Resident of the world, traveling the road of life
68152 stories
·
21 followers

The AI Bubble's Impossible Promises

1 Share

Readers: I’ve done a very generous “free” portion of this newsletter, but I do recommend paying for premium to get the in-depth analysis underpinning the intro. That being said, I want as many people as possible to get the general feel for this piece. Things are insane, and it’s time to be realistic about what the future actually looks like.


We’re in a bubble. Everybody says we’re in a bubble. You can’t say we’re not in a bubble anymore without sounding insane, because everybody is now talking about how OpenAI has promised everybody $1 trillionsomething you could have read about two weeks ago on my premium newsletter.

Yet we live in a chaotic, insane world, where we can watch the news and hear hand-wringing over the fact that we’re in a bubble, read article after CEO after article after CEO after analyst after investor saying we’re in a bubble, yet the market continues to rip ever-upward on increasingly more-insane ideas, in part thanks to analysts that continue to ignore the very signs that they’re relied upon to read.

AMD and OpenAI signed a very strange deal where AMD will give OpenAI the chance to buy 160 million shares at a cent a piece, in tranches of indeterminate size, for every gigawatt of data centers OpenAI builds using AMD’s chips, adding that OpenAI has agreed to buy “six gigawatts of GPUs.”

This is a peculiar way to measure GPUs, which are traditionally measured in the price of each GPU, but nevertheless, these chips are going to be a mixture of AMD’s mi450 instinct GPUs — which we don’t know the specs of! — and its current generation mi350 GPUs, making the actual scale of these purchases a little difficult to grasp, though the Wall Street Journal says it would “result in tens of billions of dollars in new revenue” for AMD.

This AMD deal is weird, but one that’s rigged in favour of Lisa Su and AMD. OpenAI doesn’t get a dollar at any point - it has work out how to buy those GPUs and figure out how to build six further gigawatts of data centers on top of the 10GW of data centers it promised to build for NVIDIA and the seven-to-ten gigawatts that are allegedly being built for Stargate, bringing it to a total of somewhere between 23 and 26 gigawatts of data center capacity.

Hell, while we’re on the subject, has anyone thought about how difficult and expensive it is to build a data center? 

Everybody is very casual with how they talk about Sam Altman’s theoretical promises of trillions of dollars of data center infrastructure, and I'm not sure anybody realizes how difficult even the very basics of this plan will be.

Nevertheless, everybody is happily publishing stories about how Stargate Abilene Texas — OpenAI’s massive data center with Oracle — is “open,” by which they mean two buildings, and I’m not even confident both of them are providing compute to OpenAI yet. There are six more of them that need to get built for this thing to start rocking at 1.2GW — even though it’s only 1.1GW according to my sources in Abilene.

But, hey, sorry — one minute — while we’re on that subject, did anybody visiting Abilene in the last week or so ever ask whether they’ll have enough power there? 

Don’t worry, you don’t need to look. I’m sure you were just about to, but I did the hard work for you and read up on it, and it turns out that Stargate Abilene only has 200MW of power — a 200MW substation that, according to my sources, has only been built within the last couple of months, with 350MWs of gas turbine generators that connect to a natural gas power plant that might get built by the end of the year.

Said turbine is extremely expensive, featuring volatile pricing (for context, natural gas price volatility fell in Q2 2025…to 69% annualized) and even more volatile environmental consequences, and is, while permitted for it (this will download the PDF of the permit), impractical and expensive to use long-term. 

Analyst James van Geelen, founder of Citrini Research recently said on Bloomberg’s Odd Lots podcast that these are “not the really good natural gas turbines” because the really good ones would take seven years to deliver due to a natural gas turbine shortage.

But they’re going to have to do. According to sources in Abilene, developer Lancium has only recently broken ground on the 1GW substation and five transformers OpenAI’s going to need to build out there, and based on my conversations with numerous analysts and researchers, it does not appear that Stargate Abilene will have sufficient power before the year 2027. 

Then there’s the question of whether 1GW of power actually gets you 1GW of compute. This is something you never see addressed in the coverage of OpenAI’s various construction commitments, but it’s an important point to make. Analyst Daniel Bizo, Research Director at the Uptime Institute, explained that 1 gigawatt of power is only sufficient to power (roughly) 700 megawatts of data center capacity. We’ll get into the finer details of that later in this newsletter, but if we assume that ratio is accurate, we’re left with a troubling problem.

That figure represents a 1.43 PUE — Power Usage Effectiveness — and if we apply that to Stargate Abilene, we see that it needs at least 1.7GW of power, and currently only has 200MW.

As an aside, I need to clear something up, because everybody — including myself! — has been getting this wrong.

When you read “1.2GW data center,” they are almost certainly referring to the data center’s IT load — which is the power consumed by all of the computing equipment inside, but not the cooling systems or power lost in the infrastructure bringing the electricity to the gear itself. The amount of non-IT load power required, furthermore, can fluctuate. 

Data centers need far more power than their IT load, and any time you read a “gigawatt” data center, know that they need about 30% more power than the amount of capacity the data center has.

Stargate Abilene does not have sufficient power to run at even half of its supposed IT load of 1.2GW, and at its present capacity — assuming that the gas turbines function at full power — can only hope to run 370MW to 460MW of IT load.

I’ve seen article after article about the gas turbines and their use of fracked gas — a disgusting and wasteful act typical of OpenAI — but nobody appears to have asked “how much power does a 1.2GW data center require?” and then chased it with “how much power does Stargate Abilene have?”

The answer is not enough, and the significance of said “not enough” is remarkable.

Today, I’m going to tell you, at length, how impossible the future of generative AI is. 

What Makes a Gigawatt

Gigawatt data centers are a ridiculous pipe dream, one that runs face-first into the walls of reality.  

The world’s governments and media have been far too cavalier with the term “gigawatt,” casually breezing by the fact that Altman’s plans require 17 or more nuclear reactors’ worth of power, as if building power is quick and easy and cheap and just happens.

I believe that many of you think that this is an issue of permitting — of simply throwing enough money at the problem — when we are in the midst of a shortage in the electrical grade steel and transformers required to expand America’s (and the world’s) power grid.

I realize it’s easy to get blinded by the constant drumbeat of “gargoyle-like tycoon cabal builds 1GW  data center” and feel that they will simply overwhelm the problem with money, but no, I’m afraid that isn’t the case at all, and all of this is so silly, so ridiculous, so cartoonishly bad that it threatens even the seemingly-infinite wealth of Elon Musk, with xAI burning over a billion dollars a month and planning to spend tens of billions of dollars building the Colossus 2 data center, dragging two billion dollars from SpaceX in his desperate quest to burn as much money as possible for no reason. 

This is the age of hubris — a time in which we are going to watch stupid, powerful and rich men fuck up their legacies by finding a technology so vulgar in its costs and mythical outcomes that it drives the avaricious insane and makes fools of them. 

Or perhaps this is what happens when somebody believes they’ve found the ultimate con — the ability to become both the customer and the business, which is exactly what NVIDIA is doing to fund the chips behind Colossus 2.

According to Bloomberg, NVIDIA is creating a company — a “special purpose vehicle” — that it will invest $2 billion in, along with several other backers. Once that’s done, the special purpose vehicle will then use that equity to raise debt from banks, buy GPUs from NVIDIA, and then rent those GPUs to Elon Musk for five years.

Hell, why make it so complex? NVIDIA invested money in a company specifically built to buy chips from it, which then promptly handed the money back to NVIDIA along with a bunch of other money, and then whatever happened next is somebody else’s problem.

Right?

Actually, wait — how long do GPUs last, exactly? Four years for training? Three years? The A100 GPU started shipping in May 2020, and the H100 (and the Hopper GPU generation) entered full production in September 2022, meaning that we’re hurtling at speed toward the time in which we’re going to start seeing a remarkable amount of chips start wearing down, which should be a concern for companies like Microsoft, which bought 150,000 Hopper GPUs in 2023 and 485,000 of them in 2024.

Alright, let me just be blunt: the entire economy of debt around GPUs is insane.

Assuming these things don’t die within five years (their warranties generally end in three), their value absolutely will, as NVIDIA has committed to releasing a new AI chip every single year, likely with significant increases to power and power efficiency. At the end of the five year period, the Special Purpose Vehicle will be the proud owner of five-year-old chips that nobody is going to want to rent at the price that Elon Musk has been paying for the last five years. Don’t believe me? Take a look at the rental prices for H100 GPUs that went from $8-an-hour in 2023 to $2-an-hour in 2024, or the Silicon Data Indexes (aggregated realtime indexes of hourly prices) that show H100 rentals at around $2.14-an-hour and A100 rentals at a dollar-an-hour, with Vast.AI offering them at as little as $0.67 an hour.

This is, by the way, a problem that faces literally every data center being built in the world, and I feel insane talking about it. It feels like nobody is talking about how impossible and ridiculous all of this is. It’s one thing that OpenAI has promised one trillion dollars to people — it’s another that large swaths of that will be spent on hardware that will, by the end of these agreements, be half-obsolete and generating less revenue than ever.

Think about it. Let’s assume we live in a fantasy land where OpenAI is somehow able to pay Oracle $300 billion over 5 years — which, although the costs will almost certainly grow over time, and some of the payments are front-loaded, averages out to $5bn each month, which is a truly insane number that’s in excess of what Netflix makes in revenue. 

Said money is paying for access to Blackwell GPUs, which will, by then, be at least two generations behind, with NVIDIA’s Vera Rubin GPUs due next year. What happens to that GPU infrastructure? Why would OpenAI continue to pay the same rental rate for five-year-old Blackwell GPUs?  

All of these ludicrous investments are going into building data centers full of what will, at that point, be old tech. 

Let me put it in simple terms: imagine you, for some reason, rented an M1 Mac when it was released in 2020, and your rental was done in 2025, when we’re onto the M4 series. Would you expect somebody to rent it at the same price? Or would they say “hey, wait a minute, for that price I could rent one of the newer generation ones.” And you’d be bloody right! 

Now, I realize that $70,000 data center GPUs are a little different to laptops, but that only makes their decline in value more profound, especially considering the billions of dollars of infrastructure built around them. 

And that’s the problem. Private equity firms are sinking $50 billion or more a quarter into theoretical data center projects full of what will be years-old GPU technology, despite the fact that there’s no real demand for generative AI compute, and that’s before you get to the grimmest fact of all: that even if you can build these data centers, it will take years and billions of dollars to deliver the power, if it’s even possible to do so.

Harvard economist Jason Furman estimates that data centers and software accounted for 92% of GDP growth in the first half of this year, in line with my conversation with economist Paul Kedrosky from a few months ago

All of this money is being sunk into infrastructure for an “AI revolution” that doesn’t exist, as every single AI company is unprofitable, with pathetic revenues ($61 billion or so if you include CoreWeave and Lambda, both of which are being handed money by NVIDIA), impossible-to-control costs that have only ever increased, and no ability to replace labor at scale (and especially not software engineers).  

OpenAI needs more than a trillion dollars to pay its massive cloud compute bills and build 27 gigawatts of data centers, and to get there, it needs to start making incredible amounts of money, a job that’s been mostly handed to Fidji Simo, OpenAI’s new CEO of Applications, who is solely responsible for turning a company that loses billions of dollars into one that makes $200 billion in 2030 with $38 billion in profit. She’s been set up to fail, and I’m going to explain why.

In fact, today I’m going to explain to you how impossible all of this is — not just expensive, not just silly, but actively impossible within any of the timelines set

Stargate will not have the power it needs before the middle of 2026 — the beginning of Oracle’s fiscal year 2027, when OpenAI has to pay it $30 billion for compute — or, according to The Information, choose to walk away if the capacity isn’t complete. And based on my research, analysis and discussions with power and data center analysts, gigawatt data centers are, by and large, a pipedream, with their associated power infrastructure taking two to four years, and that’s if everything goes smoothly.

OpenAI cannot build a gigawatt of data centers for AMD by the “second half of 2026.”  It haven’t even announced the financing, let alone where the data center might be, and until it does that it’s impossible to plan the power, which in and of itself takes months before you even start building. 

Every promise you’re reading in the news is impossible. Nobody has even built a gigawatt data center, and more than likely nobody ever will. Stargate Abilene isn’t going to be ready in 2026, won’t have sufficient power until at best 2027, and based on the conversations I’ve had it’s very unlikely it will build that gigawatt substation before the year 2028. 

In fact, let me put it a little simpler: all of those data center deals you’ve seen announced are basically bullshit. Even if they get the permits and the money, there are massive physical challenges that cannot be resolved by simply throwing money at them. 

Today I’m going to tell you a story of chaos, hubris and fantastical thinking. I want you to come away from this with a full picture of how ridiculous the promises are, and that’s before you get to the cold hard reality that AI fucking sucks. 

Read the whole story
mkalus
33 minutes ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete

Ping

1 Comment and 3 Shares
Progress on getting shipwrecked sailors to adopt ICMPv6 has been slow.
Read the whole story
alt_text_bot
9 days ago
reply
Progress on getting shipwrecked sailors to adopt ICMPv6 has been slow.
mkalus
36 minutes ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete

Measure Twice, Cut Once

2 Comments and 6 Shares
"Measure zero times, cut zero times." --carpenter who has achieved enlightenment and realized the wood is fine where it is
Read the whole story
alt_text_bot
11 days ago
reply
"Measure zero times, cut zero times." --carpenter who has achieved enlightenment and realized the wood is fine where it is
mkalus
36 minutes ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete
1 public comment
deebee
9 days ago
reply
Buffalo Bill would be the exception here
America City, America

Hiking

3 Comments and 8 Shares
Every downhill walk is a waterslide that might have been.
Read the whole story
alt_text_bot
16 days ago
reply
Every downhill walk is a waterslide that might have been.
steelhorse
13 days ago
The water is going to be an issue. Instead of a water slide, why not a really good Teflon coated slide? Or use that stuff Clark Griswold invented.
mkalus
36 minutes ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete
1 public comment
summerofjay
16 days ago
reply
Or conserve water by having on-demand zip-lines!
23223
marcrichter
15 days ago
I'd wager it's a closed loop
cheatmaster
15 days ago
The drones can just lift the end and the water will flow back - the slide doesn't even need power.

Google won’t fix ‘ASCII smuggling’ hack in Gemini AI

1 Share

Chatbots have various guard rails. But they’re trained on all the good stuff and the bad stuff, so slapping a filter on the front just isn’t going to work well. You can probably sneak around it to give commands to the bot.

One way to get around the filters is ASCII smuggling — where you hide your command in weird high-end Unicode characters.

Viktor Markopoulos at FireTail tested a pile of chatbots on how well they block ASCII smuggling. ChatGPT, Copilot, and Claude all catch this attack and block it successfully. Gemini, Grok, and DeepSeek do not. [FireTail]

Gemini is special — Google’s put it everywhere. For instance, in any company that uses Google Apps. Markopoulos built test attacks against Google Calendar invites. He could hit the titles, the meeting organizer’s name, and the meeting descriptions.

Markopoulos reported this to Google last month. Google said it wasn’t a security bug:

The issue you’re describing can only result in social engineering.

Because fooling people into doing things has never been a security issue, right? Yeah, turns out it has. Social engineering is most of how you hack companies. But Google has determined: just go for it. Stopping prompt injections is hard, so we’ve made a carve-out just for this one.

It’s grossly negligent for Google to shove its AI into everything, then declare an attack method isn’t really a security hole because it only involves dealing with people. After all, who deals with people at work?

Read the whole story
mkalus
1 hour ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete

OpenAI does deals to make a deal — and the money is getting worried

1 Share

The AI bubble is the bubbliest bubble that ever bubbled. We’ve covered at length how venture capital is moribund except AI and especially OpenAI, how the VCs are desperate to keep the party pumping, how the investors will spend billions of real actual dollars to get twice the imaginary dollars, and how this whole huge Enron-style imaginary money shuffle has a couple of years at most before it all falls over.

So this week we have OpenAI and AMD announcing a deal! Or a deal to do a deal. Nothing has actually happened. [WSJ]

AMD is a huge chip maker. They do the best CPUs right now. But AMD is also Nvidia’s biggest competitor for video card chips that just happen to be good for number crunching too. Such as training AI models.

I’m actually surprised that AMD haven’t done better in the AI bubble and it’s all been Nvidia. But OpenAI is here to help!

Starting in 2026, OpenAI will, allegedly, purchase up to “6 gigawatts” of AMD chips. AMD says that’s “tens of billions of dollars per gigawatt.” They will not say what the actual dollar numbers are.

OpenAI gets cheap options on up to 10% of AMD’s company value if it deploys enough AMD chips. AMD’s stock price also has to go up.

Not one penny has moved, not one chip has moved, not one share of stock has moved. OpenAI has to find the money to buy all these AMD chips.

But in the meantime, both companies can do press releases and throw around eight-digit numbers with a dollar sign in front.

It doesn’t matter that the eight-digit numbers are just made-up. Because AMD’s stock price promptly went up like a rocket, up 23.7% on the day, and up 38% in the past week as I write this.

Nvidia’s not sleeping. They’re doing a lot of circular deals themselves, where they “invest” in an AI company so that company can buy more Nvidia chips.

Last month, Nvidia also did a deal-to-make-a-deal with OpenAI. Nvidia will put “up to” $100 billion into OpenAI. And the way that works is, OpenAI buys up to 10 gigawatts of Nvidia chips, and Nvidia buys more and more equity in OpenAI. With actual dollars! As far as anything actually happens in this deal. [FT]

Both the AMD and Nvidia deals depend on OpenAI having the data centres to put all these chips in, and the electricity to power them. OpenAI is already having problems getting enough power for these fairyland data centres in the sky.

The mainstream press is finally saying the words “AI bubble” out loud. But somehow, they’re still not really bothering to connect the dots. The numbers don’t maIke sense. They’re just stupid:

  1. OpenAI sets money on fire to sell a product that loses money every time someone uses it.
  2. There is no path to profit.
  3. To keep this going, OpenAI is supposedly building more data centres — with other people’s money they haven’t got yet.
  4. These vaporware data centres will run on power that doesn’t exist yet.
  5. To power these data centres, OpenAI needs to build out generation capacity faster than anyone’s ever built out that much power.
  6. If they can even get the parts.

SoftBank is still trying to get together the money to prop up OpenAI. They’re looking under the sofa cushions for $5 billion and trying to borrow the money backed by their share holding in ARM. [Bloomberg]

The circular deals are like an industry-wide Enron. They’re all sending each other the same billion virtual dollars in a loop. [SF Gate]

I’m going to predict things again. Here’s how I think the bubble collapses:

  • First, the smaller AI companies will go down. They’ll just run out of cash and won’t be rescued. We might see some desperate mergers so their investors can pretend all that imaginary book value is still there.
  • The last one to go down will be OpenAI. It’ll run out of money so thoroughly that not even everyone put together whose accounts are held up by OpenAI equity will be able to pump in money fast enough to keep the company propped up. They finally have to admit their losses.
  • Without the flagship, the whole thing falls down.
  • Google, Microsoft, and Amazon can finally admit this never made any money. They have an excuse to account massive losses and go “oh well.”
  • Nvidia will be fine — they’re the ones who got everyone else’s money.

Now we’ve got the Bank of England weighing in and saying this is obviously dumb and broken and it’s going to crash the whole market: [Bank of England]

Downside factors included disappointing AI capability/adoption progress or increased competition, which could drive a re-evaluation of currently high expected future earnings. Material bottlenecks to AI progress … could also harm valuations, including for companies whose revenue expectations are derived from high levels of anticipated AI infrastructure investment.

That sounds understated — but if you translate from central banker, all that means “DANGER! DANGER!”

Jamie Dimon of JPMorgan, talking to the BBC, gives it six months to two years before there’s a major correction, where the whole stock market crashes. Dimon was asked about how this happens: [BBC; BBC, podcast]

BBC: If you look at some of the things that are going on in the AI tech space, you’ve got, you know, NVIDIA taking stakes in OpenAI, Microsoft taking stakes in OpenAI, the federal government taking stakes in Intel, there’s all these cross-holdings. It looks a bit weird, doesn’t it? How would you track it?

DIMON: Well, AI is a little bit different, you know, and there are some justifications for vendor finance and things like that, so you’re going to have to go through each one. Regarding America’s investments, I think, you know, I’m not going to talk about stockpiling, you know, things like crypto, I say we should be stockpiling bullets, guns, and bombs, I mean, the world’s a much more dangerous place, and I’d rather have safety than not.

That’s the sort of very normal talk you don’t really expect to hear from an investment banker. The US already has the world’s largest military — Dimon’s speaking from a personal place. After a stockmarket crash, you gotta stockpile bullets, guns, and bombs! But I’m sure Dimon won’t shoot his own foot off personally when he’s going rich guy Mad Max wanna-survivalist.

Read the whole story
mkalus
1 day ago
reply
iPhone: 49.287476,-123.142136
Share this story
Delete
Next Page of Stories