Strategic PartnersIssue 129 records
NVIDIA
This edition moves from yesterday's AI Infra Summit/DSX/Dynamo/CC operator set to September 16–17 peer-suite and grid-flexibility evidence: NVIDIA's Vera Rubin NVL72 MLPerf Inference v6.1 preview (up to 3.7x/2.5x vs GB300 by model/stack), MLCommons' record v6.1 round marking Rubin/Vera Rubin preview and new RAG/Edge Agentic tests, independent DIY AI scrutiny that scenario gains and missing wall power leave cost-per-token unproven, AEMA flexible-data-center launch with Axios/Latitude enforceability and FERC/SB6 context, CoreWeave multi-rack Vera Rubin cloud bring-up, Groq 3 LPX deterministic power controls explaining extreme tokens/MW slides, and GenAI Daily caveats that AgentX's 30x peak is early-silicon and coding-corpus only.
- Evidence records
- 9
- Cross-source patterns
- 3
- Evidence classes
- 4 vendor claim4 independent reporting1 standards or public-body guidance
- Outcomes
- 4 emerging3 cautionary1 effective1 mixed
- Source freshness
- 9 new this fortnight
- Research completed
- 2026-09-18
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Vera Rubin MLPerf multipliers are stack- and scenario-scoped
NVIDIA's Sep 16 MLPerf blog leads with up to 3.7x on Qwen3-VL (vLLM+Dynamo) and 2.5x on DeepSeek-R1 (TensorRT-LLM), MLCommons marks Vera Rubin as preview, and DIY AI's entry-level table shows Interactive peaks near ~3.7x while Offline/Server gains often sit nearer ~1.7–2.0x—with no measured system power submission to convert throughput into cost-per-token.
Operating questionWhich model, scenario, and serving stack were matched before using a Vera Rubin MLPerf multiplier in a capacity or budget plan?
Supporting evidenceVera Rubin NVL72 MLPerf Inference v6.1 preview: up to 3.7x vs GB300 on Qwen3-VLMLCommonsDIY AI
AI factory grid flexibility is policy-real but enforceability-gated
NVIDIA's AEMA launch promotes technology-neutral flexible data centers, while Axios and Latitude report the coalition's regulatory push (FERC show-cause, SB6-like interconnects) alongside public rate backlash and Sivaram's own 'verifiable and enforceable' condition—plus limits that flexibility will not fix water/land/noise/tax opposition.
Operating questionWhat measurable curtailment proof and non-power community mitigations exist before counting on faster interconnection?
Supporting evidenceNVIDIA, Google, and Emerald AI launch AI Energy Management Alliance for flexible data centersAxiosLatitude Media
Agentic tokens-per-watt claims need tiered, corpus-aware proof
Vendor MLPerf/AgentX messaging and Groq 3 LPX power deep dives advertise large tokens-per-megawatt upsides for high-interactivity agents, CoreWeave is bringing multi-rack Rubin to cloud agent loops, and GenAI Daily documents that the 30x AgentX peak is early-silicon, coding-corpus, and concurrency-tier dependent—so peak slides are not default planning numbers.
Operating questionFor which interactivity tier and agent corpus was tokens-per-megawatt measured on hardware you can actually buy?
Supporting evidenceGenAI DailyGroq 3 LPX determinism: PEP/CPS cut voltage droop >60% on Vera Rubin power stackCoreWeaveVera Rubin NVL72 MLPerf Inference v6.1 preview: up to 3.7x vs GB300 on Qwen3-VL
Evidence ledger
Full record · every source keeps its link and limitations
Vera Rubin NVL72 MLPerf Inference v6.1 preview: up to 3.7x vs GB300 on Qwen3-VL
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, evidence and limitations
- 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.
MLCommons MLPerf Inference v6.1: record participation, new RAG and Edge Agentic tests
MLCommons announced MLPerf Inference v6.1 on September 16, 2026 with record 30 submitting organizations, new End-to-End RAG and Edge Agentic Inference tests, speculative decoding support in interactive scenarios, and first peer-reviewed results for several new platforms including NVIDIA Rubin and Vera Rubin NVL72 marked in preview. Suite-level claims include up to 5.7x DeepSeek-R1 per-accelerator server gain vs v5.1 and 2.99x VLM server per-accelerator gain vs v6.0. Over 50% of submitters used a new API-centric harness toward MLPerf Endpoints.
Why it matters, evidence and limitations
- Why it matters
- Provides the independent suite framing SLED evaluators should prefer over vendor blogs alone—especially the preview label on Vera Rubin and the shift toward agentic/RAG workloads that campuses actually run.
- Evidence and measured results
- Press release dated Sep 16, 2026 lists 30 organizations, names NVIDIA Rubin/Vera Rubin NVL72 as preview, introduces RAG and Edge Agentic tests, notes largest-ever 512-accelerator submission and geographically distributed heterogeneous systems, and points to Datacenter/Edge result pages and visualizer.
- Limitations and uncertainty
- Suite-level 'up to' gains are across best per-accelerator results, not a single NVIDIA-vs-NVIDIA rack comparison; detailed scores live on result pages not fully reproduced here.
Independent read: Vera Rubin MLPerf gains are real but cost-per-token still unproven
Steven Jones (Sep 16, 2026) cross-checks public MLPerf entries and finds Vera Rubin NVL72 vs like-sized GB300 gains ranging roughly 1.74x–3.74x by workload/scenario (largest on Qwen3-VL Interactive ~3.74x; DeepSeek-R1 Interactive ~2.58x; Offline/Server often ~1.7–2.0x). Notes submissions are not MLPerf power results, lists nameplate TGP sums (72×2300W vs 72×1400W) as non-authoritative, stresses software stack differences (TensorRT-LLM vs vLLM+Dynamo; FP4 vs BF16), preview status, and that cost-per-useful-token remains unproven without wall power and pricing.
Why it matters, evidence and limitations
- Why it matters
- Directly usable scrutiny for SLED procurement: prevents treating the vendor 3.7x headline as a blanket refresh ROI and supplies a buyer request list (wall power, matched scenarios, sustained throughput, software pins).
- Evidence and measured results
- Article publishes a six-row comparison table from public MLPerf entries for VR200 NVL72 vs GB300 NVL72; states Rubin results are not power submissions; flags preview marking; contrasts 512-GPU AMD aggregate leaderboard vs per-GPU normalization; lists buyer follow-ups including measured wall power and migration economics vs existing Blackwell.
- Limitations and uncertainty
- Secondary analysis of public tables; nameplate TGP arithmetic is illustrative not measured rack power; author incentives are independent-media not lab audit.
NVIDIA, Google, and Emerald AI launch AI Energy Management Alliance for flexible data centers
On September 16, 2026 NVIDIA announced with Emerald AI and Google the AI Energy Management Alliance (AEMA) to advance data centers that dynamically manage electricity use in response to grid conditions via workload shifting, storage discharge, paired generation, or contingency response. AEMA is described as technology-neutral and performance-based (response speed, duration, predictability, emergency behavior). Principles include pre-defined ride-through/curtailment obligations, standardized metrics/data sharing, faster pathways for verifiable flexibility commitments, and cost allocation reflecting avoided upgrades.
Why it matters, evidence and limitations
- Why it matters
- Signals that large NVIDIA-centric AI factories are being designed as controllable loads—relevant to state/local permitting and university/state research data centers facing interconnection queues, without implying every SLED GPU room can or should curtail.
- Evidence and measured results
- NVIDIA blog by Josh Parker dated Sep 16, 2026 states founding announcement, flexibility mechanisms, technology-neutral performance focus, and four principle bullets on ride-through, metrics, faster pathways, and cost allocation; notes NVIDIA+Emerald already working on real-time responsive AI factories.
- Limitations and uncertainty
- Vendor/coalition advocacy post; no measured curtailment SLAs or SLED case studies in-page; membership list not fully enumerated here.
Axios: AEMA flexible-power coalition faces public backlash and enforceability tests
Amy Harder's Sep 16, 2026 Axios report describes AEMA as a Google/NVIDIA/Emerald AI-backed relaunch of a formerly inactive Advanced Energy Management Alliance, now focused on flexible AI data centers among ~20 members (Anthropic, National Grid, AES, Constellation, NRG, RWE cited). Notes Google's ~1 GW demand-response portfolio claim, public polling (84% concerned about data-center impacts on local rates; bipartisan support for developers paying for upgrades), and Emerald CEO Sivaram's insistence that faster/larger connections require 'verifiable and enforceable' flexibility.
Why it matters, evidence and limitations
- Why it matters
- Gives state/local and campus stakeholders independent context: flexibility advocacy is real, but political legitimacy hinges on enforceability and ratepayer concerns—not vendor win-win messaging alone.
- Evidence and measured results
- Axios article dated Sep 16, 2026 quotes Josh Parker and Varun Sivaram; cites AP-NORC/UChicago EPIC polling figures; describes coalition size/partners and regulatory advocacy intent; reality-check paragraph on community persuasion difficulty.
- Limitations and uncertainty
- News summary of announcement call; polling is public-opinion not engineering validation; partner list may be incomplete.
Latitude Media: AEMA targets FERC show-cause dockets, ERCOT SB6-style flexible interconnects
Lisa Martine Jenkins (Sep 16, 2026) reports AEMA will leverage prior Advanced Energy Management Alliance regulatory standing and relationships with DOE/FERC/state commissions. Near-term focus: FERC June show-cause orders on large-load interconnection rules outside ERCOT, plus advocacy for ERCOT SB6-like faster/larger connections for flexible loads and Silicon Valley Power-style performance-based standards. Notes Google's view that power—not capital/silicon—is the limiter; NVIDIA/Emerald six demonstrations with a ~100 MW Manassas, VA flexible AI factory targeted before end of 2026; flexibility does not address water/land/noise/tax concerns driving opposition (cites NYT poll 61% oppose AI data centers).
Why it matters, evidence and limitations
- Why it matters
- Maps NVIDIA-linked flexibility advocacy onto concrete U.S. regulatory venues SLED energy and higher-ed facilities teams already watch, while clarifying limits versus community opposition drivers.
- Evidence and measured results
- Article quotes Tyler Norris and Varun Sivaram; details FERC show-cause engagement, SB6 analogy, SVP program reference, Manassas 100 MW timeline, six demonstrations, and community opposition caveats including water/land/noise/tax abatements.
- Limitations and uncertainty
- Reporting on announced plans and interviews; Manassas facility not yet online; poll figures are opinion context.
CoreWeave brings up multi-rack Vera Rubin NVL72 cluster with storage cross-region writes
On September 16, 2026 CoreWeave announced multi-rack NVIDIA Vera Rubin NVL72 bring-up on CoreWeave Cloud, unifying hundreds of Rubin GPUs via Spectrum-X into a scale-out cluster for agentic AI, alongside AI Object Storage features: cross-region write acceleration (local write with background replication) and a new Archive tier (no retrieve/early-delete/read fees from Archive as stated). Single rack described as 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9, BlueField-4. Mission Control Rack LifeCycle Controller automates detection, firmware, validation, power, and cooling.
Why it matters, evidence and limitations
- Why it matters
- Shows partner cloud path to Vera Rubin beyond single-rack demos; storage latency features matter for multi-step agents. Still a cloud-provider claim—not a government deployment outcome.
- Evidence and measured results
- Press release dated Sep 16, 2026 quotes Chen Goldberg and Cohere's Cécile Robert-Michon on unified multi-region datasets; details LOTA cache claims (up to 7 GB/s per GPU, 8x lower latency vs traditional storage cluster reads); describes multiplane multirail ConnectX-9 scale-out.
- Limitations and uncertainty
- Partner marketing announcement; no independent benchmark card for the multi-rack bring-up in-page; Cohere quote is customer testimony.
Groq 3 LPX determinism: PEP/CPS cut voltage droop >60% on Vera Rubin power stack
Sep 15, 2026 NVIDIA technical blog explains Groq 3 LPX deterministic execution enabling cycle-exact schedules across 256 LPUs, unlocking Preemptive Power (PEP) and Clock Period Synthesis (CPS) that internal testing says reduce voltage droop >60% and allow high-single-digit percent baseline voltage reduction (power ~V^2). Positions LPX with Vera Rubin NVL72 and DSX MaxLPS/Intelligent Power Smoothing; claims up to 35x throughput per megawatt vs GB200 NVL72 for 2T+ parameter models at long context and high interactivity; DSX MaxLPS cited for up to 40% more GPUs / 35% higher token throughput in the same site-power envelope. H2 2026 availability framing.
Why it matters, evidence and limitations
- Why it matters
- Newly relevant alongside Sep 16 MLPerf/power narratives: clarifies that extreme tokens-per-megawatt claims depend on specialized LPX racks and factory power software—not a default CUDA node upgrade for typical SLED rooms.
- Evidence and measured results
- Blog by Weidman/Vayalapra/Raghavan dated Sep 15, 2026 details deterministic schedule, PEP/CPS mechanisms, >60% droop reduction from internal testing, V^2 power relationship examples (5–10% voltage → ~8–19% power), and platform-level 35x perf/watt claim vs GB200 for specified operating point; notes DSX MaxLPS factory reclaim figures.
- Limitations and uncertainty
- Vendor engineering blog with internal testing; 35x is operating-point specific vs GB200 not GB300; H2 2026 timing; not a SLED deployment study. Included though dated Sep 15 because check_sources showed unknown and it newly explains power claims circulating with Sep 16 MLPerf coverage.
GenAI Daily: AgentX 30x Vera Rubin result is early silicon and coding-corpus only
Sep 17, 2026 GenAI Daily summarizes SemiAnalysis AgentX results highlighted at AI Infra Summit: Vera Rubin NVL72 up to 30x throughput per megawatt vs GB300 at highest interactivity on a 1.6T MoE, with ~2x at low interactivity and ~10x at moderate concurrency, plus a stated 45x cost-per-million-tokens reduction. Notes AgentX replays 393 real coding-agent sessions (median ~142k input vs ~444 output tokens). Caveats: early silicon with ongoing software optimization; coding-only corpus; Vera CPU tool-calling contribution not yet measured; methodology independence questions because NVIDIA hosts implementation docs while SemiAnalysis designed corpus/dashboard.
Why it matters, evidence and limitations
- Why it matters
- Essential counterweight to vendor 30x agentic efficiency slides circulating since Sep 15–16—SLED teams should not plan campus agent platforms on the peak concurrency point or assume non-coding agents match.
- Evidence and measured results
- Article dated Sep 17, 2026 reports corpus size/cost/license, concurrency-dependent advantage curve, Lambda MaxLPS 19-in-16/+24% aside, and explicit early-silicon/coding-only/independence caveats attributed to NVIDIA and SemiAnalysis acknowledgments plus reporter analysis.
- Limitations and uncertainty
- Secondary reporting of NVIDIA/SemiAnalysis materials; not a lab reproduction; cost-per-token reduction depends on unstated pricing assumptions.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.