Two tracks share one codebase and one capability model. Dates assume the MVP
clock started mid-July 2026.
| When | Milestone | Notes |
|---|---|---|
| 2026-08-14 | MVP fully deployed on AWS (TKF-50) + customer install guide | Demoable on our account or installable by a partner; real vLLM wired in after the runtime team's manual deploy lands (~Jul 27) |
| mid-August | Design partner sessions start (Ronen) | Existing Mirantis customers, guided walkthroughs; feedback must be actionable fast |
| October 2026 | First release "worth its weight" — line in the sand: GTC Berlin | Adds model management (TKF-53): user management, routing, model set, inference deployment |
| March 2027 | NeoCloud (IREN) production | Thousands of GPUs; separate 20-engineer track, hiring from August |
Target: demo-ready ~mid-August 2026 for Accenture/Dell conversations.
| Week | Deliverable | Exit criterion |
|---|---|---|
| 1 | Umbrella Helm chart skeleton; vLLM serving one model (Qwen2.5/Mistral 7–8B class) on the reference GPU box | curl to vLLM's OpenAI endpoint returns tokens; helm install stands the engine up from zero |
| 2 | LiteLLM wired in front with Postgres; model registered; virtual key issuance working | End-to-end curl through the gateway with a governed key; direct engine access disabled |
| 3 | LibreChat against the gateway; admin flow: create user → issue key → send chat link | A non-engineer can log in, chat, and copy a working key |
| 4 | Portal + Control Plane in the chart (self-contained profile, controlplane#7) with Keycloak login, values-single-node.yaml, demo script, install docs |
Fresh single-server install to working demo in under 30 minutes, by someone who didn't build it |
Cut line if week 4 gets tight: demo against the hosted portal pointed at the
demo gateway instead of the self-contained Portal + Control Plane install
(ADR-005 fallback) → never cut the gateway.
Phased by business value; see the NeoCloud edition
for the full rationale.
| Phase | Scope | Rough horizon |
|---|---|---|
| P1 — Local serving | vLLM fleets behind the existing amazee.ai regional gateways; runtime + GPU telemetry (vLLM /metrics, DCGM) |
Q3 2026 |
| P2 — Catalog & registry | Curated open-weight catalog; artifact pipeline (signing/scanning); on-cluster weight caching | Q3–Q4 2026 |
| P3 — Gateway & sovereignty | Residency enforcement GA, latency/capacity-aware routing, request-level failover, dedicated-region hardening | Q4 2026 |
| P4 — Scale-out runtime | llm-d vs Dynamo decision → disaggregated serving, follow-the-sun warm pools | Q4 2026 – Q1 2027 |
| P5 — Federation | Cross-NeoCloud surge-to-partner | 2027 |
See ADRs. Open decisions worth
watching: OCI vs purpose-built model store,
llm-d vs Dynamo, Envoy AI Gateway vs LiteLLM-native routing for the multi-region gateway,
LibreChat fit for the integrated chat surface
(ADR-007).