawsuit filed in August alleges that NVIDIA did the same to numerous creators to train its "Cosmos" AI software..
The legal strategy at this point is sheer force of will, hoping that none of these lawsuits reach the point where any legal precedent is set that might define training these models as a form of copyright infringement, which is exactly what a multidisciplinary study out of the Copyright Initiative recently found was the case.
These lawsuits are progressing, with a judge in August granting plaintiffs further claims of copyright infringement against Stability AI and DeviantArt (which uses its models), as well as both copyright and trademark infringement claims against Midjourney. If any of these lawsuits prevail, it would be calamitous for OpenAI and Anthropic, and even more so for Google and Meta, whose Gemini and Llama models use datasets including millions of artists' work, mostly because it's "virtually impossible" for AI models to forget training data, meaning that they'd need to be retrained from scratch, which would cost billions of dollars and dramatically reduce their efficacy at tasks they're already not particularly good at.
I am deeply concerned that this entire industry is built on sand. Large Language Models at the scale of ChatGPT, Claude, Gemini and Llama are unsustainable, and do not appear to have a path to profitability due to the compute-intensive nature of generative AI. Training them necessitates spending hundreds of millions — if not billions — of dollars, and requires such a large amount of training data that these companies have effectively stolen from millions of artists and writers and hoped they'd get away with it.
And even if you put these problems aside, generative AI and its associated architectures do not appear to do anything revolutionary, and absolutely nothing about the generative AI hype cycle has truly lived up to the term "artificial intelligence." At best, generative AI seems capable of generating some things correctly sometimes, summarizing documents, or doing research at an indeterminate level of "faster." Microsoft's Copilot for Microsoft 365 claims to have "thousands of skills" that give you "infinite possibilities for enterprise," yet the examples it gives involve generating or summarizing emails, "starting a presentation using a prompt" and querying Excel spreadsheets — useful, perhaps, but hardly revolutionary.
We’re not “in the early days.” Since November 2022, big tech has spent over $150 billion in combined capex and investments into their own infrastructure and budding AI startups, as well as in their own models. OpenAI has raised $13 billion, and can effectively hire whoever they want, as can Anthropic.
The result of an industry-wide Marshall Plan to get generative AI off the ground has resulted in four or five near-identical Large Language Models,the world's least-profitable startup, as well as thousands of overpriced and underwhelming integrations.
Generative AI is being sold on multiple lies:
- That it's artificial intelligence.
- That it's "going to get better."
- That it will become artificial intelligence.
- That it is inevitable.
Putting aside terms like "performance" — as they are largely used as a means of generating things "accurately" or "faster" rather than being good at stuff — Large Language Models have effectively plateaued. "More powerful" never seems to mean "does more," and "more powerful" oftentimes means "more expensive," meaning that you've just made something that doesn't do more but does cost more to run.
If the combined forces of every venture capitalist and big tech hyperscaler have yet to come up with a meaningful use case that lots of people will actually pay for, there isn't one coming. Large Language Models — and yes, that's where all of these billions of dollars are going — are not going to magically sprout new capabilities because big tech and OpenAI sunk another $150 billion into them. Nobody is trying to make these things more efficient, or at the very least nobody has succeeded in doing so. If they had, they'd be shouting it from the rooftops.
What we have here is a shared delusion — a dead-end technology that runs on copyright theft, one that requires a continual supply of capital to keep running as it provides services that are, at best, inessential, sold to us dressed up as a kind of automation it doesn't provide, that costs billions and billions of dollars and will continue to do so in perpetuity. Generative AI doesn’t run on money (or cloud credits), so much as it does on faith. The problem is that faith — like investor capital — is a finite resource.
My concern is that I believe we’re in the midst of a subprime AI crisis, where thousands of companies have integrated generative AI at prices that are far from stable, and even further from profitable.
Almost every "AI-powered" startup that uses LLM features is based on some combination of GPT or Claude. These models are built by two companies that are deeply unprofitable (Anthropic is on course to lose $2.7 billion this year), and that have pricing designed to get more customers rather than make any kind of profit. OpenAI, as mentioned, is subsidized by Microsoft — both in the "cloud credits" it received and the preferential pricing Microsoft offers — and its pricing is entirely dependent on Microsoft's continued support, both as an investor and a services provider, a problem that Anthropic faces with its deals with Amazon and Google.
Based on how unprofitable they are, I hypothesize that if OpenAI or Anthropic charged prices closer to their actual costs, there would be a ten-to-a-hundred-times increase in the price of API calls, though it's impossible to say how much without the actual numbers. However, let's consider for a fact that the numbers reported by The Information estimate that OpenAI's server costs with Microsoft will be $4 billion in 2024 — which, I add, are over two-and-a-half-times cheaper than what Microsoft charges others — and then consider that OpenAI still loses over five billion dollars a year.
OpenAI is more than likely charging only a small percentage of what it likely costs to run its models, and can only continue to do so if it is able to continually raise more venture funding than has ever been raised and continue to receive preferential pricing from Microsoft, a company that recently mentioned that it considers OpenAI a competitor. While I can't say for certain, I would think it's reasonable to believe that Anthropic receives similarly-preferential pricing from both Amazon Web Services and Google Cloud.
Assuming that Microsoft gave OpenAI $10 billion of cloud credits, and it spent $4 billion on server costs and, let's say, $2 billion on training — costs that are both sure to increase with the new o1 and “Orion” models — OpenAI will either need more credits or will have to start paying actual cash to Microsoft sometime in 2025.
While it might be possible that Microsoft, Amazon and Google extend their preferred pricing indefinitely, the question is whether these transactions are profitable for them. As we saw following Microsoft’s most recent quarterly earnings, there is growing investor concern over the CapEx spending required to build the infrastructure for generative AI, with many voicing skepticism about the potential profitability of the technology.
And what we really don’t know is how unprofitable generative AI is for the hyper-scalers, because they bake those costs into other parts of their earnings. While we can’t know for sure, I imagine if this stuff was in any way profitable, they’d be talking about the revenue they were receiving from it.
They’re not.
The markets are extremely skeptical of the generative AI boom, and NVIDIA CEO Jensen Huang had no real answers about AI's return on investment, leading to an historic $279 billion drop in NVIDIA's market cap in a single day. This was the biggest rout in US markets history. The total value lost is the equivalent of nearly five Lehman Brothers at its peak value. And while the comparison ends there — Nvidia isn’t even close to failing, and even if it did, the systemic impact wouldn’t be anywhere near as damaging — it’s still an insane amount of money, and an indicator of the distorting power of AI on the markets.
At the beginning of August, Microsoft, Amazon, and Google all took a beating from the markets for their massive capital expenditures related to AI, and all three of them will face the wheel next quarter if they can't show a significant increase in revenue from the combined $150bn (if not more) that they've put into new data centers and NVIDIA GPUs.
What's important to remember here is that other than AI, big tech is out of ideas. There are no more hyper-growth markets left, and as firms like Microsoft and Amazon begin to show signs of declining growth, so too does their desperation to show the markets that they've still got it. Google, a company almost entirely sustained by multiple at-risk monopolies in search and advertising, also needs something new and sexy to wave in front of the street — except none of this is working because the products aren't useful enough and it appears most of its revenue comes from companies "trying out" AI and then realizing that it isn't really worth it.
At this point, there are two eventualities:
- Big Tech realizes that they've gotten in too deep, and out of a deep fear of pissing off Wall Street chooses to reduce capital expenditures related to AI.
- Big Tech, desperate to find a new growth pig, decides instead to cut costs to sustain their ruinous operations, laying off workers and reallocating capital from other operations as a means of sustaining the generative AI death march.
It's unclear which will happen. If Big Tech accepts that generative AI isn't the future, they don't really have anything else to wave at Wall Street, but could do their own version of Meta's "year of efficiency," reducing capital expenditures (and laying off a bunch of people) while also promising to "slow down investment" by some degree. This is the most likely path for Amazon and Google, who, while desperate to make Wall Street happy, still have their own profitable monopolies to point to — for now, at least.
Nevertheless, there needs to be actual revenue growth from AI in the next few quarters, and it has to be material rather than some vague thing about how AI is a "maturing market" or "annualized run rates," and said material contribution will have to be magnitudes higher if capital expenditures have increased along with it.
I don't think it's going to be there. Whether or not it's Q3 or Q4 2024, or even Q1 2025, Wall Street will begin punishing big tech for the sin of lust, and said punishment will savage these companies far more harshly than NVIDIA, which, despite Huang's bluster and empty platitudes, ia the only company in the market that can actually point to how AI is increasing revenue.
I worry, somewhat, that option 2 is far more possible: that these companies are deeply committed to the idea that "AI is the future," and their cultures are so thoroughly disconnected from the creation of software that solves the problems that real people face that they'll burn the entire company to the ground. I deeply worry about the prospect of mass layoffs being used to fund the movement, and nothing about the last few years makes me think they’ll do the right thing and walk away from AI.
Big tech has become thoroughly poisoned by management consultants — Amazon, Microsoft and Google are all run by MBAs — and in turn have surrounded them with similarly-specious ghouls like Google's Prabhakar Raghavan, who chased out the people that actually built Google Search so that he could run it.
These people do not really face human problems, and have created cultures dedicated to solving the imagined problems that software can fix. Generative AI must seem kind of magical when your entire life is either being in a meeting or reading an email, and I imagine the winning mindset of Satya Nadella mostly comes down to "having the tech people sort it out." Sundar Pichai could've ended the entire generative AI boom in an instant if he simply looked at Microsoft's investment in OpenAI and laughed — but no, he had to follow, because none of these men have any actual ideas, and these companies are not run by people that experience problems, let alone people that might actually know how to fix them.
They’re also desperate, and things have never gone like this for them before, other than when Meta burned billions on the metaverse. Yet this situation is so much bigger and uglier because they have put so much money and so thoroughly welded AI to their companies that removing it will be both embarrassing and damaging to their stocks, and a tacit admission that all of this was a waste.
All this could’ve been stopped earlier if the media had actually held them accountable. This narrative was sold through the same con as previous hype-cycles, with the media assuming that these companies would “just work it out,” despite the fact that it was blatantly obvious they wouldn’t. Think I’m a doomer? Well, answer me this: what's the plan here? What does generative AI do next? Is your answer that they'll "work it out," or that they "have something behind the scenes that's incredible,” you’re an unwitting participant in a marketing operation.
Sidebar: No, really, we have to stop being conned by this shit. When Mark Zuckerberg claimed we were about to enter the metaverse, large swaths of the media — the New York Times, The Verge, CBS News and CNN to name a few — humored an idea that was clearly flawed, one that looked like shit and was peddled using outright lies about the future. It was so obviously nothing other than a shitty VR world, yet The Wall Street Journal was still — over six months into the hype-cycle when it was blatantly obvious that the metaverse was bullshit — talking about it as “the future vision for the internet.” And it happened with crypto, Web3, and NFTs too! The Verge, The New York Times, CNN, CBS News — all again participating in pumping technology that so clearly didn’t do anything — though I should add that when I say “The Verge” here, I really mean Casey Newton, who has a great reputation despite being on his third straight bag-pumping, claiming in July that “owning one of the most powerful LLMs could provide the company with a basis for all manner of money-making products” while discussing a technology that only loses money and has yet to provide one truly useful, lasting product.
I believe that, at the very least, Microsoft will begin reducing costs in other areas of its business as a means of helping sustain the AI boom. In an email shared with me by a source from earlier this year, Microsoft's senior leadership team requested (in a plan that was eventually scrapped) reducing power requirements from multiple areas within the company as a means of freeing up power for GPUs, including moving other services' compute to other countries as a means of freeing up capacity for AI.
On the Microsoft section of anonymous social network Blind (where you're required to verify that you have a corporate email of the company in question), one Microsoft worker complained in Mid-December 2023 of "AI taking their money," saying that "the cost of AI is so much that it is eating up pay raises, and that things will not get better." In mid-July, another shared their anxiety about how it was apparent to them that Microsoft had "[a] borderline addiction to cut costs in order to fund Nvidia's stock price with operational cash flows," and that doing so had "damaged Microsoft's culture deeply."
Another added that they believe that "Copilot is going to ruin [Microsoft] in FY25," adding that "The FY25 Copilot focus is going to massively fall in FY25," and that they knew of "big copilot deals in [their] country that have less than 20% usage after almost a year of PoC, cuts and layoffs," adding that "Corp risked too much" and that Microsoft's "huge AI investments are not going to be realized."
While Blind is anonymous, it's hard to ignore the fact that there are many, many posts that tell a tale of cultural cancer in Redmond, with disconnected senior leadership that only funds projects if they have AI taped onto them. Multiple posts lament Satya Nadella's "word salad" approach, and complain of a lack of bonuses or upward mobility in an organization focused on chasing an AI boom that may not exist.
And at the very least, there's a deep cultural sadness to the company, with the many posts I've seen oscillating between "I don't like working here," "I don't know why we're putting so much into AI," and "get used to it, because Satya doesn't care."
Buried in The Information's article about the lack of adoption of Microsoft's Office AI features is one particularly worrying thought about the actual utilization of Microsoft's massive data center spend:
Other signs back up those estimates: Around March of this year, Microsoft had set aside enough server capacity in its data centers for 365 Copilot to handle daily users of the AI assistant in the low millions, according to someone with direct knowledge of those plans. It couldn’t be learned how much of that capacity was in use at the time.
Based on The Information's estimates, Microsoft has somewhere between 400,000 and 4 million users for its Office Copilot features, meaning that there's a decent chance that Microsoft has built out capacity that isn't getting used.
Now, one could argue that they're building with the belief that this product category will grow, but here's another idea: what if it doesn't? What if — and this is a crazy idea — Microsoft, Google and Amazon built out these massive data centers to capture demand that may never arrive?I realize I sound a little crazy suggesting this, but back in March I made the point that I could find no companies that had integrated generative AI in a way that has truly benefited their bottom line, and just under six months later, I'm still looking. The best that big companies appear to have is stapling AI functionality onto existing products and hoping that lets them sell more of them, something that doesn't appear to be working for anybody, or, like Microsoft, offering "AI upgrades" that don't seem to provide any real business value.
While there may be companies "integrating AI" that are driving some degree of spend on Microsoft Azure, Amazon Web Services and Google Cloud, I hypothesize a lot of this demand is driven by investor sentiment, with companies “investing in AI” to keep the markets happy rather than any cost/benefit analysis or actual utility.
Nevertheless, these companies have spent a great deal of time and money baking in generative AI features to their products, and I hypothesize they will face one of a few different scenarios:
- After developing and launching these features, they find customers don't want to pay for them, as Microsoft has found with its 365 Copilot, and if they can't find a way to make them pay for it now, during the AI boom, they're going to be hard pressed to once people's bosses stop asking them to "get in on AI."
- After developing and launching these features, they can't seem to find a way to get users to pay for them, or at least pay extra for them, which means that they'll have to bake AI into their existing products with no extra margin, effectively turning them into a parasite that eats away at revenue.
- This was a point raised by Jim Covello in Goldman Sachs' generative AI report — that if AI's benefits are efficiency improvements (such as being able to analyze documents faster), that's something that your competition can do too. If you look carefully, almost every generative AI integration is the same — some sort of copilot that can answer questions for customers or internally (Salesforce, Microsoft, Box), content creation (Box, IBM), code generation (Cognizant, Github Copilot), and soon "agents," the latest bullshit that will mean "customizable chatbots that can connect to other parts of websites."
- This point also illustrates one of the biggest problems with generative AI — that, while "powerful" in some vague way, said power mostly boils down to "generates stuff based on other stuff" rather than any "intelligence." This is also why so many of the AI pages on companies' websites (see Cognizant's) are gobbledygook. Their biggest sales pitch is "uhhhh, you work it out!"
What I fear is a kind of cascade effect. I believe that a lot of businesses are "trying" AI at the moment, and once those trials end (Gartner predicts that 30% of generative AI projects will be abandoned after their proof of concepts by end of 2025), they'll likely stop paying for the extra features, or stop integrating generative AI into their companies' products.
If this happens, it will reduce the already-turgid revenue flowing to both hyper-scalers providing cloud compute for generative AI applications and Large Language Model providers like OpenAI and Anthropic, which could, in turn, create more price pressure on these companies as their already-negative margins sour further. At that point, OpenAI and Anthropic will almost certainly have to raise prices, if they haven’t already done so.
Even though Big Tech could keep funding this boom — after all, they’re pretty much entirely responsible for its existence — that won’t help the startups that won't be able to afford to run their companies, after getting used to discounted rates. Though there are cheaper alternatives like independent providers running Meta’s Llama models, it’s hard to believe that they don’t face exactly the same profitability problems as the hyper-scalers.
It’s also important to note that the hyper-scalers are also terrified of pissing off Wall Street. While they could theoretically (as I fear) start doing layoffs and other cost-cutting measures as a means of improving margins, these are short-term solutions that only really work if they're able to somehow shake some money from the barren Generative AI tree.
In any case, it's time to accept that the money isn't there. It's time to stop and take stock of the fact that we're in the midst of the third delusional epoch of the tech industry. Yet, unlike cryptocurrency and the metaverse, everybody has joined the party and decided to burn money pursuing an unsustainable, unreliable, unprofitable, and environmentally-destructive boondoggle sold to customers and businesses as "artificial intelligence" that will "automate everything" without it ever having a path to do so.
So, why does this keep happening? Why have we had movement after movement — cryptocurrency, the metaverse, and now generative AI — that doesn’t seem like it was actually made for a real person?
Well, it’s the natural result of a tech industry that’s become entirely focused on making each customer more valuable rather than providing more value to the customer. Or, for that matter, actually understand who their customers are and what they need.
The products you’re being sold today almost certainly try to wed you to a particular ecosystem — one owned by Microsoft, Apple, Amazon or Google, as a consumer at least — and, in turn, increase the burden of leaving said ecosystem. Even cryptocurrency — ostensibly a “decentralized” technology — quickly abandoned its free-wheeling libertarian ideas and sought to consolidate users on one of a few big platforms like Coinbase, OpenSea, Blur or Uniswap, all backed by the same venture capital firms (like Andreessen Horowitz). Rather than being flag bearers for a new, radically independent online economic system, they were all only able to scale through the funds and connections of the same people that funded every other recent era of the internet.
The metaverse, while a total sham, was an attempt by Mark Zuckerberg to own the next iteration of the internet, one where the dominant platform is “Horizon.” And I’ll get to generative AI in a minute.
Everything is about further monetization — about increasing the dollar-per-head value of each customer, be it through keeping them doing stuff on the platform to show them more advertising, upselling them new features that are only kind of useful, or creating some new monopoly or oligopoly where only those with the massive warchests of big tech can really play — and very little is about delivering real value. Real utility.
Generative AI is so exciting (at least, for a specific type of person) because Big Tech sees it as the next great monetizer — a means of creating a toll on every single product from consumer tech to the enterprise, because the majority of generative compute goes through either OpenAI or Anthropic, which in turn flows back to either Microsoft, Amazon or Google, creating cloud compute revenue for them to continue expressing growth. The biggest innovation here isn’t what Generative AI does, or can do, but rather the creation of an ecosystem that’s hopelessly dependent upon a handful of hyperscalers, and has no prospect of ever shaking its dependence.
Generative AI may not be super useful, but it is really easy to integrate into stuff and make “new things” happen, creating all sorts of new things that a company could theoretically charge for, both for consumer apps and enterprise software companies that make millions (or billions) upselling as many customers as possible. Sam Altman was smart enough to realize that the tech industry needed a new “thing” — a new technology that everybody could take a piece of and sell — and while he might not really understand technology, Altman understands the growth-lust of the larger economy, and productized transformer-based architecture as something that everybody could sell, a magical tool that could plug into most things and make something different.
The problem is that the desperation to integrate generative AI everywhere has shined a light on how disconnected these companies are from actual consumer needs, or even running functioning businesses. Just “doing new stuff” has kind of worked for 20 years, in the sense that simply launching new stuff and forcing salespeople to hock it was enough to keep growth going, to the point that the tech industry’s leaders have bought into a truly poisonous and deeply unprofitable business.
The people running these companies — almost exclusively MBAs and management consultants that have never built a product or a technology company from the ground up — either don’t understand or don’t care that generative AI has no path to profitability, likely assuming it will naturally become profitable like Amazon Web Services (which took nine years to do so) did, despite these being two very, very different things. Things have “just worked out” in the past, so why wouldn’t they today?
I mean, besides the fact that higher interest rates have dramatically reshaped the venture capital markets, reducing VC coffers and shrinking fund sizes. And the fact that sentiment towards tech has never been this negative. And the myriad of other factors why 2024 is nothing like 2014, which are too lengthy to name in a newsletter that’s already over 8,000 words long. Those old chestnuts.
The really worrying part is that other than AI, many of these companies don’t seem to have any other new products. What else is there? What other things do they have to grow their companies? What else do they have?
Nothing. They’ve got nothing. And that, really, is the problem, because when it fails, the effects will invariably cascade down to other companies in the tech space.
Every major tech player — both in the consumer and enterprise realm — is selling some sort of AI product, integrating either one of the big Large Language Models or their own, invariably running cloud compute on one of the big tech players’ systems. On some level, every single one of these companies is dependent on big tech’s willingness to subsidize the entire industry.
I hypothesize a kind of subprime AI crisis is brewing, where almost the entire tech industry has bought in on a technology sold at a vastly-discounted rate, heavily-centralized and subsidized by big tech. At some point, the incredible, toxic burn-rate of generative AI is going to catch up with them, which in turn will lead to price increases, or companies releasing new products and features with wildly onerous rates — like the egregious $2-a-conversation rate for Salesforce’s “Agentforce” product — that will make even stalwart enterprise customers with budget to burn unable to justify the expense.
What happens when the entire tech industry relies on the success of a kind of software that only loses money, and doesn’t create much value to begin with? And what happens when the heat gets too much, and these AI products become impossible to reconcile with, and these companies have nothing else to sell?
I truthfully don’t know. But the tech industry is building toward a grotesque reckoning with a lack of creativity enabled by an economy that rewards growth over innovation, monopolization over loyalty, and management over those who actually build things.
