From the NVIDIA edition of September 16, 2026
AI Infra Summit: Vera Rubin, DSX MaxLPS, and partner tokens-per-watt results
NVIDIA · AI infrastructure / cloud hyperscale partners · United States (Santa Clara AI Infra Summit coverage)
- Publisher
- NVIDIA Blog
- Original publication
- September 15, 2026
- Source retrieved
- 2026-09-17
- Event date
- 2026-09-15
What happened
On September 15, 2026 NVIDIA published an AI Infra Summit recap claiming Vera Rubin NVL72 MLPerf Inference v6.1 preview throughput up to 3.7x vs GB300, Lambda DSX MaxLPS validation (+24% token throughput / +23% perf-per-watt within a fixed power budget), Emerald AI flexible-load demos with Silicon Valley Power, and SemiAnalysis AgentX claims of up to 30x throughput per megawatt vs GB300 for agentic coding sessions.
Why it matters
Relevant as ecosystem direction on power-constrained AI factories and agentic inference economics. Most claims are hyperscale/cloud-partner measured; SLED campuses rarely match 250 MW-class DSX sites. Treat Lambda/Emerald results as transferability signals for power budgeting—not turnkey campus outcomes.
Evidence and measured results
NVIDIA Writers post dated September 15, 2026 reports Ian Buck AI Infra Summit remarks; lists Amazon Annapurna NVHBM and d-Matrix NVLink Fusion collaborations; states Lambda ran 19 nodes in a 16-node power budget (+24% tokens/s, +23% perf/watt) and cites Vera Rubin MLPerf preview up to 3.7x GB300 and AgentX up to 30x tokens/MW vs GB300.
Limitations and uncertainty
Vendor-authored summit recap aggregating partner demos and MLPerf preview numbers without independent replication in this source. Hyperscale power envelopes and agentic coding AgentX trajectories may not map to campus/agency workloads.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-17; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Leadership wants more AI capacity without new utility interconnects or megawatt upgrades.
- Stakeholders
- CIO/CTO
- facilities/energy manager
- AI platform owner
- finance
- Discovery questions
- What is the hard site or room power ceiling today?
- Are workloads mixed training/inference with variable draw?
- Which partner-measured result (MaxLPS, MLPerf, AgentX) maps to your jobs?
- Value hypothesis
- Dynamic power allocation and next-gen racks can raise tokens per megawatt when the estate resembles partner AI-factory conditions.
- Potential engagement
- Power-envelope discovery workshop comparing current rack draw to MaxLPS-style reclaim assumptions and a scoped pilot SLO.
- Unsupported claims
- Guaranteed 3.7x or 30x campus gains
- Immediate utility interconnect relief
- Turnkey SLED DSX factory
Pre-sales engineering
Role takeaway
- Fit
- Fit for large AI factories or cloud estates evaluating Vera Rubin/DSX; limited direct fit for single-room campus DGX without partner-class power telemetry.
- Architecture and integration
- Integrate DSX MaxLPS/Flex-style power control with rack PDUs, serving stack (Dynamo), and job priority policies.
- Prerequisites
- Per-GPU/rack power telemetry
- Workload priority taxonomy
- Validated cooling headroom
- Constraints
- Hyperscale reference envelopes
- MLPerf/AgentX workload specificity
- Partner-reported—not third-party audited here
- Security
- BlueField and factory security mentioned at marketing depth only; design attestation and tenancy separately.
- Proof of value
- Measured tokens/s within fixed kW budget
- Documented shed/restore of low-priority jobs
- Side-by-side vs static power caps
Delivery
Role takeaway
- Work
- Facility telemetry enablement, power-policy definition, and controlled density pilot.
- Dependencies
- Facilities engineering
- utility coordination if Flex
- serving-platform team
- Ownership
- Joint facilities + AI platform operations.
- Skills and adoption
- Operators need power-aware scheduling literacy beyond GPU ops.
- Governance checkpoints
- Utility/program eligibility review
- Workload priority approval
- Pilot go/no-go on measured tokens/MW
- Acceptance criteria
- Agreed baseline tokens/s at capped kW
- Documented reclaim without priority-job breach
- Written transferability limits to campus rooms
- Risks
- Over-promising hyperscale multipliers
- Cooling/power trip during reclaim
- Partner demo non-reproducibility
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?
Codesign power management (DSX MaxLPS/Flex), scale-up fabric (NVLink), and Dynamo/TensorRT-LLM serving as one stack when evaluating next-gen racks; do not treat GPU SKU alone as the capacity plan.
Governance
Who approves, reviews and stays accountable for outcomes?
Grid flexible-load programs (Emerald AI / municipal utility) imply operational policies for which jobs may shed—relevant if a public institution joins demand-response, but not evidenced for typical SLED estates.
Security and privacy
What data, permissions and controls need testing?
Post emphasizes BlueField DPUs for infrastructure security at factory scale; no attestation measurements or sovereign-control evaluation.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Summit narrative is infrastructure/economics focused; no workforce accessibility evidence for public-sector staff.
Procurement
What should contracts, pricing and exit terms secure?
Useful for RFP language on tokens-per-megawatt and power reclaim, but require bidder-measured pilots—not summit headlines alone.
Operating model
Which teams own the service once it runs?
Implies factory-like ops with dynamic power and job hierarchy; most SLED IT orgs will need new facilities–platform operating agreements.
Publication history
- 2026-09-16NVIDIA · Issue 119 resources
Stable resource ID: nvidia-ai-infra-summit-vera-rubin-dsx-20260915