Winners and Losers of the AI Era

Search is, in his words, "one of the most perfect, beautiful business models of all time. And the purest aggregator of them all. Like scales in every direction, doesn't have to invest any money to do it." The problem with such businesses is that percentage margins are wonderful and there is nowhere to reinvest, so cash just piles up.

His Berkshire parallel is exact. Buffett took See's Candies profits into BNSF, a far worse-margin business whose absolute dollars were so much larger that BNSF threw off more free cash in one recent year than See's had in its entire life. "Once your capital gets so large, you start operating in a world of absolute numbers as opposed to percentage numbers."

If AI is intelligence and its TAM is basically all white collar work, and eventually with robotics... the absolute profits available here, even if the margins are lower, is so much larger that will we look back on Google search like See's Candies? It feels like that's what's happening.

That reframes the equity raise, which he first calls shocking: "it's Google. They can't raise money. Like, why are they issuing equity? Why are they reducing their upside if they believe so strongly in this?" His answer is that you end up with a smaller percentage of an astronomically larger pie, and "at the end of the day, no one's gonna be complaining."

On Berkshire's own stake: "are they actually not just an investor in Google, but a model for Google and where they're going?"

And the floor condition, said almost in passing: "Google just needs search to not die too quickly."

Asked which of the top tech companies has the most interesting setup, he does not hesitate. The mechanism is the first-best-customer flywheel: Amazon builds for itself at a scale that gets anything off the ground, then sells it outward.

He corrects the AWS origin myth along the way. It was not spare retail capacity, and retail moved onto AWS late. What drove it was wanting compute you could plug into without a meeting. Logistics ran the same play in reverse: use UPS and FedEx, build it in-house, then open it to third parties.

The sharpest observation is how Amazon launders immature silicon:

The early versions were terrible. But if you're using their managed services, like the Redshift database service, they don't tell you what the processor is underneath that... So they can put all their crappy processors underneath the services they're selling, and that gives them the volume and the capacity to iterate them and get better.

Because they were the first best customer, Graviton got better and Trainium got better, and Trainium now runs Anthropic workloads.

Microsoft is deliberately off the frontier, and that is why it has cash. Its play is middleware: sit between enterprises and the models, provide the harness, manage model churn, disintermediate the labs. He names it precisely: "It's the IBM play of the nineties."

The Gerstner history he is drawing on is unflattering and he knows it. Gerstner's insight was that IBM was mediocre at everything, which is the price of monopoly:

Once you've been a monopoly, you kinda lose your capacity to be good because you didn't need to compete anymore... And the problem is that once you lose that muscle, it's gone. You're just sort of fat and flabby.

So breaking IBM up would have exposed the mediocrity. The real asset was being big and trusted. IBM built a consulting army, inserted middleware between mainframes and the web, brought corporate America online, and bought thirty years. Middleware "saws off the sharp edges" and delivers a lowest common denominator, and enterprises buy it anyway.

He also thinks the new pricing is genuinely dangerous. The old seat licence was "a thoughtless revenue stream" attached to headcount and mentally filed as something like capex. Metering breaks that in three ways: most companies budget annually and cannot make monthly consumption decisions; the decision has to be re-made continuously rather than once; and customers start looking at the bill and asking what each component is actually worth.

They had to do it because the heavy AI user "costs way more to Microsoft than $100 a month," but they need everyone else not to think too hard, "because that breaks the model in very surprising ways."

Then the knife. Systems of record were sticky because migrating them was tedious and repetitive, which is exactly what AI is good at. And Microsoft is not really a system of record anyway:

It's user interface. It's like where you actually interact with the computer. That's the part when you see codex, claude, coworker, whatever, it is aimed like an arrow to the heart of what Microsoft has.

His verdict, held in tension deliberately: "Their strategy is sound. It's also desperate in an existential way and also in a they might pull it off because they're desperate sort of way."

Two separate arguments, and the second is the valuable one.

First, on being on the frontier at all: "there's a very good case to make that it is more reckless to not be on the frontier if you're a digital company." He credits Zuckerberg explicitly, calling Meta's position "one of the purest manifestations of founder energy for better or for worse," while carefully not endorsing the spending. Hiring a completely new team and starting from scratch is "pretty insane."

Second, and this is the part worth extracting, the models pay off in ads. They help in two distinct ways. Generation becomes a verifiable domain, because the question "is this generated image good?" resolves to "did the ad sell?":

They have this massive advantage, this huge liquid market that is a verification machine where the verifiers are humans deciding whether they click on that ad and make a purchase or not, but they're doing it at global scale.

And matching moves from crude embedding similarity to genuine prediction: "Meta's gonna look at people and say, this person probably wants to see this next." A few percentage points is billions of dollars, which he says alone justifies frontier investment.

He also makes a margin point nobody else makes. AI-generated content is accretive for YouTube, where inference could cost less than the creator revenue share, but dilutive for Meta, because Meta currently pays nothing at all. "Instagram is this unbelievable product that generates all this money for which Facebook pays $0 for content."

His diagnosis of the TikTok miss is that Meta misclassified the category. TikTok is not a social network, it is an entertainment product, and limiting content to your own social graph is an artificial constraint. The contrarian upside he offers: in a world saturated with AI, wanting human connection rises again, and social networking, meaning the group chat rather than the feed, gets important.

What visibly annoys him is the communication. Meta has the best ad business in the world, one he considers a societal good, and will not say so. "Mark Zuckerberg has the same problems that Altman does. He doesn't love ads." He argues the hole opened when Sandberg left, cost them during Apple's App Tracking Transparency, which he calls "one of the worst antitrust violations in the history of technology," and costs them now in permission from Wall Street to invest, compounded by having burned "cumulative 100 some billion dollars on Oculus."

The most relaxed take in the episode. Apple sat AI out and may have been lucky rather than smart, but the aggregator logic protects them: they own access to customers, so suppliers come to them rather than the reverse, and they can source models as needed.

The cost argument is genuinely structural. If consumers mostly want a chatbot rather than productivity, Apple can serve that, plausibly on device, and "they actually don't even need to pay for inference costs either because they're using the customer's electricity." He is careful that we are not there yet, noting Apple currently leans on Google Cloud and Nvidia chips.

He also defends them on fit, which is the most interesting part:

AI is this probabilistic endeavor. Apple is the king of deterministic products... you ship that iPhone, you ship it once, and it's gotta be good. If it's bad, it costs you billions and billions and billions of dollars. Apple's never had an iPhone recall.

Hence: "I generally prefer companies to do what they're good at. So from my perspective, I'm fine with Apple not doing AI. I want them to keep making great devices."

The risk he names is positional rather than competitive, and it is the Microsoft mobile trap. Microsoft did not miss mobile; it built mobile as a small PC because it assumed the PC stayed the centre. If ambient AI becomes the locus and Apple assumes the phone stays the centre, that is the same mistake.

The most analytically aggressive section. His opening judgment is that the position is "definitely unnatural," and that the apparent margin durability is an illusion created by where the price cuts are being booked.

The circular financing, in his reading, exists to lower the counterparty's cost of capital. And the counterparty's cost of capital falls for a specific reason:

They get a lower cost of capital because Nvidia assumed risk. This is my point before. Risk never disappears. It just appears somewhere else. Taking on risk has a price.

He sizes it as a distribution rather than a certainty. There is a world where AI never stops and Nvidia captures all the upside, and a world where a neocloud backs up, compute floods the market, and "Nvidia is paying for compute that no one wants." The expected value is somewhere in between, and that expected value is a reduction in profitability:

If you actually look at their business holistically, what that is is a price cut. Now the price cut didn't show up in margins... We have seen price cuts. They're just manifesting in these very bizarre sort of ways.

He adds, drily, that "moving stuff off the balance sheet by and large works."

The long-run threat is the hyperscalers, and the mechanism is capital rather than silicon quality. Google and Amazon are bigger, fund cheaper than neoclouds, and critically they sell TPUs and Trainium as commodities while Nvidia sells differentiation. Because nobody chooses AWS in order to use Trainium, selling it outside cannibalises nothing.

On CUDA he is direct: the moat "is dramatically diminished because the models don't care what they run on." What remains is fungibility, and he rejects Musk's stated reason for buying Nvidia: "No. You'll buy Nvidia because they're the most fungible." If your model is build-and-rent, you buy the thing that is easiest to rent out.

He reads the sovereign-cloud and enterprise push as deliberately targeting buyers too small to do porting work, exactly as Intel kept government and enterprise while AMD took the hyperscalers.

Finally, a neat inversion: Nvidia would benefit from a power shortage, because scarce power makes token efficiency decisive and Nvidia is still most efficient. "Probably the biggest problem for Nvidia over the last couple of years is I think The US has actually brought a lot more power online than expected." He suspects Huang expected power to become his moat sooner than it has. Every extra month of adequate power is another month for Trainium and TPU to close the gap.

His most original call, and the most falsifiable. TSMC's conservatism was rational for TSMC and is about to cost it. A fab is a thirty year asset, so overcapacity poisons years of returns, and the company is structurally biased toward restraint. The risk did not vanish; it landed on customers as foregone profit.

He is generous about the past. Morris Chang is "a one of one on the Mount Rushmore... of the greatest and most impactful tech executives of all time," for coming back, firing the leadership that cut spending in the great recession, and investing straight through the downturn just as the iPhone launched. His verdict on current management is cooler: "TSMC, they were pretty conservative, to be totally honest."

The consequence is the interesting part. His long-held view was that second-sourcing to Intel never made rational sense. You would endure the pain of teaching Intel to be a foundry while your competitor simply used TSMC, who are excellent and easy to work with. Nobody pays for insurance whose expected value is negative. Acute scarcity changes that arithmetic, because the alternative becomes foregone revenue.

This scarcity is what ultimately saved Intel. I expect at some point that they're gonna announce some major partner for the first time. It's gonna be a big deal. But ultimately, TSMC brought it on themselves. It's the cure for high prices is high prices thing.

The elegant conclusion is a policy one. You do not solve the Taiwan dependency with subsidies or exhortation: "The way to solve the geopolitical problem of dependence on TSMC is to come up with a compute use case that is so massive that everyone is economically incentivized to bring other people up to speed. And then we get the geopolitical insurance for free."

His cleanest commodity case study, and he is bearish on the makers' behaviour rather than their current numbers.

The history: many makers, each boom draws entrants, capacity arrives late, prices collapse, players wash out. Samsung took the market by investing into downturns, which "requires a ton of guts and a ton of discipline and a ton of money," and wiped out the Japanese. The survivors consolidated to three and became disciplined: "we're not colluding, but we all are on the same page about let's not do that." That discipline then collided with a genuine secular demand shift they were slow to believe.

Then the warning, which is the sharpest thing he says about any component maker:

I've analogized memory makers to Iran. The issue with the Strait of Hormuz is it's very effective. It's more effective if you don't use it because then it's always hanging out there as something you could do. Now they did it. Turns out it worked. But The UAE and Saudi Arabia, they're gonna build pipelines. They're gonna build new ports. They're not gonna let this happen again... My concern I have for the memory makers is they might have done the same thing.

Concretely: Apple lobbying to admit Chinese memory, and every architecture team now optimising to use less memory. "I think the memory makers probably screw themselves in the long run by creating such a massive target on their back."

Asked about five potential winners, he answers mostly on culture and structural incentive.

Never discount the power of belief. They think they're creating God. The most impactful things in history have usually been fueled by religion.

He types them: OpenAI is mainline protestant, goes to church on Sunday; Anthropic are the evangelicals, "all in. It is core to their belief." And he adds a hard-nosed rider that runs through the whole section: "The fact you need to make a business work for you to survive goes a very long way." Labs that must earn money have an advantage over divisions that need not.

OpenAI

Gets his most pointed criticism, and it is a strategy error rather than a technical one. They replayed Dropbox at a hundred times the scale. Dropbox built a beloved consumer product, found consumers would not pay enough, and rebuilt for enterprise. OpenAI "sold a lot, but they didn't sell enough."

His governing rule is that consumers will not pay for software and do not want to be productive: "I spent all day working. Why do I wanna come home and be more productive? I wanna sit on the couch and watch reels." Therefore consumer AI must be advertising, whose great structural advantage is that the advertiser bears the price increase, so monetisation scales with volume instead of fighting willingness to pay. That is Netflix's structural problem and ads do not have it.

Had they leaned into advertising immediately as soon as ChatGPT was a hit, I think they would have a killer ad product right now. I think that Google would be in much bigger trouble. I think Meta would be in much bigger trouble.

Now they are pivoting to ads late while simultaneously scrambling into enterprise "because Anthropic is kicking our ass. So I'm not quite sure what they're doing there." Where he does find them interesting is ambient and home hardware.

Anthropic

The enterprise winner in progress, and the beneficiary of belief plus commercial necessity. Anthropic recurs throughout as the swing customer for non-Nvidia silicon: Google TPUs, Trainium, and externally rented capacity. He credits both Anthropic and OpenAI with real recent acceleration, which he reads as evidence that AI improving itself is genuinely happening.

xAI, Meta's lab, SpaceX AI

Grok and Meta are "chasing" rather than at the frontier, and how they fare is his test case for whether catching up is still possible. SpaceX AI he rates weakest, but for an interesting reason: the space data centre is differentiated enough that it may not need a proprietary model at all. "So why are you wasting billions and billions of dollars in the meantime?" He likes the Cursor acquisition as a tactical fit.

China

Steady and unalarmed. Chinese labs are "very capable, very smart, and also definitely distilling these models to stay about six to nine months behind," which he considers a decent equilibrium favourable to the US. He is sceptical they close that final gap, because "getting to the leading edge in that last six to nine months is very difficult."

He finds the rhetoric bizarre. You avoid the R&D, not the inference. "Kimi is very expensive to serve. The cost per answer is significantly higher." Which leads somewhere useful: if you can point AI at optimising your own serving stack, cost to serve becomes a durable structural advantage, favouring whoever is furthest ahead.

He raises this against his own framework. Aggregation rested on zero marginal cost and on controlling demand through discovery rather than distribution: "your problem isn't that you have distribution. Your problem is you don't have demand."

Test-time compute breaks the first leg. The free-tier user asking for a recipe costs roughly what serving a web page costs. The user running a multi-day reasoning job is "not even remotely in the same universe." One product, two economies, and pricing has to straddle them. Microsoft's metering problem is this same crack in enterprise form.

A useful corrective to the pessimism. Differentiated products have structurally smaller addressable markets because willingness to pay varies: "Apple's never gonna serve the whole world by selling a device, whereas a Google can because it's free."

So: "The Internet is a commodity. It changed the world. So I don't think it'd be weird that intelligence ends up a commodity and changes the world." Intelligence commoditising is not a failure case for the technology, only for particular margin structures.

He is not convinced excellence in verifiable domains transfers to unverifiable ones, and was irritated by labs answering the objection with chess and Go: "I thought we could solve chess. I thought we could solve Go because they're knowable domains. Scale was the answer to both of those, but also both of those were bounded."

His speculation is that the missing input is traces of human thought rather than its end products. Training data holds the finished Reddit comment, not the thinking that produced it, and he wonders aloud whether that is what a Neuralink-style interface would actually be worth.

Then he defuses his own concern commercially. Even with zero further model improvement, the amount of economic activity living in verifiable, procedural work is enormous: "there's a lot of people in the world who are kind of like sentient AIs to a certain extent. They operate very well in verifiable domains. They're given jobs. They do them."

He wants a bubble that leaves something behind. Dark fibre outlived the dotcom bust, Google was built partly on buying it cheap, and the core internet still runs on WorldCom fibre. Railroad money is literally what flows into Google today through Berkshire.

GPUs depreciate too fast and data centre shells are only fine. "What is it gonna be? Power. It has to be power." A world that overbuilt power and then blew up is, to him, an excellent world to land in, because energy scarcity has constrained everything always. He is impressed by how fast the US has added supply, from behind-the-meter gas to restarting nuclear plants.

His answer to the opening question. If the US reached decisive AI-driven military superiority, he thinks China's game-theoretic response is to destroy TSMC, and "to me, this one actually isn't that complicated."

Underneath it is a decoupling argument. US dependence on China is underappreciated and not fixable outside a conflict, because sourcing domestically while a competitor sources from China is a self-inflicted disadvantage. "You do it when you have literally no choice." Even Apple, browbeaten into Indian manufacturing, is diversifying rather than leaving.

He is not soft on competing: "I do think we need to beat China. We need to be competitive." What he objects to is the response. "I despair at the extent to which... so many of our responses, particularly from a political perspective, has been to try to be like China. I think we should be going the other direction. More openness."

Net: "AI right now is kinda like the Taiwan situation. In that, the current status quo actually doesn't seem so bad." And: "So right now, I kinda like where we are."

Railroads, container shipping, memory cycles, Iran and the Strait of Hormuz, IBM and Gerstner, See's and BNSF, Dropbox, Intel and AMD. Each is illuminating and none is evidence. They are also selected after the fact to fit a conclusion, and he never presents the cases where the same analogy would have misled.

He does not size the funding gap, name a date, or specify what running out of money would look like observably. That makes the episode's headline claim unfalsifiable as stated. It is also internally awkward: the Nvidia vehicle is treated as proof the ladder is ending, when it is equally readable as proof that a new rung was found. Private credit and sovereign capital pools that did not exist in the 1870s go unmentioned.

If compute commoditises to marginal cost, as he argues at length, then Google's AI capex does not obviously earn a BNSF-like return. BNSF is a good asset because it earns a decent return on sunk capital across a near-monopoly right of way, not because railroads move a lot of freight. He argues the AI market is enormous without ever asking what return on incremental capital it generates, which is exactly the question his own commodity analysis should force. A Google AI segment subject to his shipping dynamics looks less like BNSF and more like a container line booking depreciation-driven losses that eventually turn real.

A container slot is a container slot; a DRAM chip meets a spec. It is not obvious that GPU-hours are fungible the same way once chip generation, interconnect, software stack and model-specific optimisation are involved. He half concedes this by separating Nvidia (differentiated) from TPU and Trainium (commodity), which is hard to square with data centres behaving like shipping. If capability tiers persist you get durable price dispersion rather than a single clearing price, and the clean collapse never arrives.

It is "reckless" not to be on the frontier if you are a digital company, but Microsoft is rational to skip it and Apple is right to sit it out. Those can be reconciled with enough care about timing and business mix. He does not do the reconciling.

Aggregation theory and commodity analysis are both supply-and-market-structure lenses, so they treat demand as exogenous. Three things go unpriced across the whole conversation: antitrust and regulation as an independent variable; model safety failure as a business risk, which he calls "very real" and then never lets touch a single company call; and labour-market backlash as something that could move the demand curve rather than just the cost curve. If displacement produces political intervention aimed at the ad-targeting machinery he calls a societal positive, the framework has no native way to account for it.