The AI trade, as I see it this weekend (1/n) TLDR: Demand for compute is fine, rents are rising and credit is underwriting structure instead of the label now, the physical buildout does not depend on what the labs IPO at. Two weeks of persistent selling and the loudest AI bear case since spring. My view has not changed at all. What the market is repricing is who pays for compute and at what cost of capital, not whether we need it. We do need more compute. As I’ve been pointing out a lot recently, the Silicon Data H100 Rental Index went from $2.00 per GPU-hour last December to $2.88 on Friday. That chip is three years old and two generations behind. NBIS raised on-demand H100 pricing about 17% at the start of the month, and Runpod went from $3.29 to $3.99 between early September and mid-October. Over the same stretch, token prices kept falling. When the price of the output falls and the rent on the machine rises, usage is growing faster than supply. The rental data has only gotten stronger. On earnings, my view is the same. UBS expects Q3 S&P 500 EPS growth of 28% year over year, with semis contributing about half. Meanwhile hedge funds reportedly just had their worst day since COVID, and the tape has been constant deleveraging and rotation. To me this looks like crowded positioning coming out. I think it’s reasonable to assume that a lot of funds, alongside retail, flipped a switch somewhere around March and April and went heavily risk-on semis. That isn’t a problem in itself, but that positioning comes out of the market at some point, and FUD gives the selling a reason. JPMorgan Private Bank’s note this week lays out how credit is tiering AI borrowers. The strongest AI balance sheets carry 5y CDS around 50 to 120bps. Established infrastructure borrowers are near 300. The weakest financing-dependent borrower is at 1,100. That matches CRWV, whose 5y CDS touched a record 1,194.5bps intraday on October 15, implying roughly a 64% 5y default probability, up from 452bps in June. UBS told clients the same week to favor well-funded hyperscalers and quality semis and to limit neocloud exposure. There’s a key distinction here, though. The CDS, and the 9.75% notes that traded down to a mid-teens yield, are claims on the corporate entity. That is very different from financing secured by the compute and an investment-grade customer’s contract, which is how CRWV’s March facility got rated investment grade in the first place. Credit is pricing unsecured leverage at the holding company. It is not saying GPUs are bad collateral. Credit has shifted to underwriting structure, and structure decides who keeps building. At GTC Berlin, Jensen priced the unit simply: a 1GW AI factory costs about $50B and generates about $40B a year, so one quarter of delay costs about $10B. He also asked lenders to treat Nvidia compute as a financeable asset class and said GPUs now have a useful life of eight years or more. Some will point to Nvidia's other claim from the same week, up to 30x the throughput per MW of GB300 NVL72 on an agentic workload. Read the curve, though. The advantage is about 2x at 110 tokens per second per user, 10x at 130, and 30x at 160. The big gap only exists at the very top of the interactivity range. Below that, older silicon is still productive, and that’s where it ends up. Prefill and premium latency tiers go to the newest racks. Decode, batch and cheaper tiers move to hardware that’s already paid for. Old GPUs don’t just get retired. They move down the stack until the next generation of “old” silicon replaces them. This will happen eventually but that does not break the case that GPUs last a lot longer than the market initially assumed. A lot of the bear case right now leans on the labs as well. Whether OpenAI’s ARR is $70B or approaching $50B, what Anthropic or OpenAI IPO at, or whether they IPO at all. Those questions matter for how lab equity gets priced, but they don’t determine whether the buildout continues. The buildout is driven by usage, and usage is showing up in GPU rents. The technological shift we’re in the middle of isn’t driven by how a private company annualizes a month of revenue. If a lab IPO prices below its last round, the tokens still get served. Falling token prices aren’t deflationary for compute, either. Efficiency is what makes the higher rent payable. Every gain at the model layer lowers the cost of a token and pulls in more usage. That’s why token prices and GPU rents have been moving in opposite directions (see the chart I posted yesterday). The real advances are happening below the model layer, in physical problems. At the hardware layer, that is the 30x per MW above. At the rack level, power delivery is moving to 800V DC and liquid cooling is becoming standard, because racks at these densities can’t be fed or cooled the old way. Nvidia’s DSX POD pitch cuts onsite deployment from 6 months to 14 days and runs warm-water closed loops so evaporative towers aren’t needed. At the power layer, grid connection queues in Europe’s major markets can run longer than the 6 months that design saves. None of that work stops because a lab’s valuation multiple compresses. If anything, a higher cost of capital rewards whoever turns a MW into tokens most efficiently at a given level of model quality. The real link between the labs and the buildout is narrower than the headlines make it sound. If a major lab can’t raise, its compute commitments get renegotiated, and that would hit backlog. That’s a real risk. A lab pricing its IPO lower, or its ARR coming in "lower than consensus," isn't.