Strategic PartnersIssue 168 records
NVIDIA
This edition pivots from yesterday’s Dynamo Enterprise/Flex-demo set to facilities qualification and claim diligence: NVIDIA’s Sep 21 DSX Ready program for BESS and CDUs (with an explicit site-stability caveat), independent SE Daily reporting that only five CDUs qualify at 2.5 MW class, DSX BESS self-qualification guidelines bounding review at PCS AC terminals, CoreWeave’s Sep 16 multi-rack Vera Rubin ops narrative (Racky/Valvey), SemiAnalysis’s non-verification of early CoreWeave Rubin metrics, NVIDIA’s AgentX up-to-30× per-MW vendor claim, Locsic’s coding-trace reading of AgentX, and an independent Manassas Aurora tracker still showing the 96 MW Flex-framed site under construction.
- Evidence records
- 8
- Cross-source patterns
- 3
- Evidence classes
- 3 vendor claim2 independent reporting2 independent research1 standards or public-body guidance
- Outcomes
- 4 emerging3 cautionary1 mixed
- Source freshness
- 4 new this fortnight3 recent1 older, newly relevant
- Research completed
- 2026-09-22
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
DSX Ready qualifies products—not site stability or interconnect
NVIDIA’s Sep 21 DSX Ready launch and BESS self-qualification guide both state that passing product qualification does not imply site-level stability or replace site engineering, while independent SE Daily reporting on LG’s 2.5 MW CDU shows the marketplace is still thin at high capacity—so SLED BOM screening can use Ready lists without treating them as campus certification.
Operating questionWhich site-level electrical/thermal acceptance tests remain mandatory after a DSX Ready product is selected?
Supporting evidenceNVIDIA launches DSX Ready to qualify BESS and CDU products for AI factoriesDSX BESS self-qualification: AC-terminal boundary and no site-stability implicationSeoul Economic Daily / LG Electronics
Multi-rack Vera Rubin capacity is ops- and facilities-defined
CoreWeave’s Sep 16 engineering narrative ties multi-rack Rubin bring-up to Racky/Valvey lifecycle automation for power and cooling, while the Aurora Manassas tracker still shows the much-cited 96 MW flexible AI factory under construction—reminding buyers that rack-scale success and Flex commercial claims depend on operated infrastructure, not GPU slides alone.
Operating questionDo we have rack-lifecycle automation and primary evidence the cited Flex site is actually operating before counting it as capacity?
Supporting evidenceCoreWeaveSue Data Centers (registry) / Digital Realty
Agentic Rubin efficiency claims need verification and workload scope
NVIDIA’s AgentX blog advertises up to 30× Rubin throughput per MW at a stated interactivity point, SemiAnalysis says early CoreWeave Rubin metrics remain unverified pending Q3 InferenceX submissions, and Locsic stresses AgentX is a real coding-trace exam—together warning SLED teams not to generalize preview ratios into unrelated service chat SLOs.
Operating questionIs the cited ratio independently verified, and does the buyer’s workload match AgentX coding-agentic scope at the same interactivity point?
Supporting evidenceNVIDIA technical blog: Vera Rubin up to 30× AgentX throughput per MW vs GB300SemiAnalysis / InferenceXLocsic
Evidence ledger
Full record · every source keeps its link and limitations
NVIDIA launches DSX Ready to qualify BESS and CDU products for AI factories
On Sep 21, 2026 NVIDIA launched DSX Ready, a qualification program for partner products that meet applicable DSX AI factory reference-design requirements, starting with battery energy storage systems (BESS) and cooling distribution units (CDUs). Initial qualified BESS names Hitachi Energy, LG Energy Solution and Tesla; initial CDUs name LG Electronics, LiquidStack and Vertiv. NVIDIA states category-specific requirements and review are meant to reduce integration risk, while explicitly warning that passing qualification does not replace site-level engineering or imply site-level stability. Additional infrastructure and software categories are planned over time.
Why it matters, evidence and limitations
- Why it matters
- Gives SLED facilities and procurement teams a vendor-curated shortlist language for power and liquid-cooling BOM alignment to DSX—but qualification is category-bounded and does not certify a campus/agency site, utility interconnection, or water budget.
- Evidence and measured results
- Program launch date Sep 21, 2026; two launch categories BESS and CDU; named BESS: Hitachi Energy, LG Energy Solution, Tesla; named CDUs: LG Electronics, LiquidStack, Vertiv; BESS partners run tests and submit data for NVIDIA review within a defined qualification boundary; site-level engineering/stability explicitly not implied.
- Limitations and uncertainty
- Vendor program announcement; named partners are examples at launch, not an exhaustive marketplace audit; no measured token or facility MW outcomes in this post.
Seoul Economic Daily: LG 2.5 MW CDU among five DSX Ready units at that class
Sep 22 Seoul Economic Daily reports LG Electronics won NVIDIA DSX Ready approval for a 2.5 MW coolant distribution unit, stating only five CDUs in the 2.5 MW class have that approval while more than 30 CDU products have DSX Ready status overall. The piece contextualizes ~120 kW per rack and ~2.4 MW for twenty racks as the load a single 2.5 MW CDU can handle, and notes LG’s prior 600 kW and 1 MW DSX Ready CDUs. It frames approval as improving competitive position for orders while quoting that actual orders still depend on NVIDIA’s decision.
Why it matters, evidence and limitations
- Why it matters
- Independent count of high-capacity DSX Ready CDUs helps SLED buyers judge how scarce 2.5 MW-class liquid-cooling options are—without treating Korean press supply-chain positioning as a measured campus deployment.
- Evidence and measured results
- LG 2.5 MW CDU DSX Ready approval announced Sep 22; >30 CDUs DSX Ready overall; only five at 2.5 MW class including LG; illustrative 120 kW/rack and 20 racks ≈ 2.4 MW; prior LG 600 kW and 1 MW DSX Ready CDUs; orders still contingent.
- Limitations and uncertainty
- Trade press relying on company statements and unnamed industry officials; not a third-party lab validation of CDU performance; order pipeline speculative.
DSX BESS self-qualification: AC-terminal boundary and no site-stability implication
NVIDIA’s BESS Self-Qualification Guidelines v1.0 (DA-12516-001_v01, May 28, 2026) define partner-run tests for AI load buffering, demand response, and low-voltage ride-through in grid-connected and islanded modes. The qualification boundary is explicitly at BESS AC terminals (PCS inclusive); site transformers, switchgear, generators, and campus controls are out of scope. The guide states passing qualification does not imply site-level stability, while failure across defined SCR/impedance/load envelopes is non-qualification. Newly relevant because Sep 21 DSX Ready makes these BESS packets the commercial entry path.
Why it matters, evidence and limitations
- Why it matters
- Why-now with today’s DSX Ready BESS launch: SLED energy/facilities teams can demand the guideline’s evidence packet and treat “DSX Ready BESS” as PCS-boundary performance—not campus microgrid certification.
- Evidence and measured results
- v1.0 May 28, 2026; objectives include DR, LVRT, islanded stability, black start, seamless transitions; boundary at AC terminals; excluded: site transformers/reactors/switchgear/relays/generators/campus controls; explicit non-implication of site-level stability; partners submit data for NVIDIA review.
- Limitations and uncertainty
- Vendor standard, not an independent lab certification program; document dated May 2026—relevance is via Sep 21 Ready commercialization, not a new v1.0 date.
CoreWeave: multi-rack Vera Rubin bring-up needs Racky/Valvey lifecycle automation
CoreWeave’s Sep 16, 2026 engineering blog explains that multi-rack NVIDIA Vera Rubin NVL72 (72 Rubin GPUs + 36 Vera CPUs per rack, NVLink 6, ConnectX-9, BlueField-4, 45°C liquid cooling) changes the problem from single-rack bring-up to a distributed system where hundreds of GPUs, multiple NVLink domains, networking, cooling, power, firmware, and software must behave as one. CoreWeave extends Mission Control with Racky (rack-management) and Valvey (programmable liquid-cooling control) beside a Rack LifeCycle Controller so power, cooling, and environmental sensing share the compute lifecycle. The post stresses automated staging, firmware, and outlier detection before production.
Why it matters, evidence and limitations
- Why it matters
- Shows SLED operators that Vera Rubin multi-rack capacity is gated by lifecycle automation for power/cooling/firmware—not by GPU count alone—useful when evaluating neocloud or on-prem rack-scale bids.
- Evidence and measured results
- Publish date Sep 16, 2026; per-rack 72 Rubin GPUs / 36 Vera CPUs / NVLink 6 / ConnectX-9 / BlueField-4 / 45°C liquid cooling; multi-rack “hundreds of Rubin GPUs”; Racky + Valvey + Rack LifeCycle Controller under Mission Control; automated BMC detection, firmware, staging validation.
- Limitations and uncertainty
- Operator-authored narrative; not an independent audit; performance numbers referenced via related posts rather than fully reproduced here.
SemiAnalysis InferenceX: CoreWeave early Rubin metrics not independently verified
SemiAnalysis’ July 23, 2026 InferenceX deep dive on Vera Rubin NVL72 vs GB200 NVL72 states that early VR NVL72 metrics come from CoreWeave and that SemiAnalysis has not independently verified them, while noting NVIDIA committed to submitting verifiable numbers to InferenceX by Q3 CY2026. The article analyzes architecture (including Rubin LUT-based tensor-core features), performance-per-megawatt and performance-per-TCO framings, and software-stack maturity, cautioning that leads versus Blackwell depend on baseline and operating point. Newly relevant amid Sep 2026 multi-rack/AgentX marketing that still rests on early partner measurements.
Why it matters, evidence and limitations
- Why it matters
- Supplies the independent caveat SLED buyers need when neocloud or vendor decks cite early Rubin tokens/MW: treat CoreWeave-origin figures as unverified until InferenceX/MLPerf-class submissions land.
- Evidence and measured results
- Originally published July 23, 2026; explicit “We have not independently verified them”; NVIDIA verifiable-numbers commitment by Q3 CY2026; architecture and TCO/perf-per-MW analysis; software runway discussion.
- Limitations and uncertainty
- Analyst publication with subscription/context; Q3 verification commitment may slip; not a lab report reproducing CoreWeave runs.
NVIDIA technical blog: Vera Rubin up to 30× AgentX throughput per MW vs GB300
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, evidence and limitations
- 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.
Locsic: AgentX is a real coding-trace exam—CUDA moat now software-priority speed
Locsic’s Aug 25, 2026 dissection of SemiAnalysis AgentX 1.0 describes the benchmark as built from a costly real Claude Code trace corpus, with 393 anonymized sessions open-sourced, and measurements across 1,000+ chips. It treats AgentX as a serious coding-focused, long-context multi-turn exam rather than a synthetic chatbot test, while arguing the CUDA “moat” in agentic inferencing is increasingly about how fast new workloads become software priorities (and noting AMD gaps/upstream paths). Newly relevant as an independent read on why AgentX headline ratios should not be generalized beyond coding-agentic inference.
Why it matters, evidence and limitations
- Why it matters
- Helps SLED teams interpret AgentX: credible for developer-agent tooling evaluations, weak as a proxy for citizen-facing or case-management chat without domain traces.
- Evidence and measured results
- Date 2026-08-25; $3M trace capture claim; 393 open-sourced sessions; 1,000+ chips measured; AgentX positioned as real-work coding sessions; CUDA moat reframed toward software-priority speed; follow-up to prior 30×/35× discussion.
- Limitations and uncertainty
- Independent essay interpreting SemiAnalysis; some paywalled SemiAnalysis conclusions noted as incomplete in free text; not a lab reproduction.
Sue Data Centers: Aurora Manassas 96 MW AI factory still under construction
Sue Data Centers’ Aurora AI Factory registry page (record updated Sep 9, 2026; last checked Sep 14, 2026) describes Digital Realty’s Manassas, Virginia site as under construction with 96 MW planned power, announced Oct 28, 2025, and cites DCD reporting that operations were slated for first half of 2026. It records AI use as confirmed, notes power-flexible/grid-connected framing with Emerald AI Conductor and a flexible-power certification developed with NVIDIA, Emerald AI, EPRI and PJM, and states no reception on record. Newly relevant because recent NVIDIA Flex narratives call Manassas the first dedicated DSX Flex commercial deployment while public trackers still show under construction.
Why it matters, evidence and limitations
- Why it matters
- Independent status check against vendor “first commercial DSX Flex” language: SLED planners should treat Aurora as a planned 96 MW flexible-load site still under construction—not an operating measured Flex reference today.
- Evidence and measured results
- 96 MW planned; under construction; operator Digital Realty; announced Oct 28, 2025; record updated Sep 9, 2026; last checked Sep 14, 2026; H1 2026 ops cited via DCD Oct 30, 2025; Emerald Conductor / flexible certification notes; no reception on record.
- Limitations and uncertainty
- Registry synthesizes public sources; some fields explicitly “not established”; advocacy-oriented site—cross-check primary permits/utility docs for legal reliance.
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.
- Independent research
- Research conducted outside the implementing organization.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.