CZM Rewind: The Case Against Generative AI (Part 2)

to it. Having written hundreds of thousands of words this year about the AI bubble, so many of the arguments I've made and the secrets I've exposed are contained in their own discreet little episodes or newsletters. This is my series to consolidate all of the information I've put out there in one place, and I want to make it

makes sense to anyone who listens to him. I want anyone, even who someone who doesn't even know that much about AI, to listen to the audience I've been making for the past three years, to understand why things are dire and to feel the same alarm I'm feeling, or at least understand why I'm alarmed. Because I don't like to tell you how you feel. Old school bit of FEEDBACKERD got from a listener once, and I appreciate that to this day.

Now today I'll make the case the 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 headliner problem I simply do not have. This industry is effectively in service of two companies, open Ai and Nvidia, who pump headlines out through endless contracts between them or subsidiaries or investments to give the illusion

of activity. Open Ai has now promised over four hundred billion dollars in the next four years, though honestly they might owe about a trillion dollars with all the data centers they're signed up for. All of these are egregious sums for a company that have already forecasted billions in losses with no clear explanation as to how it will afford any of this beyond we need more money and they're vague hope that there's another soft bank or Microsoft

waiting in the wings to swoop in and save the day. Now, I'm going to walk you through where I see this industry today and why I see no future for it beyond a horrible, fiery car rack. While everybody reasonably harps on about hallucinations, which to remind you, is when a model authoritativety states something that isn't true, the truth of why that's bad is far more complex and actually far

worse than it seems. You cannot rely on a large language model to do what you want, even though its highly tuned models on the most expensive and intricate platforms can't actually be relied upon to do exactly what you want. And I know some people might say, well, yes, they do every time, one hundred percent of the time. 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 the way reminiscent of a human. But it's not thinking. They don't think. They have no consciousness.

They literally you ask them something and they break down what the prompt might mean and then choose it's not thinking, 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, as I said,

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 will generate the right thing, but in an inefficient and over verbos 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, let alone every time. Model companies have successfully convinced everybody that the issue is that users are prompting the models wrong and that the

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 floor in how these fucking things work. Large language models are also uniquely expensive. Many mistakenly try and claim that this is like the dot com boom or uber, but the basic unique economics of generative AI

are insane. Providers must purchase tens or hundreds of thousands of GPUs, each costing fifty thousand to seventy thousand apiece, and the hundreds of millions or billions of dollars of infrastructure that goes around them are so expensive and hard to install, and that's without mentioning things like staffing or construction, or power or water or even permitting. Then you turn

them on and immediately they start losing new money. Despite hundreds of billions of GPUs sold, nobody seems to actually make any of it, other than Video of course, 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. Now, if you're an eager listener, I would love to hear from you on one question, then this is just something that's been bouncing around my head. Super Micro is in Video a customer of super Micro.

They're a custom in super Micro is a huge customer of Video. I've read something like seventy percent of the cost of goods sold is buying GPUs. But I read that in Video as a customer of them. But I can't find anything else. Reach out easy at better Offline dot com if you've got any thoughts there anyway, But back to those resourers. This arrangement works out great for Jensen Wang, the CEO of in Video, and terribly for

everybody else. Today, I'm gonna explain the insanity of the situation we'll find ourselves in and why I continue to do this work underterred. The bubble has entered its most pornographic, aggressive and destructive stage, where the more obvious it becomes that we're all cooked here in AI land, the more ridiculous the GENERATIVAI industry will act a dark juxtaposition against every new study that says generative AI does not work, or new story about chat gpd's uncannyability to activate mental

illness in people. And we're going to start looking at one company in Video which now dominates the stock market and has taken extraordinary and dangerous measures to sustain growth. That is, to any sane person, completely unsustainable and unrealistic on every level. But let's start simple. Nvidia is a hardware company that sells GPUs, including 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 A one hundred, h one hundred, H two hundred, and more modern GPUs like the Blackwell Series B two hundred and GB two hundred, which combines two GPUs with an Nvidia CPU. This is all complex sounding,

but I want you to have the groundwork. These GPUs cost anywhere from fifty to seventy thousand dollars and require tens of thousands of dollars more of infrastructure networking to cluster these server acts of GPUs together to provide compute and massive cooling systems to deal with the massive amounts of heat they produce, as well as 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. GPUs itself is likely the most expensive single it and 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 uses. Well, all this crap adds up. I've mentioned in Video because it has a virtual monopoly in this space, Genera IVAI effectively requires in video GPUs, in part because it's the only company really making the kinds of high powered cards that GENERAIVAI demands, and because in Video created something called Kuda. 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 some open source alternatives as well as alternatives from Intel with its Arc GPUs and AMD, in Vidia's main rival in the consumer space, these aren't nearly as mature or feature Richkuda's been around for ten fifteen years now, they really knew what they were doing. They also have bought a company called Melanox, which did the high speed

networking back in twenty nineteen, I think, for six billion dollars. Anyway, due to the complexities of AI models, one cannot just stand up a few of these GPUs 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 and the hundreds of millions or billions of dollars, especially considering they require completely different data center architecture to make them run. You've probably read a bunch of stuff

about cryptominers turning into AI data center providers. These crypto data centers have to be knocked down and replaced. They you can't just put the same GPUs in isn't going to work, and with the new Blackwell ones, the brand new ones, and then the Rubens following them, same deal. A common request like asking a generative AI model to pass through thousands of lines of code and make a change or in addition, may use multiples of these fifty

thousand dollars GPUs at the same time. And so if you aspire to serve thousands or millions of concurrent users, you need to spend big, really really really big. It's these factors, the vendor lock in the ecosystem and the fact that generative AI really 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 the executive job is so vague, they can telegraph the value of their labour by spending money on initiatives and partnerships and stratagem. 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, Satchnadella 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

or plug AI within their own software profitts. But really it was just jump on the big stupid asshole train. We are in the midst of one of the darkest forms of software in history, described by many as unwanted guest, invading their products, their social media feeds, their bosses, empty minds,

and resting in the hands of monsters. Every story of AI's 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 in generator of AI really 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 I phone and the new Netflix all at once, a panacea where the software guys pay one hardware guy for GPUs

to unlock the incredible value creation of the future. In many ways, Jenera iv AI was always set up to fail because it was meant to be Everything. Was talked about like it was everything. It's still sold like it's everything. Yet for all the fucking hype, it comes down to two companies, Open Ai and Nvidia, and in Video was

for a while living high on the hog. All CEO Jensen Huang had to do every three months would say, check out these numbers and the markets and business journalist would squeer 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 frame within the context of the cash inferna that's generated AI, and it all seems kind

of fucking ludicrous. Huang showmanship worked really well for Invidia 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 made up forty two percent of Invidia'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 by saying the

other guy's doing it. This is one of the major reasons the AI bubble is happening, because people conflated in Vidia's incredible sales with interest in AI rather than everybody buying GPUs at once. Don't worry, I'll explain the revenue side a little bit later. We're here for the long haul. Sit down, get comfort. You're gonna need to be anyway, Invidia is now facing a big problem that the only

thing that grows forever is cancer. On September ninth, twenty twenty five, The Wall Street Journal said that in Vidio's wow factor was fading, going from beating analyst estimates by nearly twenty one percent in its fiscal year Q two twenty twenty four earnings to scraping by with a pathetic measly one point five to two percent beat in its most recent earnings, something that for any other company would be a good thing because they made so much money,

But framed against the delusional expectations that generative AI has inspired, well, the figure looks nothing short of ominous. I quote the Wall Street Journal. Already, in Vidia's fifty six percent 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. In any other scenario, fifty six percent year of a year growth would lead to an abundance of dom perignon and one signing hundreds

of boobs. But this is in video, and that's just not good enough. Back in February twenty twenty four, in Video was booking two hundred and sixty five percent year of a year growth but in its February twenty twenty five earnings in Video only grew by a measly per thetic disgusting seventy eight percent year of a year. I'm being sarcastic, of course. It isn't so much that in Vidia isn't growing, but to grow year over year at

the rates that people expect is insane. Life was a lot easier when in Video went from six point oh five billion dollars in revenue in Q four fiscal year twenty twenty five to twenty two billion dollars in revenue in Q four fiscal year twenty twenty four. But for it to grow even fifty five percent year of a year from Q two FY twenty twenty six, I'm just going to truncate that now, which is forty six point

seven billion dollars to Q two twenty twenty seven. That would require them to make seventy two point three eight five billion dollars in revenue in the space of three months, mostly from selling GPUs, which make up about eighty eight percent of its revenue. Just want to be clear that in a year they would have to make seventy two billion dollars just selling pretty much GPUs and the associated hardware in the space of three months. It's it's insane.

This is this is really that it's too much. It's too much to expect. And this, by the way, we would put in video in the ballpark of Microsoft, who made seventy six billion dollars in their last quarterly earnings, and within the neighborhood of Apple, who made ninety four billion dollars in their last quarter of earnings. And they would do this predominantly making money in an industry that a year and a half ago barely made the company six billion dollars in a quarter. And the market needs

in Vidia to perform. They must, they must, as the company makes up seven to eight percent of the value of the S and P five hundred. It's not enough for Invidia to be wildly profitable, or to having a monopsony on selling GPUs, or for it to have effectively ten x their stock in a few years. No, no, no, more, more, more,

always more. Number must go up. It must continue to grow at the fastest rate of anything ever, making more and more money, selling more and more of these GPUs to a small group of companies that immediately start losing money. The moment they plugged them in. It's not brilliant, is it.

While a few members of the Magnificent Seven could be depended on to funnel tens of billions of dollars into a furnace each quarter, there were limits even for companies like Microsoft, which had brought over four hundred and eighty

five thousand g us in twenty twenty four alone. To take a step back about how people actually make money from buying these GPUs, companies like Microsoft, Google, and Amazon make their money by either selling access to large language models that people incorporate into their products, or by renting out servers full of those GPUs to run influence the thing, to generate the output, or train AI models for companies that develop and market their models themselves, namely Anthropic and

Open AI, with some smaller competitors that don't really matter that latter revenue stream. Renting out GPUs is where Jensen Wong found a solution to that horrible eternal growth problem the neo cloud, namely companies like core Weave, Lambda, and Nebius. Now,

these businesses are fairly straightforward. They own or lease data centers that they then feel full of servers that are full of Nvidio GPUs, which they then rent out on an hourly basis to customers either on a per GPU basis, are in large badges for large customers who guarantee they'll use a certain amount of compute and sign up for a long term agreement for moderan an era a time

couple years perhaps these larger commitments. A neocloud is a specialist cloud compute company that exists only to provide access to GPUs for AI, Unlike Amazon Web Services, Microsoft Azure, and Google Cloud, all of which to have healthy businesses selling other kinds of compute with AI, as I'll get into later, failing to provide much of a return on investment at all. It's not just the fact that these companies are more specialized than say AWS or Azure, as

you've gathered from the name. These are new, young, and in almost all cases incredibly precarious businesses, each with financial circumstances that will make a Greek finance minister blush. That's because setting up a neocloud is expensive, even if the company in question already has data centers as core We've did with its cryptocurrency mining operation, AI requires, as I said, completely new data center infrastructure to run and call the GPUs,

and those GPUs also need paying for. And then there's the other stuff I mentioned earlier like power, water and the other bits of the computer. CPU are the bah blah blah blah blah blah. As a result, these neoclouds are forced to raise billions of dollars in debt, which they can lateralized using the GPUs they already have, along with contracts from customers, which they then use to buy more GPUs. That's right, they buy GPUs from Nvidia. They raised dead on those GPUs and then they used that

debt to my more GPUs from Nvidia. It's enough to drive a man insane. Corwave, for example, has twenty five billion dollars in debt on an estimated five point three five billion dollars of revenue in twenty twenty five, losing hundreds of billions of dollars per quarter. Now you know who also invests in these neoclouds. You'll never guess. It's

in video. Invidia is also one of Corweed's largest customers, accounting for fifteen percent of its revenue in twenty twenty four, and just signed a deal to buy six point three billion dollars of any capacity that core Weave can't otherwise sell to someone else through twenty thirty two, an extension of a one point three billion dollar twenty twenty three deal reported by the Information. It was also the anchor investment in Corewey's IPO about two hundred and fifty million dollars.

In Vidia is currently doing the same thing with Lambda, another neocloud that Invidia invested in, which also plans to go public next year. In Vidia is also one of Landa's largest customers, signing a deal with it this summer to rent ten thousand GPUs for one point three billion

dollars over four years in the UK. And Video has also just invested seven hundred million dollars in n Scale, a former cryptomner that has never built an AI data center that that has despite having no experience, committed one billion dollars and or one hundred thousand GPUs to an open AI data center in Norway. On Thursday, September twenty fifth, Nscale announced that it closed another funding round within Video

listed as the main backer. Although it's unclear how much money it put in, it would be a safe to assume it's probably at least one hundred million dollars. In Vidia also invested in Nebius, an outgrowth of Russian conglomerate Yandex, and Nebbeus provides, through their partnership with Nvidia, tens of thousands of dollars of compute credits to companies in Videa's Inception startup program. Look, in Vidia's plan is simple, fund

these neoclouds. Let these neoclouds load themselves up with debt, with which point they buy bunches of GPUs from Video, which can be used as collateral for loans along with contracts and customers, allowing the neoclouds to buy even more GPUs from Nvidia. It is just that simple. It's infinite money, right, just money, me money. Now, you fund the company, the company buys from you, you fund them again. They've used the thing they bought to buy from more from you,

unlimited money. Except that is for ones more problem. These companies don't They don't really appear to have that many customers, and they don't appear to be making much money. As I went into a recent premium newsletter, and Video funds and sustains neoclouds as a way of funneling revenue to itself, as well as partners like super Micro and Dell resellers that take in video GPUs like I mentioned and put

them in service to sell pre built to customers. These two companies made up thirty nine percent of in video revenues last quarter. Yet when you remove hyperscalar revenue Microsoft, Amazon, Google, Open Ai, and n Video from the revenues of these neoclouds, there's barely one billion dollars in revenue can bind across

core Weave, Nebius and Lambda. Corwev's five point three five billion dollars in revenue is predominantly made up with its contracts with Nvidia, Microsoft, who are offering that compute to open Ai, Google who have hired Corewave to offer compute to open Ai. And I'm not kidding, and of course open Ai itself, which has now promised core Weave twenty two point four billion dollars in business over the next five years. This is all a lot of stuff, So

I'll make it really simple. There's no real money in offering AI compute, but that isn't Jensen Huong's problem, So we simply will force in Nvidia to hand money to these companies that they have contracts to point at so they can raise debt to buy more of those GPUs, so that in Vidia can give them more contracts they can use that to raise more money. It's it's really bad, all right, It's really bad. When I read this stuff out loud, I feel a little crazy because it's so

obviously unsustainable. Neil clouds are effectively giant private equity vehicles that exist to raise money to buy GPUs from Nvidia or for hyperscalers to move money around so they don't have to increase their capital expenditures and can, as Microsoft did earlier in the year, simply walk away from deals they don't like, with the masses of data center at

least as they walk from. Nebius recently signed a seventeen point four billion dollar deal with Microsoft, which even included the clause in its six K filing and official filing with the government the Microsoft can terminate the deal in the event that the capacity isn't built by the delivery dates.

And by the way, ah Nebius already used the contract that Microsoft gave them to raise three billion dollars to I'm not shitting you here, build the data center to actually to actually provide the compute for the for the contract. They don't have it. Yeah, now they don't have them. They don't have the fucking compute. They they haven't fucking built up. No one built up. They haven't got the compute. Mate, These fucking companies are right anyway. Anyway, sorry, sorry, I'll

stop spiraling. Let me just break down these numbers. Let's look at cort we First, Microsoft, they're sixty percent of their revenue twenty twenty four and they're providing compute mostly for open Ai. Fifteen percent of their revenue last year was in VideA, and then the rest was Meta and

then open Ai and then Google. Lambda half of their revenue comes from Amazon and Microsoft, and now one point five billion dollars of their revenue comes from in VideA, which are their current revenue by the way, and that one point five billion dollars over four years. So the current revenue, it's two hundred and fifty billion dollars. Well,

that would make in video the largest customer. I realize I'm just saying numbers here, but for real, with that contract, because Lambda only made two hundred and fifty million dollars in the first half of this year, and in Video is spreading one point five billion dollars across four years. In Video is the largest customer now Now Nebutus has got similar revenue to Lambda, but their largest customer is now is It's it's fucking Microsoft. It's just they don't

have real customers. They just have hyperscalers or in VideA themselves. And from my analysis, it appears that Core we've despited expectations to make that five point three five billion dollars this year, has only around five hundred million dollars of non Magnificent seven or open AI revenue in twenty twenty five, with Lambda estimated to have maybe a round of one

hundred million dollars in AI revenue. Otherwise, Nebius only around two hundred and fifty million dollars, and that's being generous. In much simpler terms, the Magnificent seven is the AI bubble, and the AI bubble exists to buy more GPUs because, as I'll talk about, there's no real money or growth coming out of this other than the amount that private

credit is investing. And this really is quite worrying. By the way, I had a quote here for analyst that says is about fifty billion dollars a quarter for the low end for the past three quarters. So why is this bad?

global revenue streams. Instead, they're heavily centralized around the same few names, one of which in Vidia, directly benefits from their existence not as a company doing business, but as an entity that can accrue debt and spend money on GPUs.

These neoclouds are entirely dependent on a continual flow of private credit from firms like Gold and Sachs, Who's becked, Nebbia's Corweave and Lambda, JP Morgan, Lambda Crusoe Building Applene, Texas As Open AI Data Center, and of course Corwave and Blackstone Lambda and corwav, who have in a very real sense created an entirely debt based infrastructure to feed billions of dollars directly to Nvidia, all in the name

of an AI revolution that's yet to arrive. The fact that the rest of the neoclo revenue stream is effectively

either a hyperscaler or open ai is also concerning. Hyperscalers are at this point the majority of data center capital expenditures and have yet to prove an any kind of success from building out this capacity, outside of course, Microsoft's investment in open Ai, which has succeeded in generating revenue while burning billions of dollars of revenue on well, I mean, it's not really any profit, is they have just burning money.

It's also insane when you say this stuff. I've got two more goddamn episodes of this and when I read these scripts, I'm just like, how is nobody else more freaked out? Oh well, hyperscaler revenue is also capricious. But even if it isn't, why are there no other major customers? Why across all of these companies does there not seem to be one major customer who isn't open Ai Well, the answers are quite obvious. Nobody that wants it can afford it, and those that can afford it don't need it.

It's also unclear what exactly hyperscalers are doing with this compute,

because it sure isn't making money. While Microsoft makes ten billion dollars in revenue from renting compute to open Ai via their Microsoft as Your Cloud, it does so at cost, and was charging open Ai one dollars and thirty cents per hour for each a one hundred AIGPU at a loss of two point two dollars an hour per GPU, meaning that it is likely losing money on this compute, especially as semi analysis as the total cost per hour per GPU around one dollars and forty six cents with

the cost of capital and debt associated for a hyperscaler, though it's unclear that whether that's for an H one hundred or an A one hundred GPU. In any case, how do these neoclouds pay for their debt if the hyperscalers give up or in video doesn't send them money, or more likely private credit begins to notice that there's no real revenue growth outside of circular compute deals with

neocloud's largest suppliers, investors and customers. Don't know why I said plural there, because it's just one in video, And the answer is they don't. In fact, I have serious concerns that they can't even build the capacity necessary to fulfill these deals, but nobody seems to worry or think

about them. But really, though it appears to be taking Oracle and Cruso around two point five years per gigawatt of compute capacity, how exactly are any of these neoclouds, or indeed Oracle itself able to expand to capture this revenue. Who knows, but I assume somebody is going to say

open Ai. Here's an insane statistic for you. By the way, open ai will account for in both its revenue projected thirteen billion dollars and in its own compute cost ten dollars, somewhere in the region of forty to fifty percent of

all AI revenues in twenty twenty five. As a reminder, open ai is leaked that it will burn one hundred and fifteen billion dollars in the next four years, and based on my estimates, it actually needs to raise I mean upwards are four hundred billion dollars in the next four years based on its three hundred billion dollar deal with Oracle and some recently announced one hundred billion dollar compute purchases for backup. And that alone is a very

bad sign, very very bad indeed. Especially it's with three years and five hundred billion dollars or more into this hype cycle, with few signs of life outside of well open AI promising people money, and that's not healthy or sane or normal, and it's certainly not stable, and it's going to get bad real fast. Gatget tomorrow. Thank you for listening to Better Offline. The editor and composer of

the Better Offline theme song is Matasowski. You can check out more of his music and audio projects at Matasowski dot com, M A T T O S O W s ki dot com. You can email me at easy at Better offline dot com or visit Better Offline dot com to find more podcast links and of course my newsletter. I also really recommend you go to chat dot where's youreaed dot at to visit the discord, and go to our slash Better Offline to check out our reddit. Thank you so much for listening better.