Consider dedicating a few minutes to the super interesting discussion in the thread below between @CarlotaPrzPerez, @ganeumann, @sameer_singh17, @daveg, and @progress_bureau (and yours truly). If I were a hedge fund manager with capital to deploy, I would pay close attention to one question it raises: which historical precedent does the current AI bubble belong to? Let me attempt a synthesis for you and add some of my own thoughts. The framework obviously comes from Carlota. As she wrote in her landmark "Technological Revolutions and Financial Capital" (2002), every technological revolution goes through four phases. Irruption and frenzy form the installation period, driven by financial speculation. Synergy and maturity form the deployment period, when the technology diffuses and total factor productivity rises. A turning point separates the two. The whole sequence is called a "great surge of development". Now, if you look back at history, each surge tends to produce two bubbles. The first bursts at the turning point, let's call it a “mid-cycle crash” that marks the shift from installation to deployment. The second comes at the end, when asset prices drift away from fundamentals while productivity growth stalls. Let’s call that one a “late-cycle crash.” Let’s examine the third surge, the age of steel, electricity, and heavy engineering, which Carlota dates to Carnegie’s Bessemer steel works in 1875. Installation brought cheap steel, transcontinental railways, the telephone, and the first power stations, mostly led by the US and financed by a wave of foreign lending, mostly from the UK. The mid-cycle crash came with the Baring crisis of 1890, when Argentine bonds collapsed and the Bank of England had to rescue a pillar of the City, followed by the American panic of 1893, when overbuilt railways went bankrupt one after another. Deployment followed: the Belle Époque, the great merger movement that produced US Steel and General Electric, and the electrification of factories and cities. As Sameer points out, the late-cycle crash for the third surge came in 1919–21. The First World War had pushed commodity prices and inflation to extremes. The Federal Reserve raised its discount rate to 7 percent in 1920, and wholesale prices fell by more than a third within a year. By then, a new great surge, initiated with the launch of Ford’s Model T, was already installing itself. (Obviously, the closer we get to the present, the harder it becomes to classify each bubble correctly. That’s where a capital allocator can make expensive mistakes, so keep paying attention if you don’t want to shoot yourself in the foot 😅) Again, the fourth surge, oil, automobiles, and mass production, starts with the Model T in 1908. From the thread below, it is clear that its mid-cycle crash is 1929. During the 1920s, manufacturing productivity rose by 43 percent while wages stagnated, a Great Decoupling in which the technology was ready but consumer purchasing power had yet to be institutionalized. Surplus capital flowed into speculation until the bubble burst. As Bill Janeway once pointed out to me, the Depression really took hold in 1931, with the international banking crisis and gold-standard austerity. Then came the Second World War, a major disruption that belongs in a completely different category from financial crashes. It forced radical institutional innovation. The Wagner Act and Social Security were already in place in the US, but Europe finally followed with empowered labor unions and the welfare state, while Europe and America together put the Bretton Woods system in place. This brand-new socio-institutional framework, effectively brought about in reaction to the horrors of the war and the Holocaust, turned workers into mass consumers, and deployment, which had already begun in the 1930s, at least in the US, restarted on steroids across the West and Japan, ushering in the postwar Golden Age. Then the fourth surge’s plateau finally arrived in the 1970s. US productivity growth fell from 2.8 percent a year before 1973 to 1.2 percent between 1973 and 1979, and markets for cars and appliances approached saturation. The late-cycle crash in the US was the collapse of the Nifty Fifty, blue chips such as Polaroid, Xerox, and Avon that investors had bid up to 40 or 50 times earnings as “one-decision” stocks. The oil shock of 1973 quadrupled crude prices, inflation surged, interest rates followed, and the S&P 500 lost almost half its value between January 1973 and October 1974. Then something interesting happened. The West found two aces up its sleeve. The first was globalization, with production moving offshore and China’s integration into the world economy becoming the defining feature of the era. The second was financialization, of production through shareholder value and leveraged buyouts, and of consumption through what Colin Crouch called “privatized Keynesianism”: household debt and home equity standing in for wage growth to sustain middle-class demand. Together, these two forces, globalization and financialization, bought two and a half more decades of growth for an exhausted paradigm. I suggest we call this period the “Borrowed Decades,” because previous surges have no equivalent: an almost dead techno-economic paradigm kept alive by cheap labor borrowed from abroad and cheap credit borrowed from the future. There was a price to pay, obviously. Part of it was paid in the form of a spectacular decoupling between productivity and wages across the West. Then, eventually, financialization went too far, which gave us the crisis of 2008. Here I agree with Jerry that 2008 had little to do with a technological revolution. Meanwhile, the fifth surge, semiconductors, software, and networks, had started with Intel’s microprocessor in 1971. Then came the Internet, the mid-cycle crash of 2000, with the end of the dot-com era and the collapse of the telecom bubble, and deployment, which, as Jerry says, was already underway in 2003–07 with broadband, search, and, in 2007, the iPhone. I believe 2022 marks the shift from synergy to maturity within the fifth great surge. Software multiples crashed, and ChatGPT opened a phase in which incumbents (which Jerry calls "synergy companies") moved first with massive capex—a clear sign that we had reached the late cycle. At this point, it’s important to reiterate that, viewed from this perspective, AI is the closing chapter of the fifth surge: it's nothing more than semiconductors, software, and networks on steroids—much like in the 1970s we had oil, automobiles, and mass production on steroids with muscle cars, Batillus-class supertankers, the Concorde, and so on. The problem, characteristic of the late cycle, is that such overshoots funded by abundant capital strain physical resources, which today means power, copper, and memory, with memory prices up six- to eightfold according to Sameer. This clearly suggests we are indeed in the buildout to a late-cycle crash resembling the 1970s and 1919–21, what Sameer calls “the wrong kind of bubble.” For an investor, these are very important signs. A mid-cycle crash tends to be followed by a golden age, so buying the lows of 1932 or 2002 paid off handsomely. By contrast, a late-cycle crash tends to be followed by a long grind. During the fourth surge, the Dow first reached 1,000 in 1966 and only broke clearly above that level in 1982, with inflation eroding real returns throughout. Assets built for the old demand get stranded, like the supertankers ordered before the oil shocks and scrapped within a decade of delivery. David and ProgressBureau are right on an important point: in our AI era, infrastructure is being built before the consumer use cases for agents exist, and if history is to be trusted, use cases do tend to arrive. The difference between the mid-cycle and the late-cycle is *who gets paid for what the technology delivers*: as Jerry explained in his landmark 2025 article, “AI will not make you rich,” published by @colossusmag (cc @zebriez @patrick_oshag), late-cycle buildouts have tended to reward users more than the people who financed them. This is probably what’s about to happen with AI, as the real, widespread use cases will rely on cheap open-weight models run locally on users’ devices, leaving all those data centers up for being scrapped like the old supertankers. There’s one more possibility to consider. Just as the fourth surge got its Borrowed Decades, the fifth may find its own, with a different mix of globalization and financialization. Carlota sees a possible way out through a new kind of globalization whereby the Global South buys capital goods from the West and develops its economies by exporting consumer goods. And on the financial side, the instruments of a new wave of financialization are already emerging, potentially giving birth to a world where money is programmable. Cc @mariekeflament In the 1970s and 1980s, the reset that opened the Borrowed Decades came through the Nixon shock, petrodollar recycling, and London’s Big Bang. This time, it may come through what Marieke and I called the Trump Shock, computedollar recycling, and a financial big bang yet to be determined. If so, we could see a new version of the Borrowed Decades. But this time, the sources of what is being borrowed could be different. Instead of cheap labor borrowed from globalization, we could get cheap labor borrowed from progress in robotics. And instead of cheap credit created through conventional financialization, we could get cheap credit through financialization on steroids, in the form of programmable money. In other words, the fifth surge may not be finished after all. It may find its own way to extend the life of an exhausted paradigm, just as the fourth surge did. The late-cycle crash could then open a second extension, and the investors who make money will be the ones who spot early where those next decades are being borrowed from. Meanwhile, obviously, a sixth surge is probably already underway. I have three hypotheses on my desk. One, initiated in China, is what I call the age of renewables, batteries, and electrification. Another, of which Sameer is a proponent, is a technological revolution in consumer robotics: open-weight models embedded in and running on robots that anyone can use and tinker with. A third hypothesis, which I think is top of mind for Carlota, is closer to a revolution in biotech, biochemicals, and related fields, which, by the way, is also being led by China. In fact, is there any sign that the sixth revolution won’t be led by China? I don’t see any. But that's a discussion for another day.

AI is the culmination of the age of semiconductors, computing and networks, the fifth great surge of development in @CarlotaPrzPerez’s framework. Corollary: it's NOT the opening chapter of a new technological revolution. One obvious reason is that AI is dominated by the companies that emerged as winners during the synergy phase of the current great surge. Consider what this would have looked like during the previous great surge. When Intel invented the microprocessor in 1971, marking the starting point of the fifth great surge, imagine if carmakers Ford and General Motors had caught up within *five months* and become direct competitors to Intel in building microprocessors. Replace Intel with OpenAI, and Ford and GM with Microsoft and Google, and you get a sense of how different AI is from the start of a technological revolution. At the start of a new revolution, the leading companies of the previous great surge would not normally be able to catch up so quickly and compete directly with its pioneers. The fact that Microsoft and Google were able to do exactly that is telling. AI is more of the same, all happening on steroids. cc @ganeumann @sameer_singh17

Here’s the problem with both Anthropic and OpenAI: 1️⃣ I don’t remember where or when, but Mike Green (@profplum99) observed that the impressive, fast and steep adoption of LLMs since 2022 is highly misleading: most of the usage, say 99%, consists of people asking questions they would otherwise ask Google. One can expect this share of the pie ultimately to be claimed back by Google. 2️⃣ Meanwhile, those who use ChatGPT or Claude for more than simply asking questions, say for drafting assistance, brainstorming on complex topics, coding, etc., are growing frustrated with the declining quality. Those are the 1% top users. 3️⃣ The reason quality is going down is most likely that both companies have to throttle inference because they are losing money on each request and facing an increasingly adverse fundraising environment. That’s true even for paid plans. 4️⃣ In other words, those who really need LLMs for the value they can add, the 1%, are having their usage constrained so that OpenAI and Anthropic can absorb the load of the 99% who pay nothing. It’s a loss on both counts: no money is made from the 99% (good luck catching up with Google on advertising revenue), while the 1% are growing disgruntled because of slop, itself the result of inference throttling. 5️⃣ The problem both companies face is very familiar to anyone versed in business strategy or marketing: bad segmentation, which in turn leads to bad strategic positioning. 6️⃣ That’s where the situation diverges. Anthropic can still decide to become 'The AI-Assisted Coding Company' and turn into some combination of Stack Overflow, GitHub and AWS. (They could even acquire GitHub from Microsoft.) I’m not sure they’ll find the TAM necessary to deliver ROIC to their shareholders, though, if the only market they serve is coding. 7️⃣ Meanwhile, OpenAI is in a bind. I don’t see how they catch up with Google in a two-sided, advertising-revenue business aimed at consumers, and I don’t see how they’ll manage to segment their offering and reach a point where there is sufficient willingness to pay across all segments. This is why there are now talks of OpenAI going down, and in any case, I'd say it is all but certain they’ll never generate an ROIC. Cc @edzitron 8️⃣ Last year, I wrote a long piece explaining how OpenAI could reposition itself to specialize in trading for compute: a Glencore for the AI age. That’s a very large TAM, but I’m not seeing anything that suggests they’re going in that direction. See: driftsignal.com/p/what-harry… 9️⃣ All of the above has been made clear by what @sameer_singh17 calls the "open model summer" 👇 Between Alex Karp’s remarks on CNBC, the emergence of cheap and readily available open-source models, and the gradual shift towards running local models, a battle that will likely be won by Apple, what had until now been discussed by only a few people (Sameer, @ganeumann, and yours truly) has now become obvious to everyone. I wrote about Karp's remarks here: driftsignal.com/p/the-beginn… 🔟 How do you call a bubble out of which the winners are two pre-existing “synergy companies”, Google and Apple (to borrow Jerry's exact words, in reference to @CarlotaPrzPerez's "synergy phase" of a great surge of development) and perhaps only a scaled-back Anthropic refocused on code? A late-cycle bubble. That’s what AI is. So stop talking about a new technological revolution and see AI for what it is: an incredible, value-adding final iteration of the age of semiconductors, computing and networks, which won’t live up to the hype from an frontier lab ROIC perspective but will still deliver tremendous productivity gains over the next two decades.

📈📉 Everyone in tech watches demand, but the number that ends up deciding returns is supply. That's the lesson of capital cycle theory, from Marathon Asset Management and @chancellor_e's "Capital Returns". Here's how it plays out: High returns attract capital; that capital builds overcapacity; the overcapacity eventually drags the returns back down. It plays out on a lag, so a sector tends to look strongest just before it turns. Now, @chamath 👇 says hyperscaler free cash flow is holding up and capex is simply exploding, and compares the build to Amazon investing massively in fulfilment and AWS. He's right about the accounting, but (respectfully) completely wrong on the timing. Amazon built that moat RIGHT AFTER the dotcom crash, when capital had fled and few rivals had the balance sheet or nerve to keep spending. It deployed counter-cyclically into an empty field, and its returns compounded as the competition thinned. Amazon also had surprise on its side. Cloud looked like a sideshow, so incumbents ignored it long enough for Amazon to entrench. AI has no such cover: the internet it runs on is already everywhere, so the whole field moves at once. Without that surprise, investment does not build a moat, as @ganeumann pointed out in @colossusmag last year. In fact, the hyperscalers are doing close to the opposite. They are all spending at once, at peak returns, into the same chips, data centres and power. Combined 2026 capex is tracking towards $725B, up about 77% on last year, and close to 94% of operating cash flow on @BankofAmerica's estimate, against a ten-year average nearer 40%. @EpochAIResearch has aggregate free cash flow reaching zero around the third quarter. Every strong incumbent deploying into the same assets at the cycle's top is the overinvestment signal Marathon warns about. From the outside it resembles Amazon's investment, but the difference is real: Amazon built into an empty field, while the hyperscalers are building into a crowded one. Amazon built during the synergy phase and grew into what @ganeumann calls a "synergy company". The hyperscalers are spending into the maturity phase, when the same reinvestment buys far less. The same logic also identifies where the next returns land: with whoever is capital-starved, which today means the owners of the scarce inputs. Three firms control about 95% of memory, grid power is tight and advanced foundry scarce, so they can meter supply and hold their margins much as OPEC once did. The hyperscalers occupy the losing position, pouring money into the layer below them even as the model layer above turns into a commodity, with open-weight rivals already an order of magnitude cheaper per token. None of this says when the market will turn. In fact, these downturns play out over years, and one disciplined giant may yet consolidate the wreckage. On aggregate spend, though, this looks like the top of a cycle. The moat this time accrues to whoever is capital-starved: the input owners, and the firms downstream turning cheap inference into revenue. The giants spending hardest at the peak are financing it for them.