Chips, data centers, neoclouds, compute markets.
Two full reports on CoreWeave: Atlas examines the financials and valuation scenarios. AIR investigates the business, its financing and the debates behind the stock.
Three neoclouds buy the same Nvidia silicon and sell it to the same handful of buyers. What durably separates them is the price of the money used to buy the fleet: 9% at CoreWeave, under 2% at IREN and Nebius.
AI-era rack power density — roughly 135 kW for an NVIDIA GB300 NVL72 cabinet against a typical rack near 10 kW, with about 90% of the heat rejected to liquid — is the mechanism…
The durable profit has left the GPU die for the rack-scale fabric that turns 72 chips into one machine — the thing NVIDIA's rivals can't yet replicate. The memory and networking rents stacked around it are real, and their owners have already published the date they end.
An AI data center can only be built as fast as its slowest input, and in 2026 two inputs are slow at once: memory chips, whose makers cut capacity in the 2023 slump and need two…