Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 21, 2026

Vendor claimEmergingRecent

NVIDIA technical blog: Vera Rubin up to 30× AgentX throughput per MW vs GB300

NVIDIA · AI factory benchmarking / agentic inference · Global (vendor technical blog)

NVIDIA
Publisher
NVIDIA Technical Blog
Original publication
August 24, 2026
Source retrieved
2026-09-22
Event date
2026-08-24
Read original source

What happened

NVIDIA’s Aug 24, 2026 technical blog reports SemiAnalysis AgentX results showing Vera Rubin NVL72 delivering up to 30× higher AI-factory throughput per megawatt than GB300 NVL72 at 160 tokens per second per user on an AgentX DeepSeek V4-Pro workload, and cites GB300 advantages versus H200 on other AgentX workloads. It describes AgentX as replaying prerecorded production-style coding sessions with interleaved reasoning and tool use, and points to system software (SGLang/TensorRT-LLM/vLLM, Dynamo, precision formats) as co-design factors. Newly relevant as Sep multi-rack and DSX narratives continue to lean on agentic efficiency claims.

Why it matters

Provides the primary vendor statement of the 30× AgentX per-MW figure SLED teams will see in decks—must be read as coding-agentic, point-specific, vendor-reported preview results, not general SLED chatbot SLOs.

Evidence and measured results

Publish Aug 24, 2026; up to 30× Rubin vs GB300 at 160 TPS/user on AgentX DeepSeek V4-Pro; AgentX uses prerecorded production-style sessions; GB300 vs H200 up to 15× / 80× cited for other models; stack mentions SGLang, TRT-LLM, vLLM, Dynamo, MXFP4/8.

Limitations and uncertainty

Vendor blog summarizing AgentX; extreme ratios are operating-point sensitive; not a campus productivity study.

Cross-source patterns this evidence supports

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-22; 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 see headline 30× agentic efficiency claims without workload context.
Stakeholders
  • AI platform
  • CFO
  • engineering
Discovery questions
  • Is your workload coding-agentic with tool use?
  • What TPS/user target matters?
  • Will you accept AgentX or need domain traces?
Value hypothesis
Point-specific AgentX gains may inform architecture choice if scoped tightly.
Potential engagement
Scoped AgentX-like replay on one domain with power metering.
Unsupported claims
  • 30× for all SLED chatbots
  • Guaranteed factory TCO
  • Preview equals GA software

Pre-sales engineering

Role takeaway
Fit
Relevant for agentic coding inference evaluations on NVL72.
Architecture and integration
Pin serving runtime, precision, Dynamo/KV settings, and power domains.
Prerequisites
  • power metering
  • engine pins
  • trace ethics approval
Constraints
  • Vendor-reported
  • 160 TPS/user point
  • coding-session corpus
Security
Control access to traces and evaluation clusters.
Proof of value
  • Matched-point replay
  • Power-normalized metrics logged
  • Sensitivity sweep

Delivery

Role takeaway
Work
Run scoped evaluation; document point and caveats in decision package.
Dependencies
  • benchmark lead
  • facilities metering
  • vendor SE
Ownership
AI platform evaluation owner.
Skills and adoption
Teams need agentic benchmark literacy.
Governance checkpoints
  • No 30× in unmarked slides
  • Workload class approved
  • Preview software labeled
Acceptance criteria
  • Config pin filed
  • Metrics reproducible
  • Caveats in memo
Risks
  • SLO mismatch
  • Software drift
  • Overfit to AgentX coding mix

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?

Agentic efficiency depends on MoE serving runtimes, KV-cache behavior, and power-normalized factory metrics—not peak chip TOPS alone.

Governance

Who approves, reviews and stays accountable for outcomes?

Label as vendor-reported AgentX preview; require matching interactivity point and workload class before using in business cases.

Security and privacy

What data, permissions and controls need testing?

Benchmark uses anonymized coding sessions per SemiAnalysis design; source does not address tenant isolation.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Agentic coding efficiency may affect developer tooling capacity; no accessibility features claimed.

Procurement

What should contracts, pricing and exit terms secure?

Require workload-class and point disclosure whenever AgentX ratios appear in proposals.

Operating model

Which teams own the service once it runs?

Factory KPIs may shift toward power-normalized agentic throughput—only if workloads match.

Publication history

  1. 2026-09-21NVIDIA · Issue 168 resources
Read preserved resource versions (JSON)

Stable resource ID: nvidia-agentx-rubin-30x-per-mw-20260824