# RL Environment Stack RL training needs thousands of disposable, isolated, stateful computers per run. It also needs trusted data to build tasks from. GPUs are not the constraint. > [!abstract] Thesis > Compute is a commodity. The defensible layers are: > - **Custody** of sensitive data > - **Domain knowledge** of what "correct" means > - **Verifiable lineage** from raw data to reward ![[RL Environment Stack.png|500]] Interactive version: [[RL Environment Stack.html]]. The layer under the pointer lifts the layers above it. ## The stack | Layer | Best practice (Oct 2026) | |---|---| | Sandbox | Firecracker microVMs. Boot any OCI image. Snapshot and **fork** one state into many children. | | Interface | Gymnasium API (`reset`, `step`, `state`). OpenEnv for transport. Verifiers for rubrics. Support both. | | Reward | Hardened, versioned graders. See [[Reward Integrity]]. | | Orchestration | Kubernetes with batch schedulers (Slinky, KAI). Separate CPU sandbox pool from GPU pool. | | Data lake | Immutable raw zone. S3-compatible store. Parquet + Iceberg/Delta. LeRobot and MCAP for robotics. lakeFS-style versions. See [[data lakes]]. | | Provenance | Per-hour consent and lineage record. See [[RL Data Provenance]]. | | Security | Per-tenant isolation, mTLS, Vault/HSM, [[Confidential Computing]], PQC signatures. See [[HSM x PQC]]. | ## Sandbox price anchor - Prime Sandboxes: **~$0.039 per sandbox-hour** (1 vCPU, 1 GiB RAM, 32 GiB disk). - A sovereign offer must beat this on cost, or justify a premium with residency, custody or air-gap. ## Portable bindings - One codebase, three deployments: hosted, customer cloud, air-gapped. - Cloud is a binding, not the architecture. Swap S3 → MinIO/Ceph, auth → Keycloak, secrets → Vault/HSM. - A sovereign host slots into this shape. Borrowed from vfrog. ## Market signals > [!warning] Reported figures. Verify before use. > - **Mecka**: >$100M run-rate with 40–60 staff (Jun 2026). $60M Series B (Oct 2026). > - **Prime Intellect**: $130M Series A at ~$1B (Jul 2026). 2,500+ hub environments. > - **Build AI**: open Egocentric-1M dataset (Apache 2.0). Free data caps the price of commodity hours. > - Counter-view: hours may be the wrong axis. Buyers want transfer evidence, not volume. ## Risks - Labs build environments in-house. - A new model makes a task catalogue obsolete. Tasks stop paying at ~70% pass rate. - Large sandbox vendors set commodity prices. - Reward hacking silently destroys environment value. - Biometric and cross-border data exposure. - Scope sprawl across human data and robot telemetry. ## Open questions - [ ] Real cost per sandbox-hour on owned hardware vs. the Prime anchor - [ ] Do buyers pay for sovereign custody, or only for price and scale? - [ ] Which buyers re-order egocentric data, and why? - [ ] Qatar and Kenya rules on exporting anonymised derivatives Next: [[RL Environment Growth Playbook]] ---- tags: #boat #deeptech links: [[Navon MoC]] [[Navon Sovereign Vaults]] [[AI Agents Stack]] [[Compute to data]] [[Foundational Models MOC]] sources: Prime Intellect (blog, sandboxes), Hugging Face OpenEnv blog, beam.cloud (Verifiers), Northflank, Modal, MarkTechPost (AgentENV), dealroom.co/news/160699, labellerr.com, varunnair191587.substack.com/p/scaling-blind, vfrog.ai. Retrieved 2026-10-14.