We mapped Hg Capital's entire portfolio into a Knowledge graph and studied their strategy. Hg owns four companies that sell the same product to the same buyer. The research every PE firm cites to justify buy-and-build says it stops working after two acquisitions. One company in this portfolio has made about a hundred. 29 of these 59 companies never raised a penny of outside capital before Hg showed up. See what you can learn about their strategy and let me know! Link to the graph in the comments below.

GPU prices fell 75% this year. GPU prices rose 40% this year. Both are true, and the gap between them is what people get wrong about AI infrastructure. Everyone says that the people building AI compute are the new kingmakers. I spoke with my friend Oscar Gompels, who's handing out the crowns himself and mapped 251 companies doing it. Here's what I learnt: Take the pricing first. On-demand H100 rates fell about 75% from the 2023 peak. One-year reserved contracts rose 40% between October 2025 and March 2026. Spot capacity deflates as each generation matures. Contracted capacity inflates while the shortage holds. Quote one number and you have described half a market. Now the kingmaker question. The companies renting you GPUs carry obsolescence risk, utilisation risk and refinancing risk at the same time. Post-depreciation margins across the category run 14 to 16%. Applied Digital earns over 25% leasing them the building. Broadcom booked $16.7bn of AI semiconductor revenue in one quarter, up 221%. VAST Data sells the storage layer, runs cash positive, and raised at $30bn. Almost everyone supplying the neoclouds has a better business than the neoclouds. @CoreWeave holds the only Platinum reliability rating in the sector, two ratings running, scored on security, health checks, networking and orchestration rather than price. It lists H100 capacity near $6.16 an hour against rivals at $3.85 and sells out anyway, because a failed run on ten thousand GPUs costs more than the rate card difference. It re-leased three year old H100s at 95% of original pricing and has A100s contracted through 2029. That is the strongest published evidence against the idea that a GPU is worthless after three years, and it came from an operator willing to disclose it. Its backlog reached $104bn, and the share running beyond four years doubled in a year. A week of this left me with one rule of thumb. The money flows to whoever holds something you cannot easily rebuild in a hurry: an energised site, a substation, an operating record customers renew against. Anyone can order chips. We've published a public knowledge graph on this topic, check the comments below to explore it yourself!