VAST Data

Roadmap & Vision

Fundamental

From a data platform to an operating system for AI - what ships now, what's coming, and the closed loop it all builds toward.

What ships when

VAST frames the platform as the “”, unveiled at a New York launch in May 2025 - the VAST Overview explains its layers and what ships today. This page is about time: the platform spans a spectrum from available-now infrastructure to stated vision, laid out on a time axis. Filter by maturity so you can separate what a customer can deploy today from what is still a roadmap commitment.

VAST AI OS - what ships when

Each lane is a layer of the platform, placed on a time axis by its stated date. Filter by maturity, then pick a milestone to see what it does and why it matters. Status labels follow VAST's own framing.

  1. Shipped - available now

    May 2025 → today

  2. In preview

    announced, not yet GA

    today

  3. Targeted end of 2026

    date as stated by VAST

  4. Upcoming - date TBA

    no public date yet

  5. Vision

    open-ended

RoadmapTargeted · end 2026
PolicyEngine

Zero-trust governance for autonomous agents.

A policy layer that governs what agents are allowed to read, write and do - zero-trust controls applied at the data platform so agent autonomy doesn't outrun governance. Announced at VAST Forward in February 2026 and targeted for end of 2026.

available now preview roadmap vision dated announcement

Status as of September 2026. The ~44% Sirius query-time figure comes from early benchmarks. Roadmap items and their timing are subject to change. Available-now items without a stated ship date are drawn ending at the today marker.

The engines of the AI OS

These pieces turn the platform from a place data lives into a place intelligence runs. Each is GPU- or AI-native by design - one is shipping today, one is in technical preview, and one is on the roadmap.

What each engine changes

Pick an engine to see the shift it makes on the platform.

Edge
On-prem
Other clouds
Google CloudNov 2025
One global namespace - the same data addressable everywhere, no copies
traininginferencefollow GPU capacity to whichever site has it

One consistent global namespace across edge, on-prem and every cloud, so the same data is addressable everywhere with no copies. Shipping now and extended to Google Cloud (Nov 2025) - letting training and inference follow GPU capacity.

The ~44% query-time figure comes from early benchmarks. Roadmap dates are subject to change. Status as of September 2026.

The “Thinking Machine”

The roadmap isn't a list of features for their own sake - every piece is a stage in a single closed loop. When observation, reasoning, action, evaluation and improvement all run next to the data, the platform stops being passive storage and starts continuously improving itself. Step through the loop below.

The closed loop: observe → reason → act → evaluate → improve

VAST's “Thinking Machine” vision - a continuous loop that runs next to the data instead of shipping data out to it. Every stage reads and writes the same data at the center.

1/5
the dataone platformObserveEvent Broker + DataEngineReasonVAST DB + VectorStoreActAgentEngine + MCPEvaluatePolicyEngineImproveTuningEngine

Dashed = stages whose component is not yet generally available (AgentEngine in preview; PolicyEngine and TuningEngine targeted end 2026) - today the loop is not yet closed.

Observe

New data lands - files, objects, events, telemetry.

On VAST

Event Broker + DataEngine triggers detect arrivals in real time.

The loop framing is VAST's stated vision. AgentEngine is in preview; PolicyEngine and TuningEngine are roadmap items (targeted end 2026). Status as of September 2026.

Why this matters for AI

The bottleneck in AI infrastructure has moved from raw to keeping GPUs fed with fresh, governed data. Each roadmap item attacks a different part of that problem: Sirius removes the CPU query bottleneck, removes the bottleneck, and PolicyEngine/TuningEngine let the loop run autonomously without losing control - TuningEngine automates the LoRA, SFT and RL methods covered in Training & Fine-Tuning. VAST's stated aim is to make the data platform itself the AI factory.

From ingest to insight: point products vs one platform

Switch between the two layouts and follow a piece of data from the moment it is ingested to the moment an AI team uses it.

4

systems to run

3

extra copies

3

glue hand-offs

Counts are for this sketch, not a measured deployment.

For the data center

less data movementfewer copieslower latency

One platform instead of separate storage, warehouse, streaming and vector systems - less data movement, fewer copies, lower latency from ingest to insight.

For AI teams

Fresh context for RAG + agentsGPU-speed analyticsGoverned fine-tuning on your data

Fresh context for RAG and agents, GPU-speed analytics, and a governed path to fine-tune on your own data - the full observe-to-improve loop without a fleet of glue services.

Key takeaways

In one line

VAST frames itself as an "AI Operating System" (unveiled at a New York launch in May 2025) - unifying storage, database, streaming, compute, and agents in one platform, with some pieces shipping and others on the roadmap.

Key points

  • DataSpace, VAST DB, Event Broker, DataEngine, VectorStore, and Polaris are shipping today; Sirius GPU SQL and AgentEngine are in preview; PolicyEngine and TuningEngine are roadmap items.
  • VAST runs Sirius, an open-source GPU SQL engine built on NVIDIA cuDF, to accelerate VAST DB queries on NVIDIA GPUs (in technical preview since Aug 2026), cutting VAST DB query time up to ~44% in early benchmarks.
  • PolicyEngine (zero-trust agent governance) and TuningEngine (automated LoRA/SFT/RL fine-tuning on platform data) are both targeted for end of 2026.
  • The stated end goal is a closed loop - observe, reason, act, evaluate, improve - running continuously next to the data (the 'Thinking Machine').

Questions to explore

  1. 01How many separate systems (storage, warehouse, streaming, vector store) does your data move between before it reaches a model?
  2. 02Is your SQL analytics pipeline CPU-bound today, and would GPU acceleration change what's practical?
  3. 03Do you have a governance layer that defines what an autonomous agent may read, write, or do before it runs?

Common questions

Is the data substrate actually shipping, or is it all vision?
Storage, VAST DB, Event Broker, DataEngine, VectorStore, and the DataSpace global namespace are described as shipping now; DataSpace extended to Google Cloud in November 2025. Sirius GPU SQL is in technical preview; PolicyEngine and TuningEngine remain roadmap items.
What does Sirius change for SQL workloads?
It runs VAST DB queries on GPUs using NVIDIA cuDF; in early benchmarks, query time dropped up to about 44%.
When do PolicyEngine and TuningEngine ship?
Both are targeted for end of 2026; PolicyEngine governs what agents may read, write, or do, while TuningEngine automates LoRA/SFT/RL fine-tuning from platform data.