GPU generation, not GPU count, sets what capacity is worth
TL;DR. Several items in the 11-07 US day point the same way. Deutsche Bank estimates SpaceX could exit 2026 with $77B ARR from contracted AI compute across 893K NVIDIA GPUs and 1.6GW, about $11B a year from Google alone, with GB300 capacity estimated at ~$13/hour versus ~$3.50 for H100. Microsoft is switching off H100-era servers it rented from Oracle after declining to renew a 2023 Bing contract, and asked IREN to upgrade its GB300 network (@rohanpaul_ai, @StockSavvyShay). A Dutch bank is trying to offload about $500M of data center loans, some tied to Stargate and CyrusOne (The Information). The market is pricing a GPU-hour by generation: a GB300 hour rents for roughly 3.7x an H100 hour.
No architecture diagram: this is a market note.
Why generation dominates
- Memory per NVLink domain. Large MoE models (mixture-of-experts, where each token visits a few expert sub-networks) serve best when all experts sit in one fast NVLink domain. An 8-GPU H100 node holds 640GB of HBM. A GB300 NVL72 rack holds tens of terabytes in one domain. The MoE routing page showed cross-GPU expert traffic, not FLOPs, sets MoE latency, so older nodes lose value on exactly the models now in demand.
- Renewal risk. Contracts signed on H100s reprice at renewal. IREN re-contracted some H100s at double (11-09 digest), but Microsoft walked away from Oracle's. The spread depends on who the buyer is.
- Financing reacts. Loan sales and a delayed CME GPU rental futures market (pushed 60 days into January, 11-07) both say the market lacks a price curve for aging GPUs.
Relation to prior wiki pages
- Extends compute-economics: the 11-07 note on Microsoft-Oracle framed it as a contract story. The DB estimate adds a price ratio.
- Tension with efficiency research: drafters like DSpine (Tencent, 2.7-3.8x agent throughput on a 27B model) make old GPUs more productive, but the market is discounting them anyway, because memory capacity, not compute, is the binding limit for frontier MoE serving.