A realistic total cost of ownership for replicating an enterprise AI agent platform in your own cloud account: the platform engineering, the GPU inference for a Deep Hat-class security model, and above all the specialist team you would have to hire, retain, and keep on call.
This is not a wrapper around a model API. A Kindo-class platform is a product in its own right, and every piece below is table stakes, not gold plating.
Durable agent orchestration, a model gateway with routing and failover, DLP/PII redaction, policy-based authorization, RAG (vector store, embeddings, indexing), code-execution sandboxing, SSO/SCIM, audit, and metering.
OAuth broker, a custom provider catalog, and a fleet of MCP servers that must track every vendor's API changes, forever.
H100-class GPU serving with HA and burst capacity, model upgrades, eval harnesses, red-teaming, and guardrail engineering. GPUs alone run $300–700k per year in production posture.
Every third-party component is pinned, patched, and rebuilt. CVE and breaking-change churn across the stack never stops, and neither can the team that absorbs it.
ML-infrastructure and platform engineers are bid on by AI labs and AI-native startups. Expect 4–6+ month searches, agency fees of 20–25% of first-year base, and 15–20% annual attrition in exactly the hardest seats.
| Role | FTE | Fully loaded / yr |
|---|---|---|
| Tech lead / engineering manager | 1 | $350k |
| Senior backend / platform engineers | 3 | $285k each |
| ML / inference engineers GPU serving, evals, red-teaming | 1.5 | $380k each |
| Frontend engineers | 1.5 | $255k each |
| DevOps / SRE | 1 | $285k |
| Security engineer DLP, authorization, sandbox review | 1 | $300k |
| Build-phase team | 9 | ~$2.7M / yr |
Year one is pure spend: nothing ships until month 15–18. And it never drops below ~$3M a year after that.
| Year 1 | Year 2 | Year 3 | Total | |
|---|---|---|---|---|
| Payroll (fully loaded) | $4.0M | $2.2M | $2.2M | $8.4M |
| Recruiting + backfill | $0.5M | $0.15M | $0.15M | $0.8M |
| Cloud: GPU inference + platform | $0.4M | $0.6M | $0.6M | $1.6M |
| Compliance + tooling | $0.1M | $0.15M | $0.15M | $0.4M |
| Total | ~$5.0M | ~$3.1M | ~$3.1M | ~$11.2M |
This is a people problem, not a cloud-bill problem. Four of every five dollars go to hiring and retaining a team whose market clears against you. The lean, everything-goes-right floor is ~$7M; the realistic ceiling is ~$14M.
DIY estimates always price the build. They almost never price what comes after.
A sane 24/7 rotation needs five to six engineers who each understand the whole system. You cannot staff half a rotation: ~$1.7–2M/yr just to keep the lights on, before a single new feature.
New model generations land every few months. Someone must eval, red-team, quantize, and migrate, or your security AI is stale within a year.
The platform carries its own SOC 2, pentest, and audit burden. Evidence collection, tooling, and auditor fees land on your team, every year.
DIY costs roughly $5M and 18 months before the first user logs in, then $3M+ every year after, contingent on winning a hiring market you are structurally disadvantaged in, to own an undifferentiated, perpetually-behind copy of something you could deploy this week.