From the NVIDIA edition of September 18, 2026
Lambda MaxLPS PoC: +24% token throughput and +23% perf/W inside fixed power
NVIDIA / Lambda · AI cloud / data-center power management · Global (AI cloud PoC on HGX B200)
- Publisher
- NVIDIA Customer Stories
- Original publication
- undated case study (cited in Sep 18, 2026 trade press)
- Source retrieved
- 2026-09-19
- Event date
- 2026-09-18
What happened
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.
Why it matters
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 and measured results
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.
Limitations and uncertainty
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.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-19; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Sites are power-capped and cannot add GPUs without new utility capacity.
- Stakeholders
- facilities
- AI platform
- CFO
- utility liaison
- Discovery questions
- What is the fixed facility kW for the AI hall?
- Do you meter rack and GPU power today?
- Which workloads create recoverable headroom?
- Value hypothesis
- Bounded densification can raise tokens/s inside existing interconnect if telemetry and policy gates hold.
- Potential engagement
- Two-week MaxLPS PoC on one power domain with matched baseline.
- Unsupported claims
- Guaranteed 24% on any site
- Rubin MaxLPS already measured at Lambda
- Utility upgrade avoidance forever
Pre-sales engineering
Role takeaway
- Fit
- Strong for instrumented AI clouds/labs; weak without power telemetry.
- Architecture and integration
- 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.
- Proof of value
- 16→19 packing reproduced
- tokens/s and perf/W logged
- latency/error within SLO
Delivery
Role takeaway
- 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.
- Skills and adoption
- Operators need power-policy literacy, not only GPU scheduling.
- Governance checkpoints
- No densification without telemetry
- SLO gate before permanent policy
- Rubin claims deferred
- Acceptance criteria
- Facility kW never exceeded
- SLO hold documented
- Rollback tested
- Risks
- Headroom vanishes under sync bursts
- Cooling limits bind first
- Vendor PoC not transferable
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
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.
Governance
Who approves, reviews and stays accountable for outcomes?
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.
Security and privacy
What data, permissions and controls need testing?
No confidential-computing or multi-tenant isolation claims; power-control plane becomes a new operational trust boundary.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Higher shared inference capacity can improve research tool responsiveness; no accessibility features claimed.
Procurement
What should contracts, pricing and exit terms secure?
Use as evaluation input for power-constrained expansions; do not sole-source on partner PoC percentages.
Operating model
Which teams own the service once it runs?
Shifts capacity planning from nameplate racks toward continuous power-policy operations.
Publication history
- 2026-09-18NVIDIA · Issue 138 resources
Stable resource ID: lambda-dsx-maxlps-poc-2026