For enterprise AI procurement, RFPs are standard. The questions for AI vendors converge on specific dimensions; well-designed RFP cuts through marketing to operational reality. Worth getting right.
Scope: hosting + managed inference + platforms + frontier API. Questions: cost at projected volume, residency options, custom fine-tune, SLAs, security / compliance posture, exit / portability. Score weighted by your priorities. Most important questions: data residency, cost predictability at scale, exit path.
Scope
- Hosting: dedicated GPU providers (GigaGPU, OVH, Hetzner, RunPod, etc.)
- Managed inference: Together AI, Fireworks, Replicate, AWS Bedrock, Azure Foundry
- Frontier API: OpenAI, Anthropic, Google
- Platforms: Databricks Mosaic, SageMaker, Vertex AI
Questions
Standard RFP questions for AI vendors:
- Cost at our projected volume (provide concrete numbers)
- Data residency options (UK / EU / specific regions)
- Custom fine-tune support + cost
- SLA targets (uptime, latency, throughput)
- SOC 2 / PCI-DSS / HIPAA / NHS DSPT certifications
- Audit log retention + export
- Exit path (data portability, model export)
- Roadmap visibility (deprecations, new features)
- Reference customers in our industry
- Support tier + response times
Scoring
- Weight criteria by your priorities (typically: cost, residency, exit path top three)
- Score 1-5 per criterion per vendor
- Total weighted score; review with stakeholders
- Adjust based on intangible factors (vendor stability, support quality)
Verdict
RFPs for AI vendors should focus on operational reality: cost predictability, residency, exit path, certifications. Marketing differentiation evaporates under structured questions. Self-hosted dedicated GPU is increasingly the right answer when criteria weight cost + residency + exit-path heavily.
Bottom line
Structure RFP around operational reality. See build vs buy.
