AI has become the dominant theme in investing today. It's driving the stock market, influencing the bond market, and to a remarkable degree, driving the entire US economy. But here's what concerns me. This entire AI investment boom has become increasingly dependent on the success of just a few companies. And if these companies don't deliver on their enormous promises, the consequences could be incredibly severe. To put it simply, the US could easily enter a recession and we could almost certainly see a large market correction.
Hi, this is Steve Eisman and welcome to another episode of The Real Eisman Playbook. Now, don't get me wrong, AI is an extraordinary technology and I have no doubt it will change much of the world, but the questions investors need to be asking are how quickly, at what cost, and most importantly, who's actually going to make money from it? Because right now, hundreds of billions of dollars are being poured into AI infrastructure. Companies are taking on enormous amounts of debt and investors are betting that all this spending will eventually generate extraordinary profits. So, in this video, I want to examine exactly where we stand in the AI boom. We'll look at who's financing it all, the bull, bear, and doomsday scenarios, and most importantly, the five major risks that I believe could derail the entire investment story. Not since the railroad as a new technology caused so much capex confusion, tension, and excitement.
Since for many, the whole AI story is rather overwhelming, I thought it would be a good idea to go through the major aspects of the story, and importantly, the major players. First things first, what exactly is an LLM? Well, LLM stands for large language model. Now, I'm oversimplifying here, but think of it as an incredibly sophisticated word prediction machine. You ask it a question and it predicts what word should come next to produce an answer. That's how ChatGPT and Claude work. That's different, the LLM, from AGI or artificial general intelligence. AGI is essentially the holy grail where machines could theoretically match or exceed human intelligence across a broad range of tasks. In the wonderful words of Spinal Tap, it goes to 11.
Now, we're not there yet, although the people raising billions of dollars to develop it would certainly like you to believe we're getting close. Then there's agentic AI, which I think is particularly interesting. Whereas a traditional LLM answers your question, an AI agent can actually go out and perform tasks on your behalf. My wife Valerie, for example, has used an AI agent to search for theater tickets, compare prices, and purchase the theater tickets she wanted with her permission. That's genuinely useful technology and I think it has enormous potential to change how we live and work. But here's where the investment story gets interesting. All of this requires extraordinary amounts of computing power. And that computing power lives inside data centers. Data centers are enormous facilities filled with sophisticated chips, servers, cooling equipment, and electrical infrastructure. And by the way, these things are incredibly expensive. A new large AI data center capable of supporting a gigawatt of power demand could cost around, get this, $30 billion to build and equip. That's just one data center.
Now, you've probably heard that data centers are popping up across the country and indeed across the world. But who's actually paying for all of this? Primarily, not exclusively, but primarily it's the hyperscalers. That's companies like Microsoft, Google, Amazon, and Oracle, which have spent years building the cloud computing infrastructure that businesses rely on. Except now AI is forcing them to build data centers on an entirely different scale. In 2026 alone, hyperscalers, just the hyperscalers, are expected to spend roughly $850 billion on AI infrastructure and potentially even more next year. Astonishing sums. And that leads us to probably the two most important players in the current AI story. At the center of this entire AI ecosystem sit two private companies. OpenAI led by Sam Altman and Anthropic led by Dario Amodei. They're developing the AI models that are driving much of this demand for data centers. And possibly the kingpin of this whole shebang is Nvidia. Nvidia creates the most sophisticated chips in the world. And Jensen Huang is at the epicenter of the AI buildout. So you have the AI companies developing the models, the hyperscalers spending hundreds of billions building the infrastructure to support it, Nvidia manufacturing billions of dollars in chips and financing them as well, and all kinds of debt and equity investors including private credit, banks, insurance companies, corporations who buy the debt issuances, and stockholders all providing enormous amounts of capital to finance the whole thing.
Which brings us to literally the trillion dollar question. Who is actually going to make money from all of this? Because as we saw with the railroads and later the internet, a technology can completely transform the world while leaving an enormous trail of bankruptcies behind it. And right now, the market seems to have a few very different ideas about how this AI story will play out. So before we get into the underlying risks, let's look at the four possible scenarios. Starting with the bull case, AI for the win. AI will change the world and Anthropic and OpenAI will continue to drive the stock market to new and greater heights with Nvidia setting endless records. Result, the market keeps going up. Now, this has been the dominant narrative for about 2 years. So, it's not crazy and it needs to be taken seriously because the largest tech companies in the world believe it. I give this narrative a 30% chance.
The bear case, AI is overhyped. LLMs are useful, but adoption will be very slow. And even if adoption is fairly rapid, the capital required to build the data centers and pay Nvidia and others for the chips and pay Anthropic and OpenAI their exorbitant prices for their programming won't happen. The Chinese open-weight models will put pressure on pricing. A price war will ensue. Moats vanish and the returns on investment don't measure up. The result, the bear case leads to a full economic recession and a market crash as well as an unwinding of a lot of private credit and private equity which is chasing AI. Banks are most likely involved to a degree. The only good news is that after the crash, the hyperscalers like Amazon and Google will still have rock-solid businesses and franchises and the stocks will likely be available then at a good price. I give this narrative also a 30% chance.
Then we turn to the intermediate case. Nvidia keeps financing the whole shebang. The hyperscalers keep going into debt and issuing more stock to finance the whole shebang. The data centers get built, but much more slowly and with many more hiccups than the buildout is modeling. The Chinese keep making a pretty good product that does the job for much of computing needs, which siphons off demand from Anthropic and OpenAI. And this does lead to a price war, or the construction of AI data centers is so great that it dwarfs demand. One analyst I follow, Brad Safflo of PAA Research, put out a report recently, was a good report, where he estimated that US data center capacity will be double the demand by the end of 2027. That's Brad's estimate, but if he is right, it will mean that AI capex spending will then slow in 2028. It's also possible that the political winds stay negative on allowing data centers, and there is not enough capacity coming online to meet demand. Either way, the future looks iffy. This narrative assumes that AI is a great technology, but there is simply way too much capex chasing demand. Result, the market unwinds much of its AI gains, and we have a bear market for a time. I gave this narrative a 40% chance.
And then we turn to the doomsday scenario. This is my favorite narrative because it's so ridiculous and yet seems so possible to so many people. Just because some old sci-fi novels predicted part of the future doesn't mean that all sci-fi predictions come true. Machines taking over the earth like in Terminator is not in the foreseeable future. I know that. And more importantly, the heads of Anthropic and OpenAI who are peddling this narrative know it too. So why are they spouting such nonsense to every politician who will listen? Obviously, if everyone really, really, really thought that the world was coming to an end, we wouldn't need to take these companies public or build data centers or even listen to this podcast. We'd go home and kiss our loved ones and hug the dog. They peddle this nonsense and surf the fear curve they've created because they want to scare the government into regulating the competition out of existence. I gave this narrative about a 0% chance.
Now that we've set the stage, let's turn to the AI investment story. Both the positives and the negatives. Over the past few years, AI has sent the stock market into a frenzy. In fact, from 2023 through 2025, the Magnificent 7 were responsible for roughly half of the S&P 500's total return. Hyperscalers are pouring hundreds of billions of dollars into AI data centers. At times, investors have become euphoric. But before we get to AI fundamentals, let me first explain how AI dominates the market. Today, infoch constitutes 40% of the S&P 500 with most of that related to AI in one way or another. Then add companies that are not technically in infoch but are techy like Google and Amazon and you reach 50%. The utility sector, which is 3% of the S&P, is also AI related in that it is spending 250 billion this year for the power needs of AI. Then add companies in the industrial sector that help build AI data centers like Eaton and Rockwell. Then add industrial companies that supply the power like GE Vernova and Quanta. Then add financial companies that are financing AI like the large investment banks and the alternative asset managers, and add all that up and you are well north of 65% of the S&P. The only two sectors that are not involved with AI at all, as far as I can tell, are staples and healthcare, which add up only to 14% of the S&P.
What about bonds? You say about 15% of existing corporate bonds are AI related, but over 50% of this year's issuance is AI related. So bonds don't provide the same degree of diversification that investors are looking for. Again, I want to emphasize that the entire market is not only directly related or influenced by AI, but the direction of the market hinges on the future health and success of AI. Before we get to the risks of AI, let's first examine the bull case, and it's a very powerful one. Bulls argue that the AI revolution is the largest technology buildout in history. Bulls also argue that this is only the third inning of a nine-inning game. Capital keeps flowing into compute, power, and data centers because the demand signal is overwhelming. The hyperscalers have around two trillion of multi-year AI cloud related backlog. Currently, the binding constraints have shifted from capital to power permitting and memory supply. It's a wonderful story with one simple underlying assumption: that AI is going to be transformative. It's going to transform software and medicine and everything else you can think of. AI will change it. Why else would great companies like Google, Amazon, and Microsoft be spending so much on AI capex? That's the bull case. And for the last 2 years, it has dominated the markets.
And for 2 years or so, AI bulls have been right. But that does not mean that they will be right forever. Let me first point out that everything that is being said about AI was said about the buildout of the internet, that the internet was going to transform the world. And that prediction proved to be totally true. What also turned out to be true was that the first generation of internet companies largely failed and failed spectacularly during the bursting of the dotcom bubble. It was only the second generation of internet companies like Google that truly succeeded. Thus AI may be transformative, but that does not mean that all the money currently being spent will generate an adequate return.
I want to touch on a theme that does not involve the risks to AI, but the risks to software companies from AI. This story has its own meme. It's called the SaaS apocalypse. What a great name. And what does it mean? It means that the cost of creating software because of AI has collapsed. As a result, AI is going to replace, so the story goes, all existing software with the exception of cyber security. Cyber security remains necessary because AI makes it easier to hack. The SaaS apocalypse narrative dominated software stocks from the summer of last year through the early summer of this year. Thereafter, most software stocks had a nice rally as the market realized that the problem with the SaaS apocalypse narrative is that it is overly simplistic. Being a successful software company requires more than just creating software. In fact, some software companies are so embedded in the operations of enterprises that the switching costs are too prohibitive. I think that probably applies to companies like ServiceNow and Salesforce.
However, many software companies do not have that kind of moat. Take Adobe for example. The perception is that Adobe is particularly vulnerable to AI. Adobe hit an all-time high of $630 in February 2024. Today, the stock is around $245. That's a decline of 60%. The threat to Adobe is not just idle speculation. Recently, it was announced that software engineer Brandon Thomas has built and released an open-source free alternative to Adobe's Creative Cloud Suite. Now, I have no idea how good this Adobe alternative is, but free is free.
The tentacles of the SaaS apocalypse extend beyond just the risk to software stocks. Between 2017 and 2022, private equity went on a buying spree of software companies, specifically SaaS, software as a service. Private equity was the buyer and private credit was often the lender to the buyout. Most transactions took place with 50% debt and 50% equity. That's an important figure because of the decline in public software stocks. Let's use Adobe's decline of 60% as a proxy, but make it a tad more conservative and say that the value of most private equity held software companies is down 50%, with cyber security software companies being a notable exception. A 50% decline in value means that the equity is wiped out. This will start to become clear when these loans start to be refinanced next year. Now, I don't think this private equity, private credit problem is anywhere close to being large enough to seriously hurt the economy, but it does mean that private equity investors are likely going to get hurt and hurt a lot.
And that leads me to the risks. What are the major economic risks to the AI story? Well, there are five predominant issues. I'm going to examine them in depth and then rank that risk in terms of its ability to derail the AI story. Issue one: from capital light to raising capital. By tradition, the hyperscalers, Google, Microsoft, Amazon, Oracle, are mostly software companies. Yes, they created cloud businesses years ago that required building out data centers, but the cost of those data centers was not prohibitive. So these companies were cash flow generating behemoths. They generated so much cash they barely knew what to do with it. All excess cash was used to buy back stock. Call these mostly capital light business models. AI has changed the game. The compute needs of AI are exponentially bigger than pre-AI needs. Also AI requires GPU chips and CPU chips and every other kind of chip. But the cost of a GPU chip is very, very high. The cost to buy Nvidia's latest GPU chip, Blackwell, can exceed $5,000 per chip. An AI data center can require at least 100,000 Blackwell chips, sometimes more. That means that the cost of just the GPUs is around $500 million. The cost of building the new AI data centers and filling them with the necessary equipment is prohibitive even for the hyperscalers.
Pre-AI Google, for example, generated over $80 billion in annual free cash flow. That cash flow is now gone. In fact, Google raised $85 billion in equity capital in 2026 to help fund its AI needs. This is an enormous change. The last time Google raised equity capital was 2005. Take Oracle as another example. Oracle is a large company, but at a market cap of $390 billion it is much smaller than its hyperscaler competitors, but it is still trying to play in this expensive game. To do so, it has had to take on a considerable amount of leverage. In the past 2 years, Oracle's total debt has increased from $94 billion to $155 billion. The ratings agencies are not amused. S&P downgraded Oracle to triple-B minus, which is only one level above junk. AI capex requirements have altered the business models of the hyperscalers from capital light to capital heavy. However, I don't think that this change in and of itself is enough to sink the AI story. If the AI story succeeds, then all this leverage will simply look like good investments.
One more point, the increase in leverage of the hyperscalers has fostered some financial shenanigans. In the days of Enron and the GFC, there were multiple examples of companies using off-balance sheet techniques to hide their leverage. The goal was to get the ratings agencies to believe that these off-balance sheet vehicles were not the responsibility of their creator and so the debt in the vehicle should not be counted as part of the company's leverage. The hyperscalers are starting to take a page out of that old book for their own benefit. For example, Meta is using an off-balance sheet technique to fund one of its data centers. Now, Meta is not a hyperscaler. It is creating its own LLMs and agentic AI apps. So, it is building data centers for its own use. Last year, Meta dipped its toe in the world of off-balance sheet financing in order to fund a $27 billion data center in Louisiana. The data center is financed with debt, but the debt is not on Meta's balance sheet. Instead, an off-balance sheet special purpose vehicle, SPV, was set up called Beignet Investor LLC. SPV rolls off the tongue. Hat tip though to the name, as anyone who has spent time in New Orleans knows that beignets are really delicious, but the financing is less appealing.
Technically, Meta owns only 20% of Beignet with the rest owned by Blue Owl and Pimco. So even though Meta is building the data center and has promised to rent the data center from Beignet for 20 years and bears the cost of any project delays and overruns, the debt will not sit on Meta's balance sheet. That's some pretty fancy footwork that's reminiscent of bad times past and doesn't pass the smell test. At least not to me. Perhaps with the memory of what happened to Arthur Andersen in mind, Ernst & Young, Meta's auditor, was not totally comfortable with this Beignet masterpiece of off-balance sheet financial engineering. In Meta's 2025 10-K, Ernst & Young flagged it as a quote critical audit matter unquote.
Now, I'm not going to quote the section from the 10-K because it is quite long, boring, and written in accounting language designed to obfuscate. What I'm going to do is translate that language into plain English, or at least the English that I think is what they're really saying. We at Ernst & Young are very nervous about this transaction. We are really not sure at all whether this should be off balance sheet. We relied on the company's description of the transaction, but we did not do any real work of our own to determine whether what Meta said about the transaction is really accurate. After all, we want to get paid and Meta is a big client. That's my slant on the accounting language. Admittedly, this off-balance sheet transaction is small. It's only $27 billion, but you should expect many more such transactions in the future and not just at Meta.
There is another aspect to AI financing that bears discussion: circular financing. Here, Nvidia seems to be the biggest culprit. It is either investing in or lending money to other tech companies and those tech companies then take that money and buy Nvidia chips. Here too, I don't think off-balance sheet or circular financing is enough on its own to sink the AI story. It just makes it more dangerous.
Now we talk about issue number two: power and data center buildouts. Even assuming that the AI story will work, a potential binding constraint is the lack of sufficient power and delays in building data centers. I need to hammer this point home. The AI story cannot work unless AI data centers are built. Now, the sums here are insane. Hyperscalers will spend $850 billion on building AI data centers just in 2026 and purchasing all the equipment to fill the data centers as well. But it takes time to build AI data centers. Here, politics plays a role. The AI industry is not popular. There is now a backlash against AI in many communities. People are afraid of the massive needs of AI data centers, needs for water and energy. So, the process of permitting AI data centers is being delayed in many parts of the country.
Energy is also a potential binding constraint. AI data centers utilize enormous amounts of energy. Now, the first choice is to hook up the data center to the electrical grid, but the grid is already stressed. Utilities are investing in new plants. But that construction also takes time. It takes a few years, for example, to build a gas turbine that powers a utility plant. That is why a stock like GE Vernova has done well because GEV is a large builder of gas turbines. Because utilities cannot satisfy all the needs of AI data centers, the hyperscalers are looking for alternatives. They are looking at nuclear or alternative gas turbine technology such as those provided by Bloom Energy. My suspicion is that AI data center construction is taking longer than hoped. But how much of an impact this will have is something I just don't know.
And now we turn to the third risk: concentration risk. Before we get to this concentration risk topic, let me scaffold it. Imagine there is a software company that is growing very rapidly but only has three enterprise customers. What would you say about such a company? I know what I would say: that this software company is a very risky investment. Should one of its customers go bankrupt or go elsewhere, growth would collapse. There is something like this going on throughout the AI food chain. Let's start at the top of the food chain, Nvidia. In its most recent quarter, Nvidia's revenue grew in excess of 100%. That's quite an accomplishment. That speaks to the bull case of AI. However, in Nvidia's 10-Q, it states that 70% of Nvidia's accounts receivable, which is a proxy for revenue, was from five customers. That's a lot of concentration risk for the largest company in the world. It's unclear who these five customers are. I assume it's the hyperscalers or someone adjacent to the hyperscalers.
Now let's go down the chain one level to the hyperscalers. According to reports from various Wall Street firms, something like 70% of hyperscaler AI revenue comes solely from Anthropic and OpenAI. 70% equates to around 25 to 35% of total hyperscaler cloud revenue. I cannot confirm these statistics, but they sound right given what we actually know about Oracle. Oracle has a backlog of around $600 billion and we know because the information is public that around 50% of that backlog is from OpenAI alone. And by the way, OpenAI accounts for over 20% of CoreWeave's total contracted order book. Hyperscalers are spending about $850 billion in capex this year and even more next year and that spending accounts for around half of the 2% GDP growth projected for 2026. Therefore, one must conclude that the health of the US economy is extremely dependent on hyperscaler capex and hyperscaler capex is highly dependent on the health of Anthropic and OpenAI.
Risk number four: OpenAI's risk of failure. What would it mean? Between the two, OpenAI is the weaker entity. Remember, OpenAI owns ChatGPT, which is rumored to be losing market share to Anthropic's Claude. The company has lost several key senior employees. Let's list them because the list is kind of illuminating. Number one, Denise Dresser, chief revenue officer, resigned in August 2026 after only eight months on the job. That probably means she gave up all her stock options as those usually don't vest for at least a year. Two, Brad Lightcap, former COO, resigned in August 2026 after 8 years with the company. Three, Caitlyn Kalinowski, head of robotics and consumer hardware, resigned in March 2026 over governance concerns regarding an agreement with the Pentagon. And four, Kate Rouch, chief marketing officer, and Kevin Weil, chief science research officer, both departed in April 2026. The most concerning departures are the ones that took place this August: Denise Dresser and Brad Lightcap.
Supposedly, OpenAI is getting closer to an IPO. That's the big payday for employees because it means that eventually they can sell some of their shares. That two such senior employees would leave now, that's an important data point. More importantly, OpenAI's financials, as far as we can tell, don't look so great. The Wall Street Journal reported that OpenAI's June quarter revenue reached $6.7 billion, up only 18% versus the March quarter. Compare that to Anthropic's revenue of $11-plus billion in the June quarter, which was up over 100% versus the March quarter. What's even worse than the relatively weak revenue growth is the explosion in costs. Again, according to the Wall Street Journal, OpenAI's costs reached $12.3 billion, up $3 billion versus the March quarter. So, in 3 months, revenue increased $1 billion, but costs surged $3 billion. Things are not going in the right direction.
Now what is the problem at OpenAI? Part of the problem is that OpenAI made a strategic mistake. I think it chose to chase the business-to-consumer market, B2C. It did so by giving its product away or underpricing it with the hope that later on it would raise prices. That strategy proved to be flawed and now OpenAI is chasing the business-to-business market, B2B. The problem is that it is late and playing catch-up with Anthropic, which chose the B2B market at the outset. Again, I think that OpenAI is potentially in trouble and this bears continuous monitoring. According to Ed Zitron, a commentator on Substack, to survive, OpenAI will have to raise $100 billion per year for the next several years.
Now, when you lose billions upon billions, appearances matter a lot. OpenAI is completely dependent on the kindness of strangers funding its cash flow needs. When a company is growing and very profitable, appearances don't matter nearly as much. Such a company can survive quite well. But when a company is not profitable and has an insatiable need for capital, appearances matter more than anything because if the narrative turns negative, raising capital becomes much more difficult. This is just me reading the tea leaves, but OpenAI's press is no longer all positive. The press has actually focused on the high-level departures and on the weak numbers that I just described. That's why I'm not predicting, at least not yet, that OpenAI will fail. I still think it's too early to make that kind of prediction. But it's not too early to think about it.
So, if OpenAI fails, Oracle is in immediate trouble because of the large increase in its debt. Oracle's debt rating is barely above junk. Now, what would happen to Oracle's stock price if OpenAI failed? Well, we have already had something of a preview. Last year when Oracle reported its backlog for the August quarter of 2025, what it calls remaining performance obligations, RPO, it showed an RPO explosion to $455 billion. It's now over $600 billion. In that quarter, the RPO was up over $300 billion in just 3 months. Now, prior to the earnings report, the stock was $230. In just a few days, it jumped to $330. And then analysts published reports pointing out that 50% of the RPO was from OpenAI. And the stock gave back all of its gains plus in a few months. Today, the stock is, get this, around $130, well below where it stood before it reported the August 2025 quarter. From the peak, the stock is down around 70-ish%. That decline is because the market perceives an over-reliance on OpenAI. Imagine what the market would do to Oracle's stock if OpenAI fails.
The ramifications of an OpenAI failure extend beyond Oracle. Remember I said that AI capex accounts for 50% of US GDP growth. While the other hyperscalers are not quite as dependent on Anthropic and OpenAI as Oracle, they are dependent enough. If OpenAI failed, the hyperscalers, I am sure, would cut back on their capex. And I'm starting to think that the demise of OpenAI would push the US into an immediate recession. Here's an important point. Timing is everything. Eventually, the AI industry may become more diversified and less dependent on Anthropic and OpenAI. If either company fails within, I believe, a year, the ramifications would be severe. But if that failure happens a few years from now, the ramifications could be much less important. For now, the health of Anthropic and OpenAI is crucial.
So what would happen to particular stocks and sectors if Anthropic or OpenAI failed soon? First, the hyperscalers would go down. Failure of OpenAI would mean they would pull back on capex. Cloud revenue growth would slow. These stocks, Amazon, Google, Microsoft, and Oracle, just to name a few, they'd all correct. Also, the whole tech sector would correct. But the ramifications are much broader than just that. Investment banks and large banks. These stocks are at fairly close to peak valuations. Also, the investment banking cycle is super strong partially because of the financing needs of AI. Should those needs lessen, the investment banking cycle would weaken and these stocks would correct from their peak valuations. Industrials. There is a subset of industrial companies that are major beneficiaries of AI. They are in the power space like GE Vernova and Quanta or they are in the electrification or automation spaces like Eaton and Rockwell and these stocks would decline as well.
So what will do well? This is not a stock picking question but a reallocation question. Investors will reallocate to safety sectors and subsectors. In the safety sector category, think about healthcare and consumer staples. I'd also point out that within almost every sector there exists safety subsectors. So for example within financials the property casualty sector is considered the safety subsector. Since we are talking about capital reallocation and not stock picking let me flag three safety ETFs. One, LVHD. It's the Franklin US Low Volatility High Dividend Index ETF. Number two, the SPLV, Invesco S&P 500 Low Volatility ETF. And last, KBWP, the Invesco KBW Property and Casualty Insurance ETF.
And now we come to the last and probably the most important risk. Number five: there are no long-term moats. Until recently, Google had an insurmountable moat around its search business. No one could touch it. There don't seem to be any moats around AI. One day ChatGPT is on top, the next day it's Claude, and the switching cost between models seems to be perhaps zero but certainly very low. I can't overstate the importance of this risk. Today I hear a lot of positive energy towards Claude. People literally email me that their company or university is endorsing Claude. The argument is that enterprise reliance is a moat. To this I say the minute ChatGPT or Gemini or the Chinese models offer the same quality and a lower price, pressure will be huge to switch. There goes the moat.
The costs of AI are expressed in tokens. Think of tokens as three-quarters of a word. Every time you query Claude or ChatGPT, it answers your query and there is a cost to the tokens for the query. Last year, LLM providers dramatically undercharged for their services. They charged a subscription and the cost of that subscription was dramatically lower than the cost of the tokens. This fostered something called token maxing, which is a fancy term for enterprises telling their software employees to use AI 24/7 and to hell with the costs. Earlier this year, prices were raised and enterprises found their AI costs soaring. Uber, for example, blew through its entire AI budget in just a few months. Sometime this summer, token maxing ended as all enterprises became more cost-conscious.
While the token maxing saga was going on, Chinese open-weight models started to take market share. The important point here is that open-weight models are much cheaper than the models of Anthropic and OpenAI. The flagship models of Anthropic and OpenAI can cost $2 to $10 per 1 million tokens. Open-weight models can cost around 20 cents per 1 million tokens. So much cheaper. The models of Anthropic and OpenAI are superior, but you don't need a Mercedes-Benz for every trip. And this leads to the risk: the potential for a price war because open-weight models are just so much cheaper. Add in the fact that there are no moats and the probability of a price war explodes. This is the key risk. Anthropic and OpenAI do not have the balance sheets to withstand a price war. And should a price war break out, fears will grow that the entire AI story will just slow. And in fact, there are reports that both Anthropic and OpenAI are cutting prices right now out of fear of losing market share to the cheaper models. Stay tuned.
And now we come to our conclusion. As you can see, there is a lot going on. The AI story is complex. It is also driving the US economy right now. The risks I have outlined are real and could, I emphasize, could derail the AI story. If Anthropic or OpenAI get into trouble in the next year, that would be telling. However, if it is years from now, the AI story may become so diversified by then that it won't matter. It is still too early to tell, but we will be watching very closely.
So, with that said, what should investors be doing right now? Well, the first thing I'd say is that being cautious about parts of the AI trade is not the same thing as being bearish on AI. I think AI is real. I think it's already started changing the economy, and I think an enormous amount of money will be made from it. The much harder question is who actually makes that money. Personally, I'm much more interested in the picks and shovels. Nvidia obviously sits at the center of the compute buildout. You've got companies like Micron supplying memory, Arista, Cisco supplying networking, and then on the power side, you've got companies like GE Vernova and Eaton. These businesses don't necessarily need to predict which LLM ultimately wins. They're selling the infrastructure required to build the entire thing.
Where I'd be much more careful is in businesses where investors are assuming enormous future profits despite very low switching costs. Who are we talking about here? Mainly the LLM providers themselves, but potentially to parts of software too as agentic AI becomes more capable. The question I keep coming back to personally when analyzing these businesses is very simple: how difficult is it for the customer to leave? If the answer is not very, you'd better be sure about the price you're paying.
And ultimately, that's how I'm looking at this entire AI boom. I'm certainly not trying to call it top. I'm not telling you this is 1999 or 2008 all over again. At least not yet anyway. Maybe these companies deliver everything the market is expecting and more. But when valuations assume an extraordinary future, you don't necessarily need the technology to fail for the stocks to disappoint. You just need reality to fall short of expectations. And right now, between Chinese competition, the lack of moats, the enormous commitments sitting on OpenAI and Anthropic, and the physical limitations of the power grid, there are a plethora of reasons to keep watching those expectations very carefully.
And that's the end of our lecture. But before I go, let me just say that if you want to support the show further and get additional exclusive content, you can check out my Substack at steveeisman.com. Premium membership to Substack offers a premium-only Wednesday episode, which can be a masterclass or deep dive interview. And premium subscribers receive early access to our Monday interviews and Friday weekly wraps, as well as full access to my daily chats and the ongoing conversations among members. Free membership in Substack offers access to my daily notes and previews of our Wednesday premium interview and previews of all newsletters. The link steveeisman.com is in the description. Thanks for watching.
This podcast is for informational purposes only and does not constitute investment advice. A host and guests may hold positions in stocks discussed. Opinions expressed are their own and not recommendations. Please do your own due diligence and consult a licensed financial adviser before making any investment decisions.
