Aravolta

Aravolta — Per-asset GPU fleet telemetry read directly from accelerator hardware while it runs: utilization, die/power temperature, power draw, and inferred workload pattern (training vs. inference) at serial-number granularity, carried continuously from deployment to retirement and joined to each device's maintenance record and a fitted per-asset depreciation/life-expectancy curve. This is hardware-condition ground truth on AI-accelerator collateral — a class of data that exists nowhere else, because hyperscalers publish utilization only as aggregate commentary and OEMs see only RMA rates.. …

Aravolta offers (Alternative, Geospatial, Fundamental, Reference) — est. 10^4-10^6 GPU-hours of per-device telemetry to date — a single H100 emitting utilization/thermal/power at one-second cadence generates roughly 2.6M data points/year, so even a few thousand financed devices yields billions of readings; a seed-stage book of a small number of deals implies a modest fleet so far but dense per-device history. No device or data-point count is disclosed..

Accelerator degradation and remaining-useful-life models (workload-stress-to-failure); GPU collateral risk and salvage-value forecasting for asset-backed lenders, replacing straight-line depreciation assumptions; RLHF-style regression targets for hardware-lifetime models — measured life vs. assumed life pairs; secondhand GPU pricing and resale-market models conditioned on actual burn-in history; datacenter energy and cooling efficiency research using utilization-to-power curves; an alternative indicator of AI-infrastructure utilization and overbuild, observable per device rather than announced as guidance.

The data is with 2 years of history.

Coverage spans US; Semiconductors, Merchant/Capital Markets; alternative, geospatial, fundamental, reference; fixed_income, equities.