GPU clusters for NVIDIA Omniverse digital twins

The company

Sight Machine builds industrial AI for manufacturers. The product this work supported renders a physically accurate 3D digital twin of a customer’s factory and drives it with live production data, so an operator can look at the plant rather than at a dashboard about the plant. The twin is built with OpenUSD and NVIDIA Omniverse, streamed to the browser through Omniverse Kit App Streaming, and rendered on NVIDIA A10 GPUs in Azure. The work reached NVIDIA through a partnership with Microsoft, and I worked directly with engineers at both companies. NVIDIA published the result as a case study.

The problem

Real-time 3D rendering needs GPUs, and every customer wanted the platform inside their own Azure tenancy rather than ours. That made the deliverable a whole environment, one per customer, built in a subscription we were given access to and did not control. Each one had to be stood up correctly and quickly, because the environment is what stands between a signed customer and a running product.

What I did

I owned the infrastructure for those deployments. Each customer got a dedicated AKS cluster and everything around it: the VNet, API Management, digital certificates, the workloads themselves, and a GPU node pool tainted so that only the NVIDIA pods that need a GPU are scheduled onto the hardware being paid for. Region and zone selection came first every time — GPU node sizes are not offered everywhere, and picking the wrong zone is discovered late and paid for in rework. The integration was announced in March 2025 and demonstrated publicly at NVIDIA GTC.

All of it was Terraform, applied through Atlantis so that a deployment was a reviewed pull request rather than someone’s laptop. Observability was part of the build rather than a follow-up: dcgm-exporter wired into the existing Prometheus and Grafana stack, so GPU utilisation was visible from the first day of each environment instead of being inferred from the Azure bill.

Result

The Terraform modules were started by the engineer before me. I inherited them and kept tightening them across deployments, and the time to bring up a new customer environment fell steadily as a result — faster onboarding, less of the per-customer improvisation that makes each environment slightly different from the last. I have no measured figure for the reduction and do not claim one.

The product’s own numbers are NVIDIA’s to report, and it reports them: operators at a US plant seeing up to 10% better line productivity and up to 15% higher profit margins, with downtime issues resolved in minutes rather than hours. Those are outcomes of the platform, not of my Terraform. They belong here only because the platform ran on environments I built, and none of them happen if the GPU clusters underneath are late, wrong or invisible.

Coverage

The platform this infrastructure carried was written up publicly:

• NVIDIA case study — the customer-facing account, including the reported plant results
• NVIDIA blog — OpenUSD and Omniverse digital twins for industrial AI
• Sight Machine announcement — the March 2025 Omniverse integration
• Microsoft Tech Community — the Azure IoT Operations side of the platform