{"resourceId":"lambda-dsx-maxlps-poc-2026","versions":[{"version":"external-bd78677776292829a4c3e07868808cd2c9cf4fedb6d68e081ad1e50319a6a8fa","resource":{"id":"lambda-dsx-maxlps-poc-2026","title":"Lambda MaxLPS PoC: +24% token throughput and +23% perf/W inside fixed power","organization":"NVIDIA / Lambda","sector":"AI cloud / data-center power management","geography":"Global (AI cloud PoC on HGX B200)","publishedAt":"undated case study (cited in Sep 18, 2026 trade press)","publicationDate":null,"eventDate":"2026-09-18","sourceName":"NVIDIA Customer Stories","sourceLabel":"Vendor partner case study","sourceUrl":"https://www.nvidia.com/en-us/case-studies/lambda/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["infrastructure","operating-model","developers-agents"],"finding":"NVIDIA’s Lambda case study reports a five-rack, 19-node HGX B200 cluster under DSX MaxLPS: at an 85% power policy, 19 nodes fit the prior 16-node full-power budget, lifting cluster token throughput from ~4.04M to ~5M tokens/s (+24%) and performance per watt by 23%. Hybrid 80% policy runs (10 inference + 10 training) cite +20% inference and +17% training throughput by reclaiming stranded headroom. Lambda’s fleet narrative spans Ampere through Vera Rubin, but measured MaxLPS figures are B200-scoped.","sledRelevance":"Gives SLED facilities and AI-cloud buyers a concrete fixed-power densification datapoint—but PoC scale, MLPerf-like peaks, and vendor authorship mean campus/agency sites still need matched workload and telemetry proofs before budgeting “free” nodes.","evidence":"Key takeaways: +24% cluster-wide token throughput (19 vs 16 nodes same budget); +23% perf/W; +20% hybrid inference / +17% hybrid training at 80% policy; baseline ~4.04M tokens/s GPT-OSS-120B @ 40 QPS/node; DPS monitors GPU/rack draw; policies tested at 80%/85% of max; five-rack 19-node HGX B200 cluster.","architectureImplications":"Requires rack/GPU power telemetry, DPS-managed groups, and policy targets that intentionally leave per-node headroom so additional nodes can join the same facility envelope.","governanceImplications":"Treat as partner PoC evidence: require reproducible power traces, workload pins, and latency/error SLOs before encoding densification into capacity plans or utility interconnection filings.","securityPrivacyImplications":"No confidential-computing or multi-tenant isolation claims; power-control plane becomes a new operational trust boundary.","caveats":"Vendor-published partner story; MLPerf-style peak loads may overstate recoverable headroom vs production mixed traffic; Vera Rubin fleet mention is not a measured MaxLPS result on Rubin.","streamIds":["nvidia"],"roles":{"sales":"Frame MaxLPS as reclaiming stranded facility power—not as a chip upgrade. Open with the buyer’s fixed megawatt and whether they can instrument rack draw, then cite Lambda’s 19-vs-16 node PoC (+24% tokens/s, +23% perf/W) as an existence proof for densification under an 85% policy. Bound claims to HGX B200 PoC conditions and insist on latency/stability gates before any “more GPUs without new power” sales motion.","engineering":"Fit when the site already has or can add GPU/rack power telemetry and a managed power-policy path. Architecture integration is DPS resource groups plus per-node power caps (80–85% tested) without changing running workloads; prerequisites include representative loads and wall-power metering. Proof of value: reproduce 16→19 node packing with unchanged facility budget and capture latency/error alongside tokens/s.","delivery":"Work is a scoped densification pilot: baseline static provisioning, apply MaxLPS policy, add nodes only after telemetry proves headroom. Ownership sits with facilities and platform SRE jointly. Acceptance criteria must include facility-budget compliance, SLO hold, and rollback to static caps. Risks: production burstiness eliminates headroom; cooling/electrical margins tighter than PoC; over-promising Rubin carryover."},"roleDetails":{"sales":{"customerProblem":"Sites are power-capped and cannot add GPUs without new utility capacity.","stakeholders":["facilities","AI platform","CFO","utility liaison"],"discoveryQuestions":["What is the fixed facility kW for the AI hall?","Do you meter rack and GPU power today?","Which workloads create recoverable headroom?"],"valueHypothesis":"Bounded densification can raise tokens/s inside existing interconnect if telemetry and policy gates hold.","potentialEngagement":"Two-week MaxLPS PoC on one power domain with matched baseline.","unsupportedClaims":["Guaranteed 24% on any site","Rubin MaxLPS already measured at Lambda","Utility upgrade avoidance forever"]},"engineering":{"fit":"Strong for instrumented AI clouds/labs; weak without power telemetry.","architectureIntegration":"DPS-managed groups with per-node power targets; no workload rewrite required in PoC narrative.","prerequisites":["GPU/rack telemetry","policy control path","representative loads","facility budget ceiling"],"constraints":["B200 PoC scope","MLPerf-like peaks","five-rack scale"],"security":"Power-control plane access must be hardened; source silent on CC.","proofOfValue":["16→19 packing reproduced","tokens/s and perf/W logged","latency/error within SLO"]},"delivery":{"work":"Run static vs MaxLPS A/B on one domain; document policy and rollback.","dependencies":["facilities electrical","NVIDIA/partner SE","observability"],"ownership":"Joint facilities and AI platform operations.","skillsAdoption":"Operators need power-policy literacy, not only GPU scheduling.","governanceCheckpoints":["No densification without telemetry","SLO gate before permanent policy","Rubin claims deferred"],"acceptanceCriteria":["Facility kW never exceeded","SLO hold documented","Rollback tested"],"risks":["Headroom vanishes under sync bursts","Cooling limits bind first","Vendor PoC not transferable"]}},"retrievedAt":"2026-09-19T03:12:00Z","enrichedAt":"2026-09-19T03:17:22Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Higher shared inference capacity can improve research tool responsiveness; no accessibility features claimed.","procurementImplications":"Use as evaluation input for power-constrained expansions; do not sole-source on partner PoC percentages.","operatingModelImplications":"Shifts capacity planning from nameplate racks toward continuous power-policy operations.","sourceVerification":{"openedUrl":"https://www.nvidia.com/en-us/case-studies/lambda/","referenceExcerpt":"Lambda ran 19 nodes instead of 16 within the same power budget, increasing cluster-wide token throughput by 24%","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}