In recent weeks, there's been a lot of market anxiety about the sustainability of the AI boom. This was partly driven by the outrage around Sarah Fryer, Open AI finance chief, floating the idea that a government backstop for its $1.4 trillion data center build out might be a good idea.
Fryer quickly walked back her suggestion in the LinkedIn post later that day, saying that she had meant that the government needed to play their part in combination with the private sector to contribute to America's AI growth and that Open AI was not seeking a government backstop for their
infrastructure commitments. Her statement, while attempting to calm the outrage, only confused matters even further about how the not yet profitable start up plans to pay for its massive AI data centre and chip commitments. Sam Altman tweeted on the Everything app. We do not have or want government guarantees for open AI data centres. We believe that governments should not pick winners or losers and the taxpayers should not bail out companies that make bad business decisions or
otherwise lose in the market. Then it turned into a Bill Aikman tweet at that point where he went on and on for around 20 pages. At first I was thinking, who would write a tweet that long, and then I realized he he'd probably just used ChatGPT. He knew that people would only read the first few lines, but he wanted to seem thoughtful. So how to churn out an entire novel?
The core problem for Open AI is that they've signed more than $1.4 trillion in infrastructure commitments over the last few months with the goal of building out the data centers that it says are needed to meet soaring demand. But they are nowhere near having the money required to complete those deals. FIRE gave the example of having to hold back Sora 2 for months due to compute constraints. I just want to be clear what it means when I say we're compute
constrained. It means that for example, we cannot roll out our new models when they are ready. So when Sora 2 was ready to when Sora 2 actually launched, there was probably a good 6-7 months actually gap there. OK, the agreements that they've signed have raised lots of questions around how a cash burning company with tiny revenues relative to their plant spending can possibly make such huge commitment. This was not the first time Open AI has looked to Washington for help, either.
Just a month ago, the company sent a detailed letter to the White House urging the federal government to double down ON Semiconductor subsidies, asking the tax credits be expanded to cover the entire AI supply chain, from chip fabrication to data centres and grid hardware. The company argued that broadening eligibility for taxpayer funded subsidies would lower the effective cost of capital de risk early investment
and unlock private capital. Open AI and it's data centre partners are of course amongst the largest buyers of semiconductors in the world, so any subsidy would directly
benefit them. AI is being pitched to governments around the world as being a matter of grave national security and economic importance, akin to past industrial mobilizations like the Manhattan Project and the Space Race. If AI companies can put it on that level and pitch it as being too important to fail, a government funded backstop might make sense.
The irony though is that while lobbying for taxpayer support in the name of geopolitical survival, the same businesses are pumping billions into building models that generate weird anime girlfriends, SpongeBob deepfakes, Sam Altman's Studio Ghibli style profile photo and in Elon Musk's case, a chatbot that appears to have been hard coded this week to constantly flatter him in the cringiest manner possible.
Which caused all sorts of hilarity on the everything app formerly known as Twitter this week. We'll come back to that in just a minute though. While all of this was going on, NVIDIA warned for the first time in a regulatory filing that it's customers ability to secure capital and energy for AI data centres could potentially slow
it's growth. On top of that, Amazon launched A complaint with the Public Utility Commission of Oregon that the electric utility was failing to provide sufficient power for the four new data centres that had built, highlighting the strain that rapid data centre expansion is putting on electric grids. I guess the utility agreed to hook them up to the grid, but not necessarily to provide them
with the power that they wanted. The question hanging over Silicon Valley is not so much whether AI will change the world, but whether the world can afford to build it. This brings us to Nvidia's earnings report on Wednesday night. The tech rally that has defined much of 2025, especially after deliberation day sell off in April, began to lose momentum in early autumn.
Some analysts traced the inflection point to when Open AI announced a $300 billion cloud deal with Oracle and NVIDIA pledged up to 100 billion in reciprocal investments. Those headlines, which were meant to signal confidence, instead raised questions about circular financing and the sheer scale of spending commitments. Private credit blow UPS added to the unease in markets, reviving concerns about lending standards and fraud in a market already stretched by aggressive
leverage. Valuations were lofty, and the spaghetti diagrams of interlocking deals with hyperscalers funding AI labs that fund chip makers that fund hyperscalers started looking increasingly fragile. No surprise, then, that bubble talk intensified. Nvidia's earnings report on Wednesday temporarily eased those fears. The world's most valuable company and the beating heart of the AI trade posted a 62% jump in revenue for the three months to October, far ahead of
expectations. Data center sales hit $51.2 billion, and the company raised its revenue forecast for the current quarter to $65 billion. For now, the numbers seem to just by the hype. As Robert Armstrong put it on the Unhedged podcast, the worry is not invidious price to earnings ratio. The worry is that the revenue it's earning, and the growth rate of that revenue is ultimately unsustainable. At today's pace, Nvidia's
valuation makes sense. The question is whether the growth curve can defy gravity indefinitely. Open AI's finances look even more precarious than most people realize. Microsoft's September earning filing revealed that Open AI lost roughly 11 1/2 billion dollars in a single quarter, it's worst on record. That pushes year to date losses north of $25 billion against projected annual revenue of about $20 billion.
The company has raised nearly $58 billion in equity so far and was valued at $500 billion last month. The company is talking about an IPO at a $1 trillion valuation next year, which would float the shares on an exchange and possibly bring in about $60 billion in cash. But that's just over 4% of it's $1.4 trillion infrastructure commitment. To bridge the gap, Open AI has leaned on creative deal structures.
NVIDIA has pledged up to $100 billion in reciprocal investments, while AMD granted Open AI warrants to buy 10% of its stock for a penny per share if deployment milestones are mad. Sarah Fryer, Open AICFO, explained the company's financing at the Wall Street Journal event. While she's from Northern Ireland, she must have been in Silicon Valley long enough to know that the first step in raising capital is to use the magic word. The innovation on the finance
side to pay for it is massive. She then went on to say. We've raised equity as a private company, very kind of typical path, but we've raised a lot. We're building a really healthy business. So free cash flow, CFO's favorite way to fund anything that is absolutely climbing quickly. But I think the third area we've gotten into is really working with our ecosystem to do some really interesting financing deals.
I'm particularly proud of the AMD warrant structure that we put in place just a few weeks back because it's very strong alignment of incentives. This was a really bizarre claim as Open AI can't fund anything with free cash flow. When that cash flow is negative, she then digs into explaining the AMD warrant. What we've seen is when someone comes out and says we're going to work with Open AI, they admit immediately are often seeing kind of impact on their stock
price. And so to the extent that that's going to happen, we would like to have some alignment on that. And I think Lisa and team did something incredibly creative
with that warrant structure. The warrant deal between Open AI and AMD is a strategic partnership where Open AI commits to buying billions of dollars worth of AM Dai chips and in return, AMD grants Open AI warrants to purchase up to 160 million of its shares, which is about a 10% stake in the company at a nominal price of 1 cent per share.
When the deal was announced, AMD stock went up at 24%, but the deal only vests if Open AI buys 6 gigawatts of AMD chips, hits undisclosed milestones, and AM DS share price triples. The AMD deal would bring in almost $100 billion worth of AMD stock if all of the targets were hit, including the tripling of AMD stock price. But it's tied to six gigawatts of chip purchases, which she later explains. So one GW data center build today is about a $50 billion investment. That's for one gig.
How that really breaks down is about 15 billion is for the land PowerShell and about 35 billion is for the chips. So to bring in $100 billion, they have to spend $300 billion. Nvidia's $100 billion pledge to invest in Open AI is also tied to reciprocal commitments. If all of these deals worked out, Open AI could bring in $200 billion. But that still leaves them $1.2 trillion short. And they're burning 10s of billions of dollars per year
with no end. Insight The unit economics of running the current generation of LLM is dire. As Paul Kedrosky explained it on the Odd Lots podcast. The incentive seems to be for all players to just grow the top line as much as possible, even if adding more users just leads to greater and greater losses. The models have negative unit economics, which is a fancy way of saying we lose money on every sale and try to make it up on volume.
In AI, costs rise almost linearly with usage, which is very different to traditional software. There's no marginal cost magic going on. According to Forbes, despite an invitation only roll out, Open AI may be losing around $15 million a day or $5 billion annualized on Sora 2. It's AI video generating app. Tech firms have always been creative about financing, but Open AI's approach borders on the surreal, where it's become all about trying to find infinite money glitches.
MicroStrategy, or Strategy as it's now called, is trying a similar trick with its Bitcoin investments, which I don't expect to end very well. Behind the headlines is a financing structure that looks increasingly baroque. Hyperscalers and AI labs are using special purpose vehicles so that they can borrow but keep the debt off their balance sheets. Tech firms have essentially been reinventing structured finance to build AI models so that they
can generate AI girlfriends. That is just the world that we live in and I will always love you, Sarah Fryer explained at the Wall Street Journal Offend that each GW of compute costs around $50 billion, where 15 billion is the land and infrastructure and $35 billion is the GPU's. From a financing perspective, people know how to finance data centres. They typically all have 2025, even 30 year lives. Those are easy things.
I would say today to finance chips have not been as easy to finance because number one, I think we're all still getting our arms around what is the life of a frontier chip. What this means is that the more innovation that happens with chips, the faster they can be expected to depreciate. And so the $35 billion worth of chips in a $50 billion data centre are very difficult to finance.
People don't want to own them if they might collapse in value when a new one comes out, and people really don't want to accept them as collateral on alone. That's when she put forth this idea. And so this is where we're looking for an ecosystem of banks, private equity, maybe even governmental, the ways
governments can come to bear. Meaning like a federal subsidy or. Something meaning like just first of all, the the backstop, the guarantee that allows the financing to happen, that can really drop the the cost of the financing, but also increase the the loan to value. So the amount of debt that you can take on top of an equity portion? For some federal backstop for
CHIP investment. Exactly. And I think we're seeing that. I think the US government in particular has been incredibly forward lean has really understood that AI has is almost a national strategic asset and that we really need to be thoughtful when we think about competitive competition with, for example, China. Are we doing all the right things to grow our AI ecosystem
as fast as possible? Essentially, the problem is that they want to lever up their bet on AI, but banks wouldn't want to lend, and the interest rate on a loan backed by rapidly depreciating chips would be so high that you would need the
government to back the loan. Now I can tell that this will make some of my viewers angry, but there's actually no need to get angry about something like this, as both Sam Altman and Elon Musk have both explained in the past that AGI will soon make money obsolete, so who cares now? Even if the money materializes and then suddenly doesn't matter anymore, the electrons may not open AI. Stargate project alone would require 10 gigawatts of power, which is roughly 10 nuclear power plants.
It's full build out implies 23. And that's just open AI. Google has a model. Facebook, or whatever they call themselves has one too. There's Grog, good old Grog, and Tropic, and lots, lots more. What I'm saying is we're going to need a lot of power plants. We're going to need a bigger boat, and we also have to plug in our cars and robots. Only one new nuclear power station has been built in the United States in the last 30 years.
It took a decade to complete and was the most expensive power plant ever built. Bloomberg estimates that AI driven electricity demand will more than double over the next 10 years. Utilities are already balking. Amazon has filed A complaint against Pacificorp for failing to deliver promised power to 4 Oregon data centres. Pacificorp says that it's protecting other customers from
indirect harms or translation. We can't turn the lights off in Portland so that Jeff Bezos can train a chatbot behind the meter. Gas turbines are proliferating, too, as stop gaps. Some operators are whispering about nuclear partnerships. These fixes create stranded asset risk, as a natural gas plant lasts 30 years and a GPU cluster might be obsolete in 18 months. Lenders see the mismatch, and they flinch.
Tech firms that promise to dematerialize the economy now need more concrete, copper, and electricity than the steel mills did. The cloud, which was supposed to be weightless, turns out to be very heavy. So why keep spending? Well, because the game is framed as being existential. US labs talk about sovereign AI and competition with China. Once you call something existential, the limit on spending becomes unlimited. Polkidoski described it as a meta bubble on the Odd Lots
podcast. Tech hype, Real estate, state speculation, loose credit and a potential government backstop all in one. There are some bubbly signs. I remember in 1999 seeing adverts on CNBC for a company that manufactured equipment used in the way for fabrication steps of making semiconductors. I couldn't understand at the time why they were paying for TV adverts when their customers would all know who they were and what they sell. No one watches CNBC and decides to start manufacturing computer
chips in their garage. I later worked out that they were advertising the stock, not the products. The stock fell around 80% over the next three years. Recently, I've seen a tech CEO being interviewed wearing AT shirt with his company's ticker symbol on it. Not the company name, the ticker symbol. I've noticed that every podcast I listened to over the last few weeks seems to have adverts for an AI military tech company.
And once again, I wonder if they think that their potential customers might be listening to a Bloomberg podcast or if they just want to pump the stock. I'll note that the CEO of that company constantly. He talks about burning short sellers while dumping his own stock. Even if there is a bubble, it can be impossible to know when it'll pop. As I mentioned a few weeks ago, the big tech firms funding a lot of the AI spending are so profitable in their core businesses that they can afford
this gamble. So should investors cash out of the stock market then? Well, probably not, unless they know how they'll get back in again. If you're a diversified investor with a long holding period, even if you invested the day before the 1987 crash, right before the credit crunch, or right before the COVID sell off, you might have been uncomfortable for a while. But if you stayed invested, you weren't good returns over time. The Economist estimates that should an AI crash occur.
Her it could erase 8% of US household wealth and cut consumption by $500 billion, or 1.6% of GDP. They show that at the peakofthe.com bubble, the market cap of the S&P was 124% of USGDP. When the bubble burst, tech stocks lost on average 76% of their value. Since CHAT GP TS launch in 2022, American stocks are up 71% and the S&P is worth 175% of GDP.
They point out that a crash today would have a bigger effect on ordinary Americans than it did 25 years ago, as the share of household wealth in the stock market has climbed from 17% back then to 21% today. If the stock market fell as much as it did back then, it would wipe out as much as 8% of US household wealth. Foreign investors who are heavily invested in US tech would take a significant hit,
too. The fallout wouldn't stop at Silicon Valley. Pension funds, REITs, and private credit vehicles are exposed to AI investment, too. Utilities that built gas plants for data centers could be left with standard assets. The last time America overbuilt infrastructure this aggressively was the telecom boom, and much of the dark fibre that was laid back then was never lit.
As I said a few weeks ago, the tech boom today is very different to the.com bubble of the late 90s where unprofitable start-ups were racing to IPO after a few months in business, burning cash on vague promises of eyeballs and banner ads. Today's big tech firms, Microsoft, Amazon, Google, Meta are highly profitable, well run businesses with entrenched
revenue streams. They may be pouring 10s of billions into AI, but if these bets fail, their core businesses, cloud, advertising and e-commerce remain intact and cash flow positive. The real risk sits with the private AI labs and their venture backers, not really with the hyperscalers. If anything resembles the fraud of 1999, it's crypto, not trillion dollar companies with fortress balance sheets. For AI users, this frenzy is a
gift. Competition has meant that the models improve rapidly and their prices start low. There's no reason not to use these products while they're free or almost free for AI investors. The economics are unforgiving. Better chips make models faster and make yesterday space chips worthless. Every leap forward accelerates depreciation on the collateral lenders are asked to finance. That's why banks refuse to lend. They prefer assets that last longer than a news cycle.
Sam Altman says that Open AI isn't and wasn't pitching for a government backstop, that he thinks government should build their own AI infrastructure. That might happen, but it does nothing to solve Open AI's problem financing $1.4 trillion of private data centres with non government guaranteed bonds. For now, the company is betting that the capital markets will keep playing along. If they don't, the build out debate FIRE stumbled into will return louder, sharper, and
harder to ignore. I almost forgot to include this piece, but one of the funnier news stories of the week was about Grok, Elon Musk's maximum truth seeking chatbot. It seems that the code must have been tweaked a bit this week and adjusted such that Grok's output is more in line with Musk's way
of thinking. Grok began claiming that Elon Musk is more physically fit than LeBron James, a better role model than Jesus, and that his intellect is in the same bracket as Isaac Newton's, that he was a better fighter than Mike Tyson, and that he's funnier than Jerry Seinfeld. People quickly worked out that Grok would say that Musk was amazing at everything, which led to some inappropriate questions and this headline at Four O 4 Media.
Many of the Grok responses were quietly deleted on Friday, and Musk tweeted that someone had manipulated Grok into saying absurdly positive things about him. I'm sure if he ever catches that guy, he'll be in a world of of trouble. Thanks again for tuning into the podcast, which is entirely supported by viewers like you on Patreon. If you'd like to support the podcast, I'll leave a link in the description. Have a great day and talk to you again soon. Bye.