Chips
One page per chip, four lanes per page: spot, committed, interruptible, and secondary hardware. Every figure below is the same cell the instrument page — and the surface behind it — prints.
Chip instruments · T2 Neocloud headline band
| Silicon | Identity | Spot · GTD USD/GPU-hr | Committed USD/GPU-hr | Interruptible USD/GPU-hr | Model-implied value USD/unit · model output |
|---|---|---|---|---|---|
| NVIDIA H100 → | SXM 80GB · Hopper · 2022 | $3.72 | $2.25 | $2.15 | $25,575 |
| NVIDIA H200 → | SXM 141GB · Hopper · 2024 | $4.47 | $3.65 | $2.45 | $40,893 |
| NVIDIA B200 → | SXM 180GB · Blackwell · 2025 | $6.69 | $6.42 | $3.95 | $86,023 |
| NVIDIA A100 → | SXM 80GB · Ampere · 2020 | $2.03 | $1.95 | $1.14 | $7,948 |
Construction
Spot / interruptible = the /rates T2 ladder cells (on-demand asks,
form and region pooled). Committed = the /term construction: the
longest published tenor implied off the on-demand anchor by the
within-provider spread — market intelligence, not a citable
reference rate. Model-implied value = /hardware's income-derived
model output for the one locked variant per chip (sensitivity band
beside each figure) — a model output, not a transacted price; the
posted and executed secondary-market record stays on /hardware and
each chip page. Lanes are never blended; an em dash is a cell
below the publication floor. Panel depth and dispersion live on
the surface pages.
A chip joins this index when it carries a T2 rates cell and a /hardware value lane — posted asks or the published income model. B200's secondary market has no posted asks on the record yet; its hardware figure is modeled only. B300 and GB200 carry neither lane and have no instrument page yet.