Skip to main content
Expanse is built to run inside networks with strict rules about what leaves them. Your code and your workload data stay with you. The few calls that do leave are documented field by field, and you can remove all of them by running air-gapped.

​
What stays in your network

  • Source code. Expanse can capture the code a job ran, so it can explain what the job did. It’s encrypted under keys you hold and stored in your own object storage. Expanse holds no key and can’t read it.
  • Workload data. Metrics, logs, job records and everything the collectors observe are stored in databases you own, inside your network.
  • The model. Your organisation’s model is trained on your jobs, inside your network, and serves predictions from there.
Some of Expanse’s analysis can use a language model. The endpoint is yours to choose and can run inside your network. It receives excerpts of code and job commands, and Expanse never sends them to OpenAI’s or Anthropic’s public APIs.

​
What leaves

A deployment that’s connected to Expanse calls out to Expanse’s control plane. The main calls are a daily licence check, an hourly report on the health of the deployment, and an hourly report of aggregate counts over groups of similar jobs, which you can switch off. The rest keep your clusters registered and their configuration current. None of these calls carries source code, anything your workloads wrote to their logs, commands, job names, job IDs or user names. Every call, every field it carries and the setting that controls it are listed in the deployment documentation we share once we’ve spoken. An air-gapped deployment makes no outbound calls at all. The licence is an offline file, and everything runs inside your network.

​
Access

  • No account on your systems. Expanse doesn’t hold credentials to your clusters, databases or storage. Your platform team installs and upgrades everything.
  • Your identity provider. Users sign in through your own single sign-on, over OIDC.
  • Our engineers only see what you allow. By default, Expanse’s engineers see only what the outbound calls carry. If you want them to help with something inside your network, your administrator can grant time-limited access, which expires after an hour and which you can revoke at any time.
  • Signed software. Every Expanse image is signed and ships with a software bill of materials, so you can check exactly what you’re running.
Expanse is SOC 2 Type II compliant, and we can share the report.

​
What the collectors cost a job

We measured the overhead on a real training job (ResNet-50 on one H100, under Kubernetes), running the same job many times, with Expanse on and off in random order. A deep GPU trace runs for 25 to 30 seconds when a rule detects a likely problem, such as a GPU sitting idle while the CPU is busy. The node collector used about 0.05 of a CPU core and around 50 MiB of memory while collecting. Optional profilers, which you can leave off, cost more, and the deployment documentation covers them. The same benchmark kit runs on your own nodes, with our job or yours.

​
Talk to us

Book a call

We’ll take your security team through the full outbound list and the deployment options.