A second pair of eyes before the deposit is non-refundable
A rack-scale GPU order commits seven or eight figures on a deposit you cannot cancel. This is an independent review of the fabric, management plane and operations design — delivered in writing, before the purchase order goes out.
EUR 10,000 · two weeks · written recommendation and a working session
Why the review comes first
By the time a cluster is racked, the expensive decisions are already made. The fabric topology, the oversubscription ratio, the management plane, the power and cooling envelope and the multi-tenancy model are all fixed by the bill of materials. Most of them cannot be changed afterwards without buying hardware twice.
What goes wrong
- ❌ Fabric sized for the GPU count, not for the collective traffic
- ❌ Cooling specified for nameplate, not sustained draw
- ❌ No management plane in the BOM, bolted on later
- ❌ Multi-tenancy left to "we will sort it in software"
- ❌ Vendor's reference architecture accepted without challenge
What the review gives you
- ✅ An independent read of the specification, in writing
- ✅ The questions to put to your vendor before signing
- ✅ Risks ranked by what they cost to fix later
- ✅ A defensible record for your board or your investors
- ✅ Someone with no stake in which hardware you buy
What is reviewed
Four areas, each against your own specification and your own workload — not a generic checklist.
Fabric
Topology, rail design and oversubscription against the collectives your workloads actually run. InfiniBand or Ethernet, leaf and spine counts, cable plant, and what happens to NCCL performance at the sizes you plan to train at.
Rack-scale and cooling
Direct-to-chip or sidecar, coolant loop and CDU sizing, sustained versus nameplate power, and the bring-up sequence. Whether the hall you have can carry what the BOM assumes.
Management plane
Scheduler and tenancy model — SLURM, Kubernetes or both — provisioning, image and driver lifecycle, and the health and telemetry stack. DCGM, node exporters, and what you do when a GPU degrades rather than fails.
Operations
Who runs it on day two: quotas and fair share, isolation between tenants, upgrade path, runbooks, and the staffing the design implies. The part that decides whether the cluster is usable or merely installed.
How it runs
- 1
NDA and materials
You send the specification, the bill of materials and the site constraints under NDA. A short call to agree what matters most to you.
- 2
The review
two weeks of work across the four areas above, with questions back to you and to your vendor as they come up rather than at the end.
- 3
Written recommendation
A document you can hand to your vendor, your board or your investors: findings, risks ranked by cost of late correction, and the specific questions to resolve before the order.
- 4
Working session
A call to go through it, so the document is understood rather than filed. Anything that needs a second look after that is included.
EUR 10,000, fixed. Not an hourly engagement and not a retainer. If the scale is materially different from a single-hall deployment, the fee is agreed before anything starts — but it is always a number, agreed up front, not a day rate running against an open scope.
Why an independent reviewer
Your integrator, your hardware vendor and your colocation provider all have a position on what you should buy. This review has none. I do not resell hardware, I take no referral fee from any vendor, and the recommendation is the deliverable.
Hands on, at command level
Multi-tenant GPU platforms on bare-metal OpenShift AI, SLURM multi-node, NCCL and DCGM — written up publicly, not just claimed on a slide.
On the record
Talks on multi-tenant GPU platform engineering at KubeCon Europe and Red Hat Summit, and eight technical books on infrastructure and automation.
Fixed scope by design
A defined piece of work with a defined end. No discovery phase, no scope creep, no retainer you have to cancel later.
Related Articles
GPUs Take Flight: Multi-Tenant Platform Engineering at
I presented safety-first multi-tenant GPU platform engineering with NVIDIA and Red Hat OpenShift AI at Red Hat Summit 2026 Discovery Theater in Atlanta.
AIFix NCCL_TIMEOUT: Multi-GPU and Multi-Node Troubleshooting
Fix NCCL_TIMEOUT errors in multi-GPU training and distributed inference. Covers timeout tuning, NCCL environment variables, InfiniBand/RoCE network.
Platform EngineeringPlatformCon 2026: Multi-Tenant GPUs on OpenShift AI
Lessons orchestrating multi-tenant GPUs on OpenShift AI with NVIDIA KAI: GPU sharing, workload isolation, scheduling efficiency, and cost control.
Ordering soon?
Tell me the scale, the site and the date the deposit is due. If a review does not fit that timeline, I will say so.
Request a review