Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 17, 2026

Vendor claimEmergingNew this fortnight

Vera Rubin NVL72 MLPerf Inference v6.1 preview: up to 3.7x vs GB300 on Qwen3-VL

NVIDIA · AI infrastructure / inference benchmarking · Global (vendor MLPerf submission blog)

NVIDIA
Publisher
NVIDIA Blog
Original publication
September 16, 2026
Source retrieved
2026-09-18
Event date
2026-09-16
Read original source

What happened

On September 16, 2026 NVIDIA published Vera Rubin NVL72's first MLPerf Inference v6.1 preview results: up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL (vLLM+Dynamo) across offline/server/interactive, and up to 2.5x on DeepSeek-R1 (TensorRT-LLM). A four-rack GB300 DeepSeek-R1 offline run reached 99% scaling efficiency (72→288 GPUs). Software optimizations delivered up to 1.6x Qwen3-VL gains vs v6.0. The post also cites SemiAnalysis AgentX preview of 30x vs GB300 and notes Nebius as another Vera Rubin preview submitter among 19 partners.

Why it matters

Gives SLED buyers a peer-suite preview of next-gen rack throughput and software-stack dependence (Dynamo/vLLM vs TensorRT-LLM), but preview silicon and hyperscale partner lists do not equal campus/agency availability or measured tokens-per-dollar.

Evidence and measured results

Blog dated Sep 16, 2026 cites Closed Division entries 6.1-0106/0074 for Vera Rubin vs GB300; 6.1-0073/0074 for four-rack 99% offline scaling; WAN 2.2 0.65 720p videos/s at 5.7s; partner list including Azure, CoreWeave, Dell, HPE, Lambda, OCI, Supermicro; AgentX 30x claim referenced as preview testing outside MLPerf.

Limitations and uncertainty

Vendor-authored summary of its own submissions; 'up to' spans scenarios/models/stacks; AgentX 30x is not an MLPerf result; no measured system power or rack price.

Cross-source patterns this evidence supports

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-18; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.

Sales

Role takeaway
Customer problem
Buyers need credible next-gen inference throughput evidence without treating preview blogs as purchase specs.
Stakeholders
  • CIO/CTO
  • AI platform owner
  • procurement
  • facilities/power
Discovery questions
  • Which model and latency scenario dominate production?
  • Is the serving stack vLLM/Dynamo or TensorRT-LLM?
  • Is Vera Rubin capacity actually quotable for your region?
Value hypothesis
Peer-suite preview results can bound expected rack-scale uplift when stack and scenario match.
Potential engagement
Scenario-matched readout workshop comparing buyer workload card to published MLPerf entries.
Unsupported claims
  • Guaranteed 3.7x on any model
  • Lower cost per token proven
  • Broad commercial availability now

Pre-sales engineering

Role takeaway
Fit
Fit for teams already on NVL72-class fabrics evaluating Rubin; poor fit for single-GPU or non-NVLink estates expecting the same multipliers.
Architecture and integration
Reproduce with disaggregated prefill/decode, NVFP4, and full rack NVLink; pin Dynamo/vLLM or TensorRT-LLM versions.
Prerequisites
  • NVL72 topology
  • Chosen inference framework expertise
  • Access to preview/partner capacity
Constraints
  • Preview system
  • Scenario-specific multipliers
  • No power submission
Security
No attestation or tenancy data in source—keep security evaluation separate.
Proof of value
  • Match within 20% of cited scenario on agreed model
  • Document framework/precision pins
  • Separate AgentX claims from MLPerf gates

Delivery

Role takeaway
Work
Baseline current fleet, pin software, run scenario tests, and gate any refresh on availability.
Dependencies
  • Cloud/OEM capacity
  • Model owners
  • Network fabric team
Ownership
AI platform engineering lead.
Skills and adoption
Team needs Dynamo/disaggregation and MLPerf scenario literacy.
Governance checkpoints
  • Preview risk acceptance
  • Procurement exception for non-GA hardware
  • Scenario sign-off
Acceptance criteria
  • Published entry IDs referenced in test plan
  • Interactive and server results both captured
  • No AgentX metric in contractual acceptance
Risks
  • Preview silicon variance
  • Software stack mismatch
  • Capacity slip

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?

Expect disaggregated prefill/decode, NVFP4, and full NVL72 NVLink domain to be prerequisites for approaching published multipliers; Jetson Thor Edge-Agentic path is a separate form factor.

Governance

Who approves, reviews and stays accountable for outcomes?

Treat as preview evidence—require MLCommons entry IDs and scenario matching in procurement evaluations before generational refresh decisions.

Security and privacy

What data, permissions and controls need testing?

No CC or tenancy measurements in this post; do not infer security posture from throughput leadership.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Higher inference capacity can improve shared research tooling responsiveness; no accessibility features claimed.

Procurement

What should contracts, pricing and exit terms secure?

Use as evaluation input with mandatory scenario/stack matching; do not sole-source on headline multipliers.

Operating model

Which teams own the service once it runs?

Encourages treating inference economics as continuous software optimization, not one-time hardware purchase.

Publication history

  1. 2026-09-17NVIDIA · Issue 129 resources
Read preserved resource versions (JSON)

Stable resource ID: nvidia-vera-rubin-mlperf-v61-20260916