NVIDIA DGX Cloud review

NVIDIA DGX Cloud is identified in the pricing information as cloud GPU compute for organizations, with an instance described as containing eight GPUs. That places it in the infrastructure layer: a buyer would obtain compute for AI workloads rather than, on the facts available, call a model API. The NVIDIA homepage covers AI, accelerated computing, robotics and simulation, but does not explain DGX Cloud or its deployment model.

That distinction matters when mapping it to a stack. The homepage does not establish whether DGX Cloud is a managed cluster, virtual machines, a hosted service with a control plane, or another arrangement. It also does not describe job scheduling, storage attachment or connections to model-serving systems. We would treat it as a cloud GPU capacity purchase, not assume it is a managed ML platform or API. Ask NVIDIA who provisions and operates the compute, what software is included, and which networking and storage options are available. The known instance size is eight GPUs; the GPU model and memory configuration are not stated.

NVIDIA’s homepage does not publish a DGX Cloud API, SDK, supported language, CLI or integration protocol. It lists Apache Spark and RAPIDS among its AI links, but does not describe them as DGX Cloud integrations. The homepage also requires JavaScript to view all its content, so it is not enough to establish the product’s developer workflow.

Before choosing a build path, ask NVIDIA for the current DGX Cloud documentation and confirm how teams provision instances, submit and monitor jobs, manage credentials, mount data, and access logs. Confirm whether those steps use a web console, command-line tooling, an API, or a cloud-provider interface; do not assume a REST API or SDK is available. Ask which languages and frameworks are supported on the instance, and whether NVIDIA or the customer maintains the software environment. Until those interfaces are confirmed, integration effort and compatibility with an existing CI/CD pipeline are unknown.

The pricing information checked on 2026-11-10 gives a low-confidence historical launch price of $36,999 per instance per month. It describes each instance as containing eight GPUs. This is a historical reference, not a confirmed current quote: the current price and annual billing terms are not published. There is no free plan listed, and no free-trial terms are given.

The stated unit is per instance/month, not per request, seat or token. The information does not give usage limits beyond the instance configuration, or rates for extra compute, storage, networking or support. It also does not publish overage charges or add-ons. That means we cannot calculate a production bill from workload volume or estimate the cost of scaling beyond the listed instance. Ask NVIDIA for a current written quote and clarify minimum commitment, billing increments, idle-instance charges, included services and any separate data-transfer or storage fees before budgeting.

The NVIDIA homepage does not state a DGX Cloud software licence or identify source code that customers can run themselves. It also does not explain whether the service is available only through NVIDIA, whether customers can deploy the same environment on their own infrastructure, or what data-export options are available. We cannot classify it as open source or confirm that it is hosted-only from the published details.

Before committing workloads, ask which components are proprietary, what rights apply to any included software, and whether workloads, model artifacts and configuration can be exported in standard formats. Confirm whether you can move jobs to another GPU provider without changing deployment scripts or rebuilding environments. Migration effort will depend on the actual provisioning interface, software stack and data location; those details are not explained on the homepage. Get the exit and data-retention terms in writing.

The known instance size is eight GPUs, but that alone is not enough to plan production capacity. The available pricing information does not publish other usage limits, and the homepage does not explain DGX Cloud’s security controls, deployment options or operational model. Reliability commitments, availability targets, incident response and support coverage also need confirmation before a production dependency is approved.

Ask NVIDIA how instances scale, what happens when capacity is unavailable, and whether workloads can be paused or resumed without losing state. Confirm GPU model, networking, storage performance, data residency, identity integration, encryption, audit logs and compliance coverage against your requirements. Establish who patches the software and handles infrastructure incidents. Finally, get current pricing and any billing rules for idle capacity, data movement and attached storage. Without those answers, teams cannot reliably estimate cost, assess security fit or design a recovery plan.