Strategic PartnersIssue 188 records
NVIDIA
This edition pivots from yesterday’s Dynamo 1.5.0 GA / ESS News BESS triad / Flex-API-vs-Utility-Dive set toward three operator threads: (1) Singapore public-safety evidence—Straits Times Sep 23 HTX first-in-country GB300 operationalisation inside NGINE paired with NVIDIA’s Sep 22 AI Day note that HTX is researching Nemotron 3 Super/Omni under locally controlled deployment; (2) DSX Ready cooling-side diligence—Vertiv’s 2.3 MW CoolChip as first Ready CDU with an explicit site-engineering caveat, plus SiliconReport’s compact BESS+CDU supplier lists restating no site-stability guarantee; (3) Flex/AEMA independent scrutiny now that The Register’s Sep 16 DSX “walled garden”/DR-is-old critique is accessible (prior 403) alongside ts2’s Sep 17 warning that AEMA’s 100 GW is a scenario and 130 kW trials ≠ GW campuses. Supporting constraints: Dynamo Enterprise NGC artifacts remain pinned at 1.4.2 after OSS 1.5.0 GA, and NVIDIA’s Sep 22 CC blog claims 96.1–98.2% DeepSeek-R1 throughput retention on 8×B200 with TRT-LLM adaptations—vendor-scoped against prior independent CC-naive penalties. Patterns: sovereign public-safety stacks are ops-defined; Ready lists ≠ site cert; Flex/AEMA need contracts and are stack-gated.
- Evidence records
- 8
- Cross-source patterns
- 3
- Evidence classes
- 4 independent reporting4 vendor claim
- Outcomes
- 4 emerging4 cautionary
- Source freshness
- 6 new this fortnight2 undated
- Research completed
- 2026-09-24
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Sovereign public-safety GB300/Nemotron paths are agency- and ops-defined
Straits Times’ Sep 23 independent report that HTX operationalised GB300 inside NGINE for multimodal/embodied public-safety AI, paired with NVIDIA’s Sep 22 AI Day blog that HTX is researching Nemotron 3 Super/Omni under locally controlled deployment, shows hardware and model narratives converging on agency-owned control planes—not GPU slides alone. Transferable as a diligence pattern for U.S. SLED public-safety AI factories, not as a measured U.S. outcome.
Operating questionWho owns the control plane, data residency boundary, and workload acceptance tests before counting GB300/Nemotron as an operational public-safety capability?
Supporting evidenceThe Straits Times / HTX (Singapore Home Team)NVIDIA AI Day Singapore: HTX researching Nemotron 3 Super/Omni for public safety
DSX Ready CDU/BESS lists are BOM screens—not site certification
Vertiv’s Sep 21 announcement that its 2.3 MW CoolChip is the first DSX Ready CDU explicitly states qualification does not replace site/configuration engineering, while SiliconReport’s program summary lists BESS (Hitachi/LGES/Tesla) and CDU (LG/LiquidStack/Vertiv) suppliers and restates that Ready does not guarantee site-level stability. SLED teams can screen BOMs with Ready lists without treating logos as campus energization approval.
Operating questionWhich site thermal/electrical studies and facility-fit sign-offs remain mandatory after a DSX Ready CDU or BESS SKU is selected?
Supporting evidenceVertivSiliconReport
Flex/AEMA are DR tools needing contracts—and DSX is stack-gated
The Register’s newly accessible Sep 16 analysis frames DSX Flex/MaxLPS as commercially motivated reclaim of stranded power and “another walled garden,” while ts2’s Sep 17 AEMA analysis stresses that 100 GW is a planning scenario, 130 kW trials are not GW proof, and advocacy cannot award interconnection terms without baselines, telemetry, and penalties. Software demos and coalitions do not replace utility agreements or multi-vendor power-control plans.
Operating questionWhat audited curtailment terms and stack scope (including non-NVIDIA accelerators) are written before Flex/AEMA language enters an interconnect or capacity plan?
Supporting evidenceThe Registerts2.space (Marcin Frąckiewicz)
Evidence ledger
Full record · every source keeps its link and limitations
Straits Times: HTX first in Singapore to operationalise NVIDIA GB300 for public-safety AI
On Sep 23, 2026 The Straits Times reported that Singapore’s HTX is the first organisation in Singapore to deploy NVIDIA’s GB300 (commercially available H1 2026) inside NGINE, its sovereign GPU AI infrastructure, to train larger multimodal and embodied-AI models and run inference at scale for public-safety robotics. HTX CAIO Ang Chee Wee, speaking at NVIDIA AI Day Singapore, stressed code/data remaining under local control. Existing HTX AI uses include immigration facial recognition, Report Lodging Co-Pilot (R-Cop) police-report drafting, and scam-site takedown, with “close to a million man-hours” saved annually. ~S$100M is committed to a late-September Home Team Humanoid Robotics Centre (H2RC). Sea Limited separately announced ASEAN-first Vera Rubin adoption plans—commercial, not HTX.
Why it matters, evidence and limitations
- Why it matters
- Provides a dated public-safety agency operationalisation signal for Blackwell-class GB300 under a sovereign-infra narrative—useful diligence for U.S. SLED public-safety/homegrown AI factory conversations, but Singapore-specific, quote-driven, and without published rack counts, SLOs, or independent performance audits transferable as U.S. measured outcomes.
- Evidence and measured results
- First-in-Singapore GB300 operationalisation claim (Ang Chee Wee, AI Day Sep 23); NGINE multimodal+embodied training/inference; sovereign code/data control quote; existing AI uses: immigration facial recognition, R-Cop report drafting, scam-site takedown; ~1M man-hours/year saved claim; H2RC ~S$100M late Sep; Sea Vera Rubin ASEAN-first commercial mention.
- Limitations and uncertainty
- Independent press summarizing official remarks; no published GB300 cluster size, tokens/s, power, or independent audit; man-hours savings predate GB300; Vera Rubin Sea note is commercial roadmap.
NVIDIA AI Day Singapore: HTX researching Nemotron 3 Super/Omni for public safety
NVIDIA’s Sep 22, 2026 AI Day Singapore blog states HTX is embarking on research using Nemotron 3 Super (complex reasoning/agentic workflows) and Nemotron 3 Nano Omni (unified vision/audio/language) to advance public-safety AI under locally controlled deployment. The post frames four Southeast Asia focus areas (government operations, citizen services, critical infrastructure/public safety, startups/research) and lists regional Nemotron adopters (NCS/VSS, ST Engineering NeMo, YTL AI Labs, ThaiLLM/iApp ~43k-user Thanoy chatbot, AIS SEA-LION, Viettel fine-tunes, FPT personas). Sea is described as ASEAN-first Vera Rubin adopter for agentic services across Garena/Monee/Shopee—commercial roadmap language.
Why it matters, evidence and limitations
- Why it matters
- Pairs with independent Sep 23 HTX GB300 reporting to show NVIDIA’s model-layer narrative (Nemotron Super/Omni research) alongside agency hardware operationalisation—useful for SLED public-safety solution mapping, but vendor-authored, research-not-production for HTX models, and regional commercial claims are not U.S. measured outcomes.
- Evidence and measured results
- HTX researching Nemotron 3 Super + Nano Omni for public safety; locally controlled AI framing; four SEA focus areas; NCS VSS blueprint; ST Engineering NeMo; Thai iApp Thanoy ~43k users; Viettel Nemotron 3 Super VMLU ranking claim; Sea ASEAN-first Vera Rubin adoption plans.
- Limitations and uncertainty
- Vendor blog; HTX described as “embarking on research,” not completed production model deployment; regional partner claims lightly evidenced; Sea Rubin is forward-looking.
Vertiv 2.3 MW CoolChip CDU qualifies as first NVIDIA DSX Ready CDU
On Sep 21, 2026 Vertiv announced its 2.3 MW CoolChip liquid-to-liquid CDU qualified as an NVIDIA DSX Ready CDU—described as the first CDU to qualify specifically under the program and among initial CDU providers. The unit supports direct-to-chip and rear-door heat-exchanger applications for end-of-row or perimeter install. Vertiv states DSX Ready uses NVIDIA’s CDU self-qualification suite for functional requirements, and explicitly: qualification does not replace site/configuration engineering; customers must still determine facility fit, system configuration, and operating requirements.
Why it matters, evidence and limitations
- Why it matters
- Extends yesterday’s BESS-heavy Ready diligence to the cooling side with a named 2.3 MW SKU and an on-the-record site-engineering caveat—useful for SLED BOM screening of liquid-cooled AI rooms without treating Ready as campus certification.
- Evidence and measured results
- 2.3 MW CoolChip CDU; first CDU specifically DSX Ready; liquid-to-liquid; direct-to-chip + rear-door HX; end-of-row/perimeter; CDU self-qualification suite; explicit non-replacement of site engineering / facility-fit determination.
- Limitations and uncertainty
- Vendor PR; no third-party re-test published here; “first CDU” is Vertiv’s characterization; MW rating is product capacity, not a measured campus deployment.
SiliconReport: DSX Ready launch lists BESS (Hitachi/LGES/Tesla) and CDUs (LG/LiquidStack/Vertiv)
SiliconReport summarizes NVIDIA’s DSX Ready program launch covering BESS and CDU categories. It reports initial BESS suppliers Hitachi Energy, LG Energy Solution, and Tesla, and CDU suppliers LG Electronics, LiquidStack, and Vertiv. It restates NVIDIA’s caution that BESS qualification operates within a defined boundary without replacing site-level engineering or guaranteeing site-level stability, notes partner self-tests for BESS versus a CDU self-qualification suite, and concludes operators must still independently validate site-level engineering and facility fit.
Why it matters, evidence and limitations
- Why it matters
- Independent compact listing of both Ready categories and the site-stability caveat—pairs with Vertiv’s CDU PR to keep SLED BOM diligence from collapsing “Ready” into campus certification, and complements (without repeating) yesterday’s ESS News BESS-only deep dive.
- Evidence and measured results
- Categories: BESS + CDU; BESS: Hitachi Energy, LGES, Tesla; CDU: LG Electronics, LiquidStack, Vertiv; BESS self-test+NVIDIA review; CDU self-qual suite; site-level engineering/stability caveat restated; additional infra/software categories planned.
- Limitations and uncertainty
- Secondary trade summary citing NVIDIA; undated page; not independent lab re-tests; overlaps thematically with prior ESS News but adds CDU triad and compact caveat.
The Register: DSX Flex/MaxLPS reclaim stranded power—but demand response is old and DSX is a walled garden
Tobias Mann’s Sep 16, 2026 Register analysis—newly accessible after prior HTTP 403—argues NVIDIA’s DSX addresses stranded datacenter power that limits GPU sales. DSX MaxLPS and DSX Exchange APIs aim to coordinate compute with DSX Ready assets (e.g., Vertiv, Schneider, BMS); Lambda’s 19-vs-16 node densification (+24% throughput) is recounted. DSX Flex with Emerald AI and Silicon Valley Power is framed as demand response that utilities may use to approve more capacity if curtailment is credible. Critical workloads continue while non-essential work pauses or migrates. Competing accelerators (e.g., AMD Instinct) likely need alternative management stacks; NVLink Fusion partners may need extra software integration. Closing line: DSX is “another walled garden” keeping customers on NVIDIA equipment.
Why it matters, evidence and limitations
- Why it matters
- Supplies the independent Flex/MaxLPS critique that was blocked in the 2026-09-22 run—essential counterweight to vendor Flex demos and yesterday’s Utility Dive contract theme: software flexibility can be commercially motivated, DR-familiar, and stack-gated.
- Evidence and measured results
- Stranded ~20% headroom framing; DSX Exchange API across Ready providers/BMS; Lambda 19-in-16 / +24%; Flex+Emerald+SVP curtailment narrative; DR “isn’t exactly new”; utility less-conservative allocation quote (Dion Harris); pause/migrate non-essential work; AMD path needs alternatives; “another walled garden” conclusion.
- Limitations and uncertainty
- Opinionated tech press; Lambda figures vendor-origin; interview-based; dated Sep 16 but newly retrievable after prior 403 block (why-now for this edition).
ts2 analysis: AEMA flexibility needs contracts; 100 GW is a scenario; 130 kW trials ≠ GW campuses
Sep 17, 2026 analysis of AEMA (Emerald AI/Google/NVIDIA, launched Sep 16) argues flexible connections only work if curtailment is dependable under the conditions that blocked firm service. It frames AEMA as extending demand response, cites Google’s 1 GW contracted DR (not a single campus shed), and stresses AEMA’s 100 GW headline is a Duke University planning scenario (≈98 GW at 0.5% average curtailment across 22 balancing authorities)—not spare local electricity. Strongest technical evidence cited: Emerald/EPRI/National Grid/Nebius five-day 130 kW London trial (22 events, <1 min response, up to 40% reductions) — four orders of magnitude below 1 GW. Contracts need baselines, auditable telemetry, event limits, and penalties; Latitude Media is cited that AEMA repurposes a largely inactive 2014 alliance and is advocacy—not a regulator.
Why it matters, evidence and limitations
- Why it matters
- Independent policy/engineering diligence that pairs with Register’s walled-garden Flex critique and yesterday’s Utility Dive binding-agreement theme: SLED interconnect strategies need tariff language and witnessed MW/time/SLO evidence, not coalition membership or national GW scenarios.
- Evidence and measured results
- AEMA launch Sep 16; DR extension framing; Google 1 GW contracted DR across utilities; Duke ~98 GW @ 0.5% curtailment / 2.1h avg event; 130 kW Nebius London trial details; EPRI DCFlex ≥10 demos planned; contract elements (baseline/telemetry/limits/penalties); Latitude advocacy characterization; FERC large-load tariff review context.
- Limitations and uncertainty
- Secondary analysis synthesizing AEMA/Duke/Emerald/Latitude/FERC; not a utility tariff text; quantitative claims depend on cited studies’ assumptions.
Dynamo Enterprise NGC artifacts still pinned to 1.4.2 after OSS Dynamo v1.5.0 GA
As of this edition’s retrieval, the NVIDIA Dynamo Enterprise NGC collection—artifacts eligible for NVIDIA Enterprise Support, stated to have no functional/binary differences from corresponding OSS—still lists only version 1.4.2 for Dynamo Frontend, kubernetes-operator, SGLang/TensorRT-LLM/vLLM runtimes, and the Dynamo Platform Helm chart, each modified 08/29/2026 UTC. No 1.5.0 Enterprise artifacts appear despite Dynamo v1.5.0 OSS GA on Sep 18, 2026 (archived prior edition). Why-now: after last night’s Dynamo 1.5.0 GA coverage, supportability lag is the operator-facing constraint.
Why it matters, evidence and limitations
- Why it matters
- Directly informs SLED change-control: campuses on AI Enterprise should not assume 1.5.0 GA features are support-eligible until -enterprise artifacts and matrix pins catch up.
- Evidence and measured results
- Enterprise collection description (support-eligible; no functional/binary diff vs OSS); containers: Frontend/k8s-operator/SGLang/TRT-LLM/vLLM all 1.4.2 modified 08/29/2026 UTC; Helm Dynamo Platform (Enterprise) 1.4.2 same date; no 1.5.x listed at retrieval.
- Limitations and uncertainty
- Catalog snapshot at retrieval time; Enterprise 1.5.0 may appear later without blog fanfare; page undated beyond artifact mtimes; does not itself measure production reliability.
NVIDIA: Confidential DeepSeek-R1 on 8×B200 retains 96.1–98.2% throughput with TRT-LLM CC adaptations
NVIDIA’s Sep 22, 2026 technical blog reports production-oriented confidential inference on Blackwell using confidential VMs, confidential GPUs, and encrypted NVLink. With TensorRT-LLM CC-aware adaptations (pageable memory instead of pinned due to encrypted bounce-buffer H2D path; %globaltimer autotune; NVLS-multicast-unavailable algorithm selection), a controlled comparison on eight B200 GPUs serving DeepSeek-R1 retained 96.1–98.2% of CC-off output-token throughput and added 1.2–4.3% mean TPOT overhead across concurrency 1–16. The post stresses CC does not remove performance engineering and urges workload-specific CC-on vs CC-off benchmarks.
Why it matters, evidence and limitations
- Why it matters
- Fresh vendor counter-measurement to prior archived independent Serialized Bridge / Tinfoil findings of much larger naive CC penalties—SLED regulated deployments should treat 96%+ as TRT-LLM-adapted DeepSeek-R1 on 8×B200, not universal CC overhead, and still re-bench their engine.
- Evidence and measured results
- Date Sep 22, 2026; 8×B200; DeepSeek-R1; CC-on retains 96.1–98.2% output-token throughput; +1.2–4.3% TPOT; concurrency 1–16; bounce-buffer → pageable memory; globaltimer autotune; NVLS multicast unavailable under B200 CC; call to benchmark exact workload.
- Limitations and uncertainty
- Vendor-controlled benchmark; single model/engine/hardware card; prior independent literature showed much higher penalties on CC-naive stacks—figures are not contradictory if scopes differ; no multi-tenant noisy-neighbor test.
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.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.