{"resourceId":"aicr-mass-ai-hub-vast-storage-snapshot-only-20261008","versions":[{"version":"external-02258cda502ebaf03655687364b886dd71a782077e7927027195231b3456b8df","resource":{"id":"aicr-mass-ai-hub-vast-storage-snapshot-only-20261008","title":"Massachusetts AI Hub's state-funded AICR cluster runs on ~14.3 PiB of VAST all-flash storage, with snapshot-only protection and no off-site backups","organization":"Massachusetts Green High Performance Computing Center (MGHPCC) with Boston University, Harvard, MIT, Northeastern, UMass and Yale; Commonwealth of Massachusetts (Massachusetts AI Hub)","sector":"State government-funded shared AI compute / higher education research computing","geography":"United States (Holyoke, Massachusetts)","publishedAt":"2026-10-08","publicationDate":"2026-10-08","eventDate":null,"sourceName":"AICR Documentation (docs.aicr.ai)","sourceLabel":"AICR operator system description","sourceUrl":"https://docs.aicr.ai/system-description/","evidenceClass":"public-sector-association","outcomeClass":"mixed","topics":["infrastructure","data-security","governance-procurement","operating-model"],"finding":"AICR's system description (page dated October 8, 2026) lists 31 B200 nodes (248 GPUs), 19 RTX PRO 6000 nodes (152 GPUs) and 32 service nodes, with ~14.3 PiB effective VAST all-flash storage rated at 360 GB/s, encryption at rest and GPU Direct Storage over NFS-RDMA on InfiniBand NDR400. Home and work areas have 7-day snapshots; scratch has none and is purged; the operator warns there are no off-site backups. UMass's AICR page (modified October 20, 2026) marks the system Beta, proposal-only with per-project caps, and bars interactive inference.","sledRelevance":"Direct state and higher-education relevance: a $31M Commonwealth of Massachusetts grant via the Massachusetts AI Hub, matched by six universities, built a shared AI resource whose data layer is VAST. It is the first U.S. state-funded VAST deployment found by this stream, and shows how a public consortium actually operates the storage: protection, quotas and use limits.","evidence":"Operator documentation from the public-private university/state consortium that runs the system (configuration and policy, not an evaluation), corroborated by the UMass Unity AICR page and MGHPCC's December 2025 announcement of the state grant and hardware providers. The 360 GB/s figure is a stated system rating; no measured application throughput, utilization, baseline or user outcomes are published.","architectureImplications":"Single VAST namespace serves two GPU classes via NFS-RDMA and GPU Direct Storage on a cluster InfiniBand fabric separate from the rail-optimized B200 fabric. Protection relies on in-system snapshots; there is no second copy, so a platform-level failure or deletion beyond snapshot retention has no off-site recovery path.","governanceImplications":"Proposal-based, capped allocations and an acceptable-use rule against interactive inference shape demand on the storage layer. Users are told to back up irreplaceable data themselves, shifting data-protection duty to research groups. Allocation processes are still evolving (Beta status).","securityPrivacyImplications":"Encryption at rest is enabled per the operator. Lack of off-site backups and snapshot-free scratch limit resilience against ransomware or administrative error; the source does not describe tenancy, key management or regulated-data support.","caveats":"Documentation can change without notice and is not an independent evaluation. Hardware vendor for compute/service nodes is Dell (covered in the Dell stream); this record concerns only the VAST storage layer. The $31M figure and university match come from MGHPCC's December 2025 post, not the October pages. Interpretation (not source finding): snapshot-only protection is an operator design choice, not a VAST limitation.","streamIds":["vast-data"],"roles":{"sales":"AICR is a state-funded reference: the Massachusetts AI Hub chose VAST as the data layer for a shared multi-university AI cluster. Use it with state AI-hub planners and university consortia to discuss shared GPU resources, but stay bounded: there are no published performance or outcome metrics, the system is Beta, and protection is snapshot-only. Discovery should probe how a buyer will fund backups, allocations and operations, not just the initial grant.","engineering":"Engineers can reuse AICR as a design pattern: about 14.3 PiB effective VAST all-flash, a 360 GB/s rating, NFS-RDMA with GPU Direct Storage on a dedicated storage InfiniBand fabric serving B200 and RTX PRO 6000 pools. The notable gap is protection: snapshots only, none on scratch, no off-site copy. Any comparable design should size a replication or backup target, define snapshot retention per tier and validate rated throughput against real multi-node training I/O.","delivery":"Delivery teams supporting a state or consortium AI resource should note the operational model: proposal intake twice monthly, per-project caps, quotas per tier, purge rules for scratch and monthly maintenance windows. Because users must back up irreplaceable data themselves, onboarding needs clear data-protection guidance and possibly a backup service. Acceptance should cover measured throughput per GPU pool, snapshot restore tests and documented recovery objectives before the system leaves Beta."},"roleDetails":{"sales":{"customerProblem":"States and university consortia funding shared AI compute need a data layer and an operating model that many institutions can use safely.","stakeholders":["State AI hub or economic development office","Consortium/MGHPCC-style operator director","University research computing leads","State or campus CISO"],"discoveryQuestions":["Who funds backups and operations after the initial grant?","What recovery objectives apply to shared research data?","How will allocations, caps and acceptable-use rules be enforced on storage?"],"valueHypothesis":"A single all-flash namespace may simplify serving multiple GPU pools for many institutions; AICR shows adoption and configuration, not measured benefit.","potentialEngagement":"A scoped review of a planned state or consortium AI resource covering storage protection tiers, backup options and allocation-linked quotas.","unsupportedClaims":["Any measured throughput or research-productivity gain at AICR","Cost savings versus alternatives","Production maturity (system is listed as Beta)"]},"engineering":{"fit":"Fits shared multi-pool GPU clusters needing one high-bandwidth namespace with GPU Direct Storage.","architectureIntegration":"VAST all-flash via NFS-RDMA on cluster InfiniBand NDR400, separate rail-optimized IB for B200 GPU traffic, Slurm partitions per GPU class.","prerequisites":["Storage InfiniBand fabric","Snapshot retention policy per tier","Decision on off-site replication or backup target"],"constraints":["No off-site backups","Scratch has no snapshots and is purged","360 GB/s is a rated figure, not measured application throughput"],"security":"Encryption at rest is enabled; tenancy, key management and regulated-data handling are not documented, so they must be confirmed before sensitive data is placed on the system.","proofOfValue":["Measure sustained per-pool I/O under concurrent multi-node training","Run snapshot restore drills for home and work tiers","Test recovery from simulated deletion beyond snapshot retention"]},"delivery":{"work":"Operate shared storage tiers for a multi-institution AI cluster: quotas, purge, snapshots, maintenance and user guidance.","dependencies":["Consortium allocation process","Member university support teams","State funding for operations"],"ownership":"MGHPCC as operator with member-university research computing teams supporting users.","skillsAdoption":"Staff need VAST administration plus allocation and data-protection communication skills; users need guidance to self-manage backups.","governanceCheckpoints":["Approval of backup/recovery posture","Acceptable-use policy enforcement review","Transition from Beta to general allocation"],"acceptanceCriteria":["Proposed: documented recovery point and time objectives per tier","Proposed: successful snapshot restore test each quarter","Proposed: measured throughput reported per GPU pool"],"risks":["Irrecoverable data loss without off-site copies","Scratch purge surprises for users","Policy churn while in Beta"]}},"retrievedAt":"2026-10-30T22:07:58-05:00","enrichedAt":"2026-10-30T22:10:57-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Shared access lowers barriers for member institutions; MGHPCC and the state frame AICR as supporting workforce development, but users must have scaling-ready code to qualify.","procurementImplications":"State grant plus university match funded acquisition through an integrator; operations and data-protection costs (e.g., backup) appear outside the documented design and should be planned explicitly in similar procurements.","operatingModelImplications":"Consortium operator with per-project caps, twice-monthly proposal reviews and monthly maintenance; users carry backup responsibility.","peerReview":"unknown","sourceVerification":{"openedUrl":"https://docs.aicr.ai/system-description/","referenceExcerpt":"AICR storage relies on snapshots only — there are no off-site backups.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}