By industry

How we earn it

Library

DatasheetModelyo architecture at a glanceTalk to an expert

Confidential GPU

AI on sensitive data. Protected while it runs.

Run sensitive AI workloads on H100, H200, and B300 GPUs, with protection for data and models in use and a verifiable runtime.

Explore GPUsHow protection worksHow it works

Included in the service

  • Confidential VM · Intel TDX
  • Protected GPU memory
  • Dual attestation
  • Infrastructure operations
  • Confidential VM · Intel TDX
  • Protected GPU memory
  • Dual attestation
  • Infrastructure operations

Room to grow. Protection built in.

Choose the GPU and usage plan that fit your workload. Pricing includes an Intel TDX environment and infrastructure ready for GPU confidential computing.

01 / StartTrial ready

NVIDIA H100

A starting point for confidential inference and model fine-tuning. 80 GB HBM3, 3.35 TB/s.

HBM3 memory
80 GB
Memory bandwidth
3.35 TB/s
Region
US-West
Scale
1–2 GPUs

$3.08per GPU-hour, on demand

Talk to us
02 / ExpandReservation ready

NVIDIA H200

More memory for larger private model workloads. 141 GB HBM3e, 4.8 TB/s.

HBM3e memory
141 GB
Memory bandwidth
4.8 TB/s
Region
US-West / India
Scale
1–8 GPUs

$4.80per GPU-hour, on demand

Talk to us
03 / ScaleQuote available

NVIDIA B300

Blackwell Ultra for advanced inference and dedicated clusters. 288 GB HBM3e, 8 TB/s.

HBM3e memory
288 GB
Memory bandwidth
8 TB/s
Region
US-East / US-West
Scale
1–8 GPUs

$6.50per GPU-hour, on demand

Talk to us

Start with your workload.

More memory supports larger models and caches. Memory bandwidth affects how quickly data moves during computation.

1x

NVIDIA H100

A starting point for confidential inference and model fine-tuning.

80 GB HBM3

1.9x

NVIDIA H200

More memory for larger private model workloads.

141 GB HBM3e

3.2x

NVIDIA B300

For advanced inference and dedicated clusters.

288 GB HBM3e

CVM runtime

Trust starts at runtime.

Each component helps protect your workload, from the confidential VM to the GPU memory and the signed evidence that proves it.

Talk to an expert

  • Docker workloads run inside an Intel TDX confidential VM with GPU passthrough. Firmware measurements allow the runtime to be checked before the workload starts.

    Intel TDXDockerGPU passthrough

  • NVIDIA Confidential Computing protects model weights, activations, and KV cache in GPU memory, alongside the protected execution environment on the CPU.

    Model weightsActivationsKV cache

  • Signed evidence from Intel TDX and NVIDIA enables combined verification of the VM and GPU configuration, providing visibility into the execution environment.

    Signed evidenceShared verifier

  • The confidential environment can support AI inference, agents, and model training, with sensitive data processed inside the protected runtime.

    Private inferenceAgentsTraining

GPU confidential computing

Protection where your model computes

NVIDIA Confidential Computing protects GPU memory alongside the protected execution environment on the CPU.

Talk to an expert

Dual attestation

Two sources of evidence. One verification process

Intel TDX and NVIDIA each produce signed attestation evidence. Combined verification checks both the confidential VM and GPU configuration.

Talk to an expert

Run your first confidential workload.

Start with an H100 trial, check the attestation report yourself, then grow into reserved capacity or a dedicated cluster.

Talk to an expert