{"resourceId":"pitt-crcd-fy2026-throughput-ai-ops-2026","versions":[{"version":"external-311657e349efee4cacf00e29ee3061d0601e50b9a167d2ae5258c554be940621","resource":{"id":"pitt-crcd-fy2026-throughput-ai-ops-2026","title":"Pitt CRCD reports record FY2026 jobs and hours; LMaaS still prospective","organization":"University of Pittsburgh Center for Research Computing and Data","sector":"University research computing","geography":"United States (Pennsylvania)","publishedAt":"FY2026 CRCD annual report PDF (publication day not stated on document)","publicationDate":null,"eventDate":null,"sourceName":"University of Pittsburgh CRCD","sourceLabel":"Self-published research computing annual report; pdftotext inspected","sourceUrl":"https://www.crcd.pitt.edu/sites/default/files/assets/CRCD_FY2026_Annual_Report%20Final%20PDF.pdf","evidenceClass":"vendor-claim","outcomeClass":"mixed","topics":["infrastructure","operating-model","accessibility-workforce","developers-agents","governance-procurement"],"finding":"Pitt reports record research-computing throughput and storage-cost cuts, plus AI-assisted ops analysis, while private Language Model as a Service remains a planned—not measured—offering.","sledRelevance":"Newly relevant measured campus research-computing telemetry contrasting prospective national AI capacity awards; U.S. public research university setting. Does not prove scientific reproducibility or LMaaS efficacy.","evidence":"FY2026: 6,588,805 jobs and 3,425,520 compute hours (GPU 1,151,851 jobs / 542,690 hours); 184 grants / $95.6M expenditures; ~$4M upgrades; warm-tier storage $65→$30 per TB/year (~54%); 13 RTX 6000 Pro Blackwell nodes (104 GPUs) and 2 H200 nodes (16 GPUs); performance review used Claude on curated scheduler data; LMaaS and cloud bursting listed as services in development. No controlled evaluation of Claude analysis quality or released LMaaS outcomes.","architectureImplications":"Interpretation: separate high-performance all-flash paths from warm-tier cost storage; plan hybrid burst and eventual private model endpoints beside GPU queues.","governanceImplications":"Interpretation: publish how AI-assisted ops analyses are validated before they change allocation policy; keep LMaaS eligibility and cost rules explicit before launch.","securityPrivacyImplications":"Interpretation: private open-source LMaaS claims still need data-class controls, key management, and no-training assurances before sensitive research data use.","caveats":"Operator self-report without independent audit or clear PDF publication date. Throughput is not scientific outcome. LMaaS unreleased. Claude analysis methodology thin.","streamIds":["research"],"roles":{"sales":"Interpretation: Speak with research computing and VPR stakeholders who need measured utilization—not only award announcements. Ask how jobs, GPU hours, and storage costs are reported to leadership, and what would count as LMaaS success. Offer a bounded telemetry and service-definition workshop comparing Pitt-style metrics to local dashboards. Value hypothesis is clearer operating baselines before private-model spend. Unsupported: guaranteed storage savings elsewhere, Claude-driven ops quality, or LMaaS productivity gains.","engineering":"Interpretation: Fit is campus HPC with specialized HTC/SMP/GPU/MPI queues, tiered storage, and planned hybrid burst plus private LLM endpoints. Integrate scheduler telemetry for wait-time analysis; keep AI ops tools on curated extracts. Prerequisites: allocation accounting and storage tiers. Constraints: self-reported figures and unreleased LMaaS. Security: isolate model endpoints and scheduler data. Proof: reproduce job/hour dashboards and storage unit-cost math on local data—not vendor slides.","delivery":"Interpretation: Research computing owns platform metrics and future LMaaS ops; PIs own scientific use. Work includes publishing annual throughput, validating AI-assisted reviews, and piloting private models with eligibility rules. Dependencies: storage vendors, GPU supply, and identity. Train staff on interpreting wait-time analyses. Checkpoints: finance sign-off on storage pricing and pre-LMaaS security review. Acceptance: documented FY metrics, published rates, and a LMaaS pilot checklist. Risks: mistaking job counts for discovery impact and launching LMaaS without cost controls."},"roleDetails":{"sales":{"customerProblem":"Leadership needs credible research-computing utilization and cost signals before funding private AI inference or more GPUs.","stakeholders":["Research computing leadership","VPR / research administration","Finance / sponsored projects","Faculty advisory committees"],"discoveryQuestions":["Which utilization and cost metrics are reported annually today?","How are GPU hours attributed to grants?","What acceptance criteria would green-light private LMaaS?","Who validates AI-assisted ops recommendations?"],"valueHypothesis":"Transparent job, hour, and storage-cost telemetry can ground capacity and LMaaS decisions better than award announcements alone.","potentialEngagement":"Bounded metrics-and-service workshop to draft a campus research-computing scorecard and LMaaS eligibility sketch.","unsupportedClaims":["Transferable $65→$30/TB savings without local cost model","Proven quality of Claude-based ops analysis","Measured scientific gains from LMaaS"]},"engineering":{"fit":"Campus multi-cluster HPC with GPU nodes, tiered storage, and planned private generative-AI service.","architectureIntegration":"Retain HTC/SMP/GPU/MPI separation; place AI training/inference on GPU partitions; warm-tier for bulk; eventual LMaaS beside identity and project quotas.","prerequisites":["Scheduler accounting","Storage quota policy","IdP integration","Curated ops datasets for analysis"],"constraints":["Self-reported FY figures","LMaaS not yet in production","GPU supply/cost volatility"],"security":"Protect scheduler extracts; sandbox LMaaS; prohibit sensitive data until classification and logging exist.","proofOfValue":["Reconcile job/hour totals to scheduler logs","Unit-cost storage math","LMaaS pilot latency/error vs commercial API on approved workloads"]},"delivery":{"work":"Publish annual metrics, validate AI-assisted bottleneck reviews, and stage a private-model pilot with cost and eligibility rules.","dependencies":["Storage and GPU vendors","Identity team","Research compliance"],"ownership":"CRCD-equivalent operates infrastructure; PIs own scientific claims.","skillsAdoption":"Train facilitators on reading utilization dashboards and when not to place data on LMaaS.","governanceCheckpoints":["Finance approval of storage rates","Pre-LMaaS security review","Annual metrics publication"],"acceptanceCriteria":["FY job/hour figures reproducible from logs","Published storage rates","LMaaS pilot checklist signed"],"risks":["Job counts mistaken for discovery impact","Unvalidated AI ops changes","LMaaS cost overruns"]}},"retrievedAt":"2026-09-17T03:04:08Z","enrichedAt":"2026-09-17T03:07:50Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Interpretation: lower storage prices and teaching-cluster UI changes can widen access; still teach data classification before LMaaS use.","procurementImplications":"Interpretation: use local unit-cost and utilization evidence when negotiating flash, warm-tier, and GPU buys; treat LMaaS build-vs-buy as a measured pilot.","operatingModelImplications":"Interpretation: research computing publishes telemetry; research offices interpret grant impact; AI ops recommendations need human acceptance before policy changes.","sourceVerification":{"openedUrl":"https://www.crcd.pitt.edu/sites/default/files/assets/CRCD_FY2026_Annual_Report%20Final%20PDF.pdf","referenceExcerpt":"Across all clusters, CRCD completed 6,588,805 jobs and delivered 3,425,520 compute hours in FY2026","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}