AI Hub roadmap dependency matrix

A nonconfidential internal planning aid for the AI Hub feature set. Rows show target outcomes by horizon. Select a row to inspect its owner, exit condition, evidence, functional prerequisites, and release gates. Select a horizon heading to filter the matrix.

Updated September 7, 2026 · The local HTML file is canonical; the unauthenticated Cloudflare deployment is a mirror.

Relationship view Build view: needs and unlocks · gate view: gated by and gates
✓Shipped ●In build ◌Planned ◇Future candidate NEW agreed gap SPLIT separated capability Deployment: managed cloud · BYOC · private data center · internal validation

Deployment parity: AI Hub provides the same product behavior across managed cloud, BYOC, and private data centers. Operating responsibilities, compatible infrastructure, connectivity, and release readiness differ. All means shared applicability, not simultaneous acceptance. Agentic applications orchestrate their tasks and call AI Hub for inference; broader application-workflow orchestration is not a confirmed commitment.

March 31, 2027 evidence gate: managed-cloud customer validation and customer-weight acceptance block both April offers. The BYOC reference blocks BYOC only. Private data center proof of concept does not block either offer.

Feature coverage notes

December 2026 validates one internal Atlas workload. Paid managed-cloud preview opens February 1, 2027. March 31 is the blocking evidence checkpoint; rows explicitly due then sit beside the April milestone, not in February entry scope. Managed cloud and qualified BYOC target April 15 independently. The private-data-center proof of concept is due April 30. Future rows retain their prior directional planning horizons in their details and do not represent a release commitment. Evidence references describe source records; existing implementation claims are not new release acceptance.

Model-class coverage

Model classAPI surfaceAvailabilityLater candidates
Chat or instruct model/v1/chat/completionsDecember: selected Atlas chat behavior, operator catalog, and first external integration. Customer weights by March 31.LoRA, preference tuning, reinforcement learning
Completion model/v1/completionsDecember only if consumed by the selected Atlas workload; otherwise separately qualified with no date commitment.Continued pretraining and full-parameter tuning
Embedding model/v1/embeddingsFebruary 1: qualified operator catalog. Customer weights by March 31.Contrastive tuning
Reranker/v1/rerankFebruary 1: qualified operator catalog. Customer weights by March 31.Pairwise and listwise tuning
Vision-language model/v1/chat/completions with image partsFuture; prior Q1 2027 planning horizon, not a preview-entry commitment.Multimodal tuning