Hidden Gold: Networking Stack Value

OpenAI’s $122 billion raise and 10GW Project Stargate buildout establish a new infrastructure standard: AI-native facilities built around liquid cooling, high-voltage power architecture, and 800G+ networking now define the performance and valuation baseline, while hardware refresh cycles compress toward 18 months and force faster repricing of GPUs, racks, optics, and supporting gear across the secondary…

In GPU decommissioning, the primary security risk is no longer limited to SSDs. In H100 and B200-class nodes, sensitive data can persist in High Bandwidth Memory (HBM) and adjacent GPU memory structures after standard software wipes or incomplete reset procedures, while VBIOS and BMC firmware layers introduce an additional persistence and supply-chain integrity risk if…

For data center operators, remarketing mid-life GPUs such as the NVIDIA A100 is increasingly a fiduciary duty rather than an optional recovery exercise: inference demand has pushed secondary-market resale values up by roughly 300%, and treating still-deployable accelerators as zero-value ITAD or recycling output destroys capital that can instead fund next-generation H100 and B200 deployments….

If you’re modeling H100s like traditional server hardware, you’re probably overestimating residual value. I call it the “Digital Lettuce” problem: the asset looks great on Day 1 — then it wilts fast. Why H100 depreciation is accelerating New GPU generations and platform refreshes compress resale windows Demand shifts quickly — workloads, power/cooling constraints, export rules,…

CoreWeave’s $8.5 billion DDTL 4.0 shows that GPUs are now being underwritten as investment-grade infrastructure assets, with a credible recovery floor that matters to lenders such as Blackstone. The core collateral is no longer abstract AI capacity but the physical stack itself: H100 and B200 clusters, plus the 800G and 1.6T networking, optics, and high-speed…