Confidential GPU inference in practice: a Reddit dev compares Chutes, VoltageGPU and GCP attestation
Polstick1971 · reddit · 2026-10-13
A developer wants to run open-weight video models (e.g. Kandinsky 6.0 Pro) on remote NVIDIA GPUs from a MacBook Air M4, paying only per use — but with a cryptographically verifiable privacy boundary, not just HTTPS or vendor promises.
Before sending private inputs, he wants to verify that: the CPU/GPU are genuinely in confidential computing mode, the deployed inference workload is the intended one, the encryption key belongs to the attested workload rather than an intermediary, and the provider cannot inspect plaintext prompts or generated videos.
He compares three options:
- Chutes: RTX PRO 6000 96GB at $1.80/hour with TEE evidence endpoints and E2EE, but unclear whether the standard client independently verifies attestation or custom video workers can be fully attested
- VoltageGPU: RTX PRO 6000 Confidential VM at $3.80/hour with Intel TDX and NVIDIA attestation, awaiting support on measured boot and custom guest images
- Google Cloud: H100 Spot with dstack or Confidential Space — more complete workload attestation chain but higher complexity
He plans a minimal remote worker running the full video inference pipeline, with the Mac handling orchestration, and asks for firsthand experience on cold starts, limitations, and real costs.
More from Infra
- Optical supply chain has never seen this kind of demand, OCP channel meetings confirm — BenBajarin · 2026-10-13
- RedPill unifies TEE-backed private inference providers behind one verifiable OpenAI-compatible gateway — bgmshana · 2026-10-13
- d-Matrix Raptor NVLink Fusion rack: 144 XPUs, 2.3 TB 3D-DRAM, 7.2 PB/s bandwidth — firstadopter · 2026-10-13
- Gelsinger: AI speeds chip design but chips still take quarters; the whole flow needs rethinking — ai · 2026-10-13
- Nvidia reportedly halts GeForce RTX 5090 production to prioritize AI data center GPUs — SumitGup · 2026-10-13
- Jefferies sees a 2008-style credit angle in the AI boom as hyperscaler RPO backlog nears $2.3tn — SumitGup · 2026-10-13