Steve Eisman — The AI Debate Gets More Complicated: Microsoft Has a Win, Meta Stumbles

Charter is the cable company behind Spectrum — home internet and TV. It was Eisman's January recommendation on two legs: the shares were extraordinarily cheap, and its heavy network spending was about to fall away, turning into free cash he expected the company to use buying back its own stock. That only works if the business itself stabilises.

It hasn't. Charter keeps losing broadband customers, and the losses are getting worse rather than better: after an encouraging fourth quarter, the first quarter deteriorated and the second lost another 172,000 subscribers, "much worse than expected." Management insists the worst is behind them; he no longer believes it. "Bottom line, I give up. Simply put, I made a mistake."

The important part is the rule he applies to himself. Owning it now would mean owning it only for the cheap price, since the fundamental leg of the argument has failed — and that is what he calls thesis creep: quietly swapping the reason you bought something for a different reason so you never have to sell. "I hate thesis creep." So he is selling, while explicitly leaving the door open — "if the fundamentals ever turn, I could come back."

Two details worth copying. First, he ignores the price action: the stock is up 15% in a week, but "the rally in the stock in my view has nothing to do with Charter" — it is money fleeing AI names and landing anywhere else, which tells you nothing about the business. Second, he hasn't sold yet, on principle: "I recommended the stock to my viewers and I strongly believe that I should not sell until I inform my viewers of my opinion change."

Meta owns Facebook, Instagram and WhatsApp, and is spending enormous sums to build AI. The quarter was "quite bad" for a reason that has nothing to do with its advertising business, which grew 28%: costs grew twice as fast, up 55%, with research and development alone jumping from $13 billion to $22 billion. Profit per share fell 13% and missed.

The number that alarms him is free cash flow — the cash left after paying for everything, including the data centres. Meta earned "a mere 784 million, which is basically nothing." For a company of this size that is essentially breaking even in cash terms, which "shows how capital-intensive this AI game has become."

Then two forward-looking problems. Guidance for next quarter came in below expectations, and the capital-spending range was narrowed by raising the floor — the company is committing to spend more, not less. And an off-balance-sheet number most reports skipped: $279 billion of future lease commitments related to AI that don't yet appear on the balance sheet, up 53% in three months. Those are obligations to pay, just not shown as debt yet.

The contrast with Microsoft is the episode's core point. Microsoft spends just as heavily but owns a cloud business (Azure) that is growing 43% and pays for the spending. "Meta does not have that business and is trying to play just in the LLM space, which is expensive and not yet lucrative enough."

Microsoft entered this print hated from both directions — as a software company AI might replace, and as a data-centre builder whose spending has exploded — which is why the shares were down 19% on the year before reporting.

The quarter answered the second worry. Profits rose 30%, revenue 18%, and the number that matters most, Azure cloud revenue growth, accelerated to 43% from 40% the prior quarter. Acceleration is the key word: growth speeding up means the AI spending is producing revenue, not just cost.

The blemish is the same one that has punished everyone else — free cash flow fell 23% because of the building. Eisman notes it and moves on: "still, I would characterize this as a very good quarter." Under his new framework Microsoft is the ideal shape — a hyperscaler with a genuine moat (the sheer cost of entry) plus established software revenue that doesn't depend on winning the model race.

Bloom Energy makes fuel cells — small and medium-sized generators that convert natural gas straight into electricity on-site. That matters because the constraint on new data centres is increasingly power: the grid can't deliver enough of it fast enough, so an operator that can generate its own electricity next to the building skips the queue.

The results were extraordinary: profit per share up 680%, revenue past $1 billion for the first time and up 166%. The shares are up more than 90% this year.

His valuation framing is the useful part. On next year's expected profits the stock costs 73× earnings, which sounds absurd. But because profits are growing so fast, the same price is 37× the 2027 estimate and 23× the 2028 estimate — the multiple decays as the earnings catch up. That reframes the question from "is it expensive?" to "will the growth actually arrive?", and he answers accordingly: "if the AI story keeps going, I would expect Bloom stock to continue to perform. But again, the AI story has to keep going." The position is a conditional one, and he says so.

FICO owns the credit score American lenders use. Eisman is short it — he profits if the shares fall. The thesis is that a rival, VantageScore, is about to take significant share in mortgage lending, where FICO has been able to raise prices at will because there was no alternative.

On the surface the quarter was excellent: profit per share up 42%. But he takes it apart. The growth came from years of price increases plus lower-than-expected costs, not from selling more — and revenue actually missed, with soft guidance for the quarters ahead.

The rule this illustrates is the whole short: "a company whose entire monopolistic business model is potentially under assault can show no signs of weakness." A monopolist is valued on the belief that it fully controls its own pricing and volume. Any evidence to the contrary — a revenue miss, a cautious forecast — is disproportionately damaging, because it is the first data point consistent with the bear case. The stock fell 17% in a day.

Quanta is the contractor utilities hire to build and maintain power plants and transmission lines. It is the least glamorous way to own the AI build-out and, on these numbers, one of the most direct: data centres need electricity, electricity needs new generation and grid, and someone has to physically build it.

He has owned it for a long time and calls the quarter "unbelievably powerful": profit per share up 71% and well ahead of forecasts, revenue up 41% to $9.6 billion, and — the part that matters most — management raised its guidance for both profit and revenue for the full year. Raising guidance means the order book, not just the quarter, has improved.

Meritage builds homes, mostly entry-level, and is Eisman's standing housing recommendation. The quarter was genuinely mixed: new orders fell 9% and revenue fell 14%, both slightly worse than expected, and profits were down 30% from last year.

What he weights instead is what the company did with its cash. Profit per share still beat by 4% because of better margins and tighter overhead, and free cash flow came in much stronger than expected — because management is deliberately buying less land and spending the money buying back shares while the stock is cheap. It repurchased $100 million in the quarter, 2% of the company, and 5% of all shares outstanding since January.

That is the mechanic worth understanding: in a weak housing market a builder can shrink its share count faster than its business shrinks, so profit per remaining share holds up. Land spending is discretionary; buying it back at a discount to what the land and homes are worth is the better use of the money.

PayPal processes online payments. The quarter technically beat expectations — profits down 1%, revenue up 3% — and he dismisses it in three words: "but so what?" A business growing 3% while Visa's payment volume grows 10% is losing ground, which is his running point about payments: outside Visa and Mastercard it is a brutal, moat-less industry.

The live event is a takeover bid: Stripe and the buyout firm Advent have offered $60 a share, and PayPal's board says that's too low. Eisman's reaction is a governance judgement, not a valuation one: "I hate it when managements play chicken. PayPal's business is under assault from large players like Apple and Google. The company should sell." A board holding out for a better price is making a bet that the business will be worth more later — and here he thinks the trend runs the other way.

Visa runs the rails that card payments travel over, taking a small fee on each transaction. Eisman has owned it for years as one of only two "impregnable franchises" in payments.

The quarter was "powerful": profit per share up 20% and ahead of forecasts, revenue up 14% and also a beat. The number he actually reads it for is total payment volume — the dollar value of everything spent on Visa cards — which rose 10%. That is a direct measure of consumer spending, unfiltered by any one retailer's problems: "no signs here that the consumer is slowing down." Paired with Procter & Gamble's flat organic sales in the same episode, it is his K-shaped economy in two data points.

Mastercard is the other half of the payments duopoly and reported the same picture: profit per share up 21%, revenue up 14%, and payment volume up 8%. Two independent networks reporting high-single to low-double-digit volume growth is stronger evidence about the consumer than either would be alone — which is why he checks both.

Procter & Gamble sells everyday household staples — Tide, Pampers, Gillette. Because everyone buys these things, its sales are a clean read on the ordinary consumer rather than the affluent one.

Profit per share fell 3%. The worse number is organic revenue growth of 0% — "organic" strips out acquisitions and currency effects, so it measures whether people actually bought more. Zero, in a year with inflation, means volumes are falling. Set against Visa's +10% payment volume, that is the K-shape: the top half spending freely while the mass-market consumer holds still. "Signs of the K-shaped economy are everywhere."

Starbucks has been a multi-year turnaround attempt that repeatedly failed to turn. This quarter it did: profit per share up 70%, and — the metric that counts for a restaurant chain — same-store sales up almost 8%.

Same-store sales measure how much more the existing shops sold versus a year ago, stripping out new openings, so it can't be flattered by expansion. Note that total sales actually fell slightly (the chain is running fewer stores), which is exactly why the same-store number is "the figure analysts care the most about": the stores it kept are performing.

Robinhood is the retail trading app. It beat comfortably — profit per share of 62 cents against 43 expected, revenue up 32% — and the stock still fell after hours because one line looked bad: crypto revenue dropped 38%, to $100 million.

Eisman treats that as noise. The revenue growth came from options, equities and prediction markets, meaning the business diversified away from crypto rather than losing it — the same result the market briefly read as a problem. By the next morning "the stock had reversed and was up." A useful reminder that an after-hours reaction is a first draft, and that a declining segment inside a growing whole is only bad if it's the segment the thesis rests on.

Apple's headline numbers were fine: profit per share up 29%, total revenue in line — "no problems here." The disappointment is in the two lines investors actually own the stock for.

Services — the App Store, iCloud, subscriptions, the high-margin recurring revenue that justifies Apple's premium valuation — missed expectations. And China, its most important growth market and the one most exposed to local competition and politics, was also weak. Shares fell about 4% after hours.

He owns it. The tension worth holding is that Dan Ives made the bull case four days earlier on this same podcast — Apple as the toll booth on consumer AI, monetising 2.5 billion devices without spending on models. This quarter is a reminder that the toll booth still has to keep collecting: if services growth stalls, the "we don't need to spend" argument gets harder to make.

Amazon's retail business is large and slow; the part that moves the stock is Amazon Web Services, its cloud computing arm. AWS revenue growth "jumped to 37%," which is the same signal Microsoft gave with Azure at 43%, and profit per share was "up huge." The stock rose more than 7% after hours.

In this episode's framework Amazon is a hyperscaler with a real moat — the cost of building at this scale is itself the barrier — and, crucially, one whose cloud growth is accelerating rather than merely absorbing capital. That is the distinction he draws between the companies whose AI spending is producing revenue and those, like Meta, where it is only producing costs.

Google appears on both sides of the new dividing line. As a hyperscaler — a builder and operator of the giant data centres everyone else's AI has to run inside — it is one of the four with "real businesses here," protected by the fact that "the amount of money it takes to be a hyperscaler is insane. And that expenditure itself is a moat."

As an LLM provider — a maker of AI models sold to customers — it is in the vulnerable group with Anthropic and OpenAI, facing cheap Chinese open-source alternatives and customers who switch between models to cut costs. The difference is that Google (like Microsoft) has "multiple revenue streams from established businesses which are very unlikely to simply disappear," so a weak model business is survivable rather than fatal.

One more role: Google is his benchmark borrower in the credit market — the company lenders "know can pay the money back," against which CoreWeave's 9% cost of debt is measured.

Oracle is the fourth hyperscaler in his list, and the illustration of the systemic risk he now watches. Its order book — contracted future revenue — is over $600 billion, and "around half is from OpenAI."

That single fact links the whole chain: if the model companies get into trouble, the backlogs of the data-centre companies stop being assets. It is the reason his selloff trigger is not a stock or an economic number but the financial health of two private companies.

Worth reading alongside Gil Luria's argument on the same backlog four days earlier: Luria's point was that the market values that backlog at zero, making the shares attractive. Both can be true — a backlog valued at zero is cheap and concentrated in one customer.

CoreWeave rents out AI computing capacity, and it has to borrow to buy the chips and build the buildings. It is raising $2.6 billion of debt at a yield of more than 9% — very expensive money in a market where the largest technology companies borrow at a small fraction of that.

Eisman's interest is in what the price says rather than the company itself: "there is something of a credit cycle here as fixed income investors are discriminating between the large companies like Google that they know can pay the money back and smaller newer companies like CoreWeave that are more risky."

That is a meaningful change. When credit markets stop treating an entire theme as one risk and begin charging very different prices within it, the era of easy financing for the weaker participants is ending — and bond investors usually notice before equity investors do.

Carried over from the Ives/Luria interview and endorsed here: "companies that have not invested in their products, I think, are in big trouble. That applies to some public companies" — with Salesforce named as the example. The mechanism is that corporate technology budgets have to fund AI spending by cancelling something else, and the first cancellations are the products customers feel they aren't getting value from.

He extends the same logic to software owned by buyout firms, which is where he expects the real damage: private equity "has been milking those companies as opposed to investing in them."

Anthropic makes one of the two leading closed-source AI models. This episode gives it a specific and unusual role: not just a company he's sceptical of, but the thing he would watch to time a market-wide sell-off.

The problem is structural. Closed models can't stop customers leaving, because "enterprises are switching between models and using cheaper open-source Chinese models in order to control costs" — so there is no moat, "or at best the moats are shallow." And unlike Google or Microsoft, Anthropic has no other business to fall back on: "they don't have the breadth of revenue streams."

Why that becomes everyone's problem: the data-centre companies' order books are largely made up of promises from these two firms. "If the lack of moats begins to cause them problems, then the entire AI ecosystem could go through a correction phase."

He also repeats, in his own voice, the accusation Gil Luria made four days earlier — that the widely-promoted story of AI destroying jobs "could be propaganda propagated by Anthropic and OpenAI so that the federal government will step in and regulate AI to the benefit of Anthropic and OpenAI. Demanding regulation based on a false narrative would be a very disturbing way to create moats." His own view is that AI produces "job dislocation, but net job creation."

OpenAI is the other closed-model provider, and the one that makes the concentration concrete: roughly half of Oracle's $600-billion-plus order book is a commitment from OpenAI. That means a large share of one listed company's future revenue depends on the finances of a private one that most investors cannot examine.

The same moat problem applies — cheap open-source competition, customers switching to control costs, and no established side businesses to cushion it. Which is why Eisman's answer to "what would cause a real sell-off?" is not a Fed decision or an earnings season but "the health of Anthropic and OpenAI."

Stripe is the large private payments company that, with the buyout firm Advent, has bid $60 a share for PayPal — an offer PayPal's board says undervalues it. Eisman disagrees with the board. Stripe appears here as the acquirer, not as a rated view; note it also featured in earlier episodes as part of the consortium taking on Circle in stablecoins.

SOXX is a fund that holds the major semiconductor stocks, so its price is a single gauge for the whole chip complex. He uses it to size the damage: down 23% from its June 22 peak and down 4% in the week, against a NASDAQ that is down 7% from its January 2 high.

The gap between those two numbers is the point. Semis have fallen more than three times as far as the broad index, which is what "investors have displayed their nervousness by selling tech stocks and all AI related plays" looks like in practice — the selling is concentrated in the AI supply chain, not spread evenly.