Matthew Paul (@MatthewJPaul_) header
Matthew Paul (@MatthewJPaul_) profile picture

Portfolio Manager at Twill Capital. Building https://nitter.cf/t.co/jjVTxRvJmR. Writing Volatile Thoughts. Views my own. ↓

Joined November 2021
Photo by Matthew Paul (@MatthewJPaul_)
Google-backed data centers borrow at 7.1%. Neoclouds built on Nvidia hardware pay 9.3%. 220 basis points. Call it the compute spread. At tens of billions per gigawatt, that's the kind of edge that used to require a process-node lead. The chip war has become a capital war. What Google built, Nvidia's $500B response, and the one thing that closes the spread: volatilethoughts.substack.co…
680
AI disruption is not happening to sectors. It is happening to functions. Corporate finance, FP&A, accounting, internal legal, and the HR that supports them are the 1950s typing pool. The mechanical middle compresses. Slowly. But the fuse is lit. Full piece ↓
1
488
Disney reportedly spent $705M on Andor — above any Star Wars film’s production budget. Best reviews of the Disney era. No. 1 on Nielsen finale week. Still doesn’t pencil. The error wasn’t lore. It was not using vertical video as the modern comic book. Full case below.
1
1
844
Equity is the most expensive capital a great business can spend. The decision to spend it is information. The market's instinct will be to treat equity-funded AI capex as one category: either conviction or desperation. That instinct is wrong. The question is not whether a company is selling stock to buy compute. It is what the marginal dollar of compute earns once it arrives. For Alphabet, the same dollar of infrastructure feeds several businesses at once: Search, Cloud, Gemini, and TPU capacity. For Meta, the calculus is different. Compute may be essential to defend the ad business, but the asset being financed still feeds essentially one income statement. Compute that defends an existing moat may be necessary. Necessary is not the same as accretive. I wrote about Alphabet's raise, Meta's buildout, and how investors should grade the next AI equity raise that crosses the tape. Link in comments:
1
270
The market still talks about AI infrastructure as a chip story. I think that is becoming incomplete. GPUs are scarce, but powered land may be scarcer. Chips can be manufactured, shipped, and repriced. Energized sites require interconnection, transmission, permits, utility cooperation, cooling, fiber, and local politics. Capital can buy more GPUs. It cannot make a grid study finish faster. That distinction matters for investors. The next phase of AI infrastructure may be less about who can buy chips, and more about who can actually energize them. I wrote about powered land, Bitcoin miners, data centers, and the infrastructure options embedded in real capacity.
1
222