AI stocks are deep in sell-off: Nvidia down 24%, AMD 38%, Micron 46%, Broadcom nearly 40% from highs.
Anthropic's IPO, one or two weeks away and likely top-three ever, rests on an S-1 showing heavy losses and one profitable quarter.
Philipp Müller rejects Michael Burry's house-of-cards thesis, saying Anthropic's numbers show the chip-to-inference chain is profitable but fragile.
AI stocks slide as Anthropic's giant IPO nears
This is a reading of the AI stock sell-off on the eve of the Anthropic IPO, recorded on a Friday when the numbers could still move. The two hosts read the drawdown as a collision of politics and expectations rather than a verdict on AI economics, since the build-out needs sites and power that the public increasingly resists. The open question is whether the model makers' losses and the neoclouds' debt can be carried by a value chain that does earn money.
Takeaways
The Anthropic IPO can no longer be stopped. Höpfner says the offering is coming and is likely one of the three largest ever, alongside Saudi Aramco , with attention on which way the stock spikes the day after it lists .
The AI sell-off follows a huge run-up. Nvidia is down 24 percent from its highs, AMD 38 percent, Micron 46 percent and Broadcom almost 40 percent . AMD is still 90 percent above its start-of-2026 price and Micron 130 percent, so the losses sit on a huge run .
Anthropic's S-1 shows losses and fast growth. Anthropic is heavily loss-making with one profitable quarter, while OpenAI also loses money and the rest of the value chain profits . Net revenue retention is a factor of 10, with customers going from 20 to 200 , but a further fivefold from a 1.5tn valuation is unlikely .
Burry's collapse thesis is contested. Burry has said for about two years that Nvidia is overvalued, reading the value chain as entangled and balance sheets as prettied up through depreciation . Müller disagrees and points to Anthropic's numbers as evidence that the chain from chip to model to inference is profitable .
Data centres have become a political problem. In a survey Müller found, 38 of 39 respondents said they did not want data centres , and the broad population is not amused . If data centres are seen as the devil, sites and energy get harder to find .
Neocloud debt has no dot-com precedent. Unlike the dot-com era, which was all equity, the neoclouds building capacity are extremely indebted because a GPU can be borrowed against , with CoreWeave guaranteeing with its balance sheet . Höpfner says cheap money keeps flowing into riskier assets .
Anthropic's plan to cut hyperscaler dependence
Anthropic wants to reduce its dependence on hyperscalers from 80 percent to between 20 and 40 percent .
It plans to buy another 20 to 40 percent of capacity from neoclouds and build the remainder itself .
Data centres as real estate
Müller says data centres are real estate with only GPUs inside, and the economics and financing instruments are essentially the same .
Höpfner counters that the GPUs inside can create a value tomorrow that is twice today's, which renting does not deliver .
Tap a name to go there · tap a line to see why they are linked
