Secure cluster design
Isolation designed in from the start, from fabric partitioning to management-network separation, then verified once the cluster is live.
For organisations building shared GPU clustersContinuous tenant-isolation verification for GPU clusters.
We design secure multi-tenant GPU clusters and continuously prove that every tenant stays isolated, from the fabric to the firmware.
Now onboarding design partners.
Why a razorbill
A shared GPU cluster is a crowded colony.
Workloads for banks, governments and AI labs, side by side on the same cliff.
Every tenant should hold its own ledge.
What can one tenant see, reach or change about another?
A canary checks every boundary.
Every time a node is provisioned or recycled, and on a schedule.
The problem
Shared GPU clusters run workloads for banks, governments and AI labs side by side.
Recent independent testing of GPU cloud providers found tenants able to:
Exposure
See other tenants' infrastructure
Boundary crossing
Read across tenant boundaries
Lateral access
Reach management networks
Often through default settings rather than sophisticated attacks.
Configurations drift. Nodes get recycled. A one-off audit is out of date the day after it ends.
Providers need continuous evidence, and their customers increasingly ask for it.
How it works
A canary tenant that runs from an ordinary customer's position.
Runs every time a node is provisioned or recycled, and on a schedule.
It asks one question
What can one tenant see, reach or change about another?
It checks every shared boundary.
07 boundariesFabric
InfiniBand partitions and keys, RoCE and Ethernet segmentation.
Management plane
BMC, IPMI and Redfish reachability.
DPUs and SmartNICs
Operating mode and host access.
Orchestration
Kubernetes and Slurm exposure, network policy, shared control planes.
Storage and monitoring
Per-tenant separation in the backend, not just the dashboard.
Node reuse
Disk wipe, DPU reflash and firmware state between tenants.
Software
Known-vulnerable drivers, container toolkits, runtimes and firmware.
→ when
Every time a node is provisioned or recycled, and on a schedule.
On failure
Alert with the exact fix
Failures raise an alert with the exact fix.
Then
Re-test
Every fix is followed by a re-test.
Every handoff
Isolation report
Each tenant handoff produces an isolation report the provider can share with that customer.
What we offer
01 / Design
02 / Audit
03 / Verify
From first design to every node handoff.
Baseline isolation audit
A full, scoped assessment of an existing cluster, with findings, fixes and a re-test.
Continuous verification
Ongoing checks on every node handoff, with per-tenant reports for your customers.
Vendor-neutral
NVIDIA&AMD
InfiniBand&Ethernet
Kubernetes&Slurm
How we work
Testing on your terms.
-
No test runs without a signed, written scope.
-
Your engineers stay in the loop.
-
Findings go only to you. Nothing is ever published without your written consent.
-
Checks run inside your environment, so your data stays in-country.
About
Saad Khan
FounderAlcora Labs was founded by Saad Khan. He has five years of experience in AWS and cloud infrastructure, a background in electrical engineering, and comes from Instec, a cybersecurity firm that has operated across Pakistan and the Gulf since the 1980s.
Named after the razorbill: a seabird that nests in crowded colonies, each pair holding its own ledge.
Prove every tenant stays isolated.
Now onboarding design partners.