Child of #695 (plan Phase 2b, amended by the RULING 2026-09-14: hyperscalers only).
The decisive risk of the whole plan: every hyperscaler Windows GPU exposes GRID/vGPU drivers, and Sunshine has the #1848 D3D11 failure class on GRID plus wrong-display capture with multiple VDD monitors. There is no consumer-driver fallback anymore. If the capture chain fails on GRID and no IDD workaround exists, #695 stops. Therefore run this gate FIRST, before the reflector spend (#705) and campaign (#706).
Rent hourly: AWS g5.xlarge Windows (about $1.2 to 1.3/h) preferred on price, Azure NVads A10 v5 (full GPU size) as backup. A few hours per attempt, budget under EUR 100 total.
Validate: VDD (MttVDD) installs and enumerates on Windows Server, Sunshine creates its D3D11 device on the GRID driver, capture pins to the correct VDD output with 2 monitors present, NVENC sessions open, a Moonlight client pairs and streams. Also install hydraprobe server and WireGuard, peer at the hub, and use the box for first tunnel path sanity checks. Record every deviation in the variant runbook draft.
Gate G2b: capture chain works on Windows Server + GRID for 24 hours. NO-GO: evaluate an alternative IDD build as a documented variant with its own sub-issue; if none works, report and stop #695 per the ruling.