Strategic PartnersIssue 148 records
NVIDIA
This edition pivots from yesterday MaxLPS PoC/Flex densification set to newly archived operator, policy, and measurement evidence: Pinterest Sep 10 Dynamo+B200 VLM production stack (projection embeddings, LMCache, multimodal KV routing), Data Center Frontier Sep 10 DSX×Lancium/Cloverleaf/Trane/Eaton infrastructure partnerships with powered-land≠installed-capacity caveats, Clarity AEMA ledger requiring curtailment terms before interconnect amid thin demonstrated depth, arXiv controlled Blackwell confidential-computing benches (~1–3% configured vs ~30–40% naive), official MaxLPS overview taxonomy (MaxP/MaxQ/MaxLPS; illustrative 1.35x), AI Enterprise Infra 8.2 support-matrix Government Ready pins, SemiAnalysis InferenceX Sep 14 Rubin AgentX tok/s/MW results with MaxLPS non-isolation, and Sep 18 DataCenterNews efficiency-deals coverage spanning Annapurna, d-Matrix, Pinterest, and Groq.
- Evidence records
- 8
- Cross-source patterns
- 3
- Evidence classes
- 3 independent reporting2 independent research2 vendor claim1 academic research
- Outcomes
- 3 emerging2 effective2 mixed1 cautionary
- Source freshness
- 5 new this fortnight2 undated1 recent
- Research completed
- 2026-09-20
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Production Dynamo multimodal gains are payload- and KV-architecture gated
Pinterest Sep 10 engineering post shows Dynamo+B200 VLM serving delivering large TTFT/E2E wins when projection embeddings, LMCache tiers, and multimodal KV-aware routing are in place, while InferenceX Sep 14 AgentX read shows Rubin tok/s/MW multipliers that swing dramatically by engine and P90 interactivity—so multimodal/agentic efficiency is an architecture and SLA pin, not a GPU-only slide.
Operating questionWhich payload contract (pixel vs embedding), KV tier, engine, and P90 interactivity were measured before citing Dynamo or Rubin efficiency for a multimodal assistant?
Supporting evidencePinterest EngineeringSemiAnalysis / InferenceXDataCenterNews
DSX site/power partnerships are pipeline and projection-heavy—not deployed capacity
Data Center Frontier separates Lancium >15 GW powered-land pipeline from installed NVIDIA IT load and labels Trane/Eaton efficiency/copper figures as projections, while NVIDIA MaxLPS overview presents illustrative 400 GPU/MW and 1.35x token graphics with Rubin testing still in progress—and Sep 18 DataCenterNews repeats Flex/MaxLPS partner claims without converting them into energized megawatts.
Operating questionHow many megawatts are leased, energized, and instrumented for MaxLPS/Flex before partner GW or overview graphics enter a capacity plan?
Supporting evidenceData Center FrontierNVIDIA DSX MaxLPS overview: MaxP vs MaxQ vs MaxLPS and illustrative 1.35x tokensDataCenterNews
Flexible-load interconnect claims need utility-adopted curtailment metrics
Clarity AEMA ledger requires ride-through/curtailment/contingency terms before connection and finds only a 256-GPU 25%-for-3-hours demo plus contracted—not emergency—1 GW demand response, with no named utility adopter; DataCenterNews Sep 18 Flex recount (hundreds of utility signals) remains operator/vendor reported—so SLED interconnect strategies need witnessed MW/time/SLO evidence, not coalition membership.
Operating questionWhat curtailment depth, response time, and duration are written into a utility tariff or interconnect agreement—and witnessed on your mix?
Supporting evidenceThe ClarityDataCenterNewsData Center Frontier
Evidence ledger
Full record · every source keeps its link and limitations
Pinterest builds production VLM serving on Dynamo + Blackwell B200 with projection embeddings
Pinterest Engineering (Sep 10, 2026) details a production VLM stack on NVIDIA Blackwell B200 GPUs and NVIDIA Dynamo (DynamoGraphDeployment on EKS), with vLLM, LMCache KV offload, encoder/prefill/decode disaggregation, and multimodal KV-aware routing. Custom PinCLIP projection embeddings avoid pixel-path vision encoding: vs pixel inputs they report ~85x average TTFT, ~7.3x E2E latency, and ~1.1x TPOT speedups (peaks ~369x TTFT / ~44x E2E), enabling ~250-image contexts with latency comparable to ~10-image pixel requests (~25x visual context). B200 showed >2x latency improvement vs Hopper in preliminary Assistant benchmarks. AIPerf DAG benchmarks mirror multi-turn multimodal traffic.
Why it matters, evidence and limitations
- Why it matters
- Strong transfer for SLED multimodal assistants (image+text) on managed Kubernetes: shows Dynamo is an operator stack requiring payload design, KV tiers, and disaggregation—not a drop-in NIM. Limited direct public-sector deployment evidence; relevance is architectural pattern and measured serving levers.
- Evidence and measured results
- Dynamo operator via Helm; DynamoGraphDeployment CRDs; OpenAI Chat Completions + Model Router; LMCache GPU/CPU/NVMe tiers; multimodal frontend image decode; PinCLIP projector path; TTFT p99 up to 5.92x with MM KV routing (~1.42x average); LoRA hot-loading; AIPerf DAG for agentic workflows; use cases: Assistant, rerank, OCR, safety.
- Limitations and uncertainty
- Single large operator; metrics are Pinterest-internal benchmarks on Assistant/VLM traffic; not a SLED agency deployment; embedding path depends on proprietary PinCLIP.
Data Center Frontier: DSX extends into Lancium/Cloverleaf/Trane/Eaton—pipeline ≠ deployed GW
Sep 10, 2026 Data Center Frontier reports NVIDIA pushing DSX beyond GPUs into site/power/cooling partnerships: Trane+Eaton DSX-aligned medium-voltage power/cooling reference (vendor claims up to 15% efficiency, 30% install-cost, 80% copper reductions), Cloverleaf minority investment moving DSX into pre-construction site design, and Lancium partnership (>15 GW powered-land pipeline; 4 GW under lease) with MaxLPS/Flex and strategic investment. Piece explicitly separates powered-land pipeline from installed NVIDIA capacity and labels Eaton/Trane savings as projections, not fleet measurements; also notes IREN/Australia GW announcements and financing platforms as capital-side extensions.
Why it matters, evidence and limitations
- Why it matters
- Useful caution for SLED facilities and economic-development AI campus talks: DSX GW headlines are mostly land/power pipeline and reference designs. Direct SLED transfer is limited to evaluation checklists (MaxLPS-ready design, Flex demonstrability) rather than citing partner GW figures as available campus capacity.
- Evidence and measured results
- Trane/Eaton Aug 17 reference design claims; Cloverleaf Aug 21 minority investment; Lancium Aug 24 >15 GW pipeline / 4 GW leased; MaxLPS up-to-40% Rubin + 45C cooling reiterated; Emerald AI/SVP Flex pilot cited; IREN Sweetwater; Australia up to 2 GW by 2027; financing platforms named.
- Limitations and uncertainty
- Trade press synthesizing vendor announcements; efficiency/copper figures are vendor projections; Lancium 15 GW is development pipeline.
Clarity: AEMA wants curtailment terms fixed before interconnect—evidence still thin
Clarity AEMA cluster synthesizes NVIDIA/Emerald/Google launch coverage and theneuron.ai scrutiny: principles call for ride-through, curtailment, and contingency-response obligations before a facility connects, plus standardized metrics (response speed, duration, predictability, emergency behavior). Measured evidence cited remains a 256-GPU, 25% power cut for three hours with SLOs held, and Google 1 GW demand-response as contracted capacity—not a single emergency reduction. No utility/grid operator is named as having adopted AEMA principles; AEMA is described as a reboot of a 2014 demand-response group. Household bill relief is treated as a separate arrangement from faster interconnection for operators.
Why it matters, evidence and limitations
- Why it matters
- Directly relevant to SLED facilities seeking faster interconnect via flexible AI load narratives: forces explicit curtailment depth/time commitments and warns against treating vendor demos or contracted GW as proven emergency performance. Limited relevance to software procurement alone.
- Evidence and measured results
- Claim ledger items on principles, 25%/3h/256-GPU test, Google 1 GW contracted DR, no named utility adopter, first-of-its-kind vs 2014 reboot, technology-neutral mechanisms (shift/storage/generation/contingency).
- Limitations and uncertainty
- Secondary synthesis of Axios/Heatmap/theneuron/NVIDIA; derived 4 GW sizing check labeled insufficient; adoption score low by Clarity own rubric.
arXiv: Benchmarking Confidential Computing on Blackwell—~1–3% when configured, 30–40% when not
Paired CC-on vs CC-off measurements on one Xeon + 8x B200 host under Intel TDX + NVIDIA Confidential Computing find well-configured confidential inference at ~1–3% throughput tax (e.g., MoE TP4 ~1.5%, TP8 ~3%, single-GPU with full CUDA graphs + patches <1%), while stock/eager stacks see ~30–40% penalties from avoidable host-submission and bounce-buffer costs. GPU compute, energy draw, and usable memory largely unaffected (fixed ~700 MB carveout). Training on 8 GPUs costs ~10–13% via encrypted collectives. Authors localize cost to encrypted PCIe/command-path and NVLE; cross-node RDMA/disagg not measured.
Why it matters, evidence and limitations
- Why it matters
- Highly relevant for SLED regulated inference: updates the prior Tinfoil/NVIDIA CC debate with controlled Blackwell methods. Shows CC can be near-parity if frameworks are CC-aware—procurement must specify engine/build/concurrency, not a single overhead percentage.
- Evidence and measured results
- Controlled same-host comparisons; MiniMax-M2.7 MoE medians; vLLM piecewise-graph ~185 submissions predicting ~2.28 ms/step overhead; eager vs full-graph probe 2.41x to 1.04x; NVLE collective tax; training 10–13%; limitations: no security audit, no cross-node measurement, attestation startup not timed.
- Limitations and uncertainty
- Single-host lab; framework versions dated; cross-node disagg out of scope; authors affiliated with Confidential.ai.
NVIDIA DSX MaxLPS overview: MaxP vs MaxQ vs MaxLPS and illustrative 1.35x tokens
Official DSX MaxLPS overview defines MaxLPS as coordinating facilities/site design, Dynamic Power Software (DPS), and advanced perf/W techniques to maximize throughput per megawatt in a fixed site-power envelope, claiming up to 40% more GPUs in that envelope. It contrasts static MaxP/MaxQ GPU settings with factory-wide dynamic MaxLPS, and shows an illustrative Vera Rubin NVL72 inference scenario of 400 GPUs per 1 MW and 1.35x token throughput vs MaxP (PUE 1.1), while stating Vera Rubin testing remains in progress. Emphasizes planning for 45C cooling even when Day-1 training mixes offer less headroom than inference.
Why it matters, evidence and limitations
- Why it matters
- Clarifies product taxonomy for SLED facilities evaluating MaxLPS after Sep 18 Lambda PoC coverage: distinguishes marketing densification from design prerequisites (45C, lifecycle fill). Newly archived docs page; treat illustrative 1.35x/400 GPU/MW as planning graphics, not measured customer results.
- Evidence and measured results
- MaxP/MaxQ/MaxLPS mode comparison; up-to-40% more GPUs claim; 45C cooling callout; illustrative 1 MW scenario 400 GPUs / 1.35x tokens vs MaxP; note Vera Rubin testing in progress; stack diagram across facilities/hardware/infra software/platform software.
- Limitations and uncertainty
- Vendor docs; undated page; illustrative scenario; Rubin testing in progress disclaimer on the same page.
AI Enterprise Infra 8.2 support matrix pins operators and Government Ready columns
The version-pinned AI Enterprise Infra 8.2 support matrix lists component versions and a Government Ready column: Data Center GPU Driver 595.91.07, Fabric Manager 595.91.07, GPU Operator 26.3.3 (Supported for Gov Ready with footnote to operand list; ARM N/A for Gov Ready), Network Operator 26.4.1, NIM Operator 3.1.2 (Supported across x86/ARM/Gov Ready), DOCA networking 3.4.0 Supported for Gov Ready, Run:ai 2.26, BCM 11.33.1. Platform tables include HGX/DGX B200/B300 and GB200/GB300 NVL72 bare-metal notes, OpenShift 4.18–4.22, Kubernetes 1.32–1.36, and cloud instance matrices. Footnotes restrict some HGX vGPU/KubeVirt combinations.
Why it matters, evidence and limitations
- Why it matters
- Actionable for SLED regulated Kubernetes baselines: yesterday 8.2 release notes are complemented by the matrix Government Ready flags and orchestration/OS pins needed for ATO-oriented stacks. Does not itself confer FedRAMP authorization.
- Evidence and measured results
- Table 4 infrastructure software with Gov Ready column; Kubernetes bare-metal and virtualized tables; GB200/GB300 bare-metal-only notes; footnotes pointing to Gov Ready operand docs; vSphere HGX B200/B300 1:1 and multi-vGPU only (no fractional).
- Limitations and uncertainty
- Vendor matrix; undated page body; Gov Ready footnotes defer to other docs; some cells N/A.
SemiAnalysis InferenceX: early Rubin AgentX shows large tok/s/MW gains; MaxLPS not isolated
Sep 14, 2026 InferenceX port reports first verified AgentX agentic-inference results for Vera Rubin NVL72 on early pre-release software: large TCO and tok/s/MW advantages vs GB300 depending on interactivity target and engine (examples include ~67x throughput per owning-TCO vs GB300 Dynamo TRTLLM at 170 TPS on a dense NVFP4 path; ~7.2x tok/s/MW vs GB300 SGLang at 150 TPS; engine label critical—at 170 TPS vs GB300 TRTLLM the MW multiplier is ~62.9x but vs SGLang ~5.56x). Authors argue GTC 3x per MW sandbagged vs measured up to ~7x. MaxLPS is described as first-class dynamic power shifting to fit more GPUs from real workload power profiles, with PowerX integration forthcoming—AgentX curves are not presented as isolated MaxLPS measurements.
Why it matters, evidence and limitations
- Why it matters
- Independent counterweight to vendor AgentX/MaxLPS marketing for SLED capacity planning: multipliers are SLA- and engine-scoped, early-software, and not MaxLPS-isolated. Limited direct public-sector deployment; relevance is procurement literacy for tokens/MW slides.
- Evidence and measured results
- AgentX real-world coding trajectories; owning vs 3-year rental TCO frames; tok/s/MW table across TPS targets; MaxLPS qualitative description; early TRTLLM Rubin software caveat; subscriber-only continuation noted.
- Limitations and uncertainty
- Public port of newsletter; early pre-release software; NVIDIA collaboration acknowledged; some fleet economics are modeled; MaxLPS not separately ablated.
DataCenterNews Sep 18: power-efficiency deals span Flex, MaxLPS, Annapurna, d-Matrix, Pinterest
Sep 18, 2026 DataCenterNews recounts Ian Buck AI Infra Summit framing around tokens per megawatt and catalogs efficiency-related collaborations: Emerald AI + Silicon Valley Power flexible-load programme responding to hundreds of utility signals; Lambda MaxLPS PoC (+24% tokens/s, +23% perf/W, 19 vs 16 nodes); Amazon Annapurna Labs NVHBM custom HBM work; d-Matrix NVLink Fusion with Vera CPUs + Raptor XPUs; Pinterest Blackwell + Dynamo for conversational visual discovery; Vera Rubin + Groq 3 LPX claims up to 35x tokens/MW vs GB200 on long-context; AgentX up to 30x tok/MW and 45x lower cost per million tokens vs GB300 on DeepSeek V4 Pro; NVLink 6 reliability features for large factories.
Why it matters, evidence and limitations
- Why it matters
- Same-day complementary reporting to yesterday Flex/MaxLPS piece with additional ecosystem partners (Annapurna, d-Matrix, Pinterest, Groq). Useful as a map of claims for SLED scanners; still vendor-sourced metrics via trade press. Limited direct agency deployment evidence.
- Evidence and measured results
- Flexible-load hundreds of signals; Lambda 19/16 nodes 4 to 5M tokens/s; Annapurna NVHBM; d-Matrix NVLink Fusion; Pinterest Dynamo; Groq 3 LPX 2529 tok/s/user on 100K-context Qwen example; AgentX 30x/45x claims; NVLink 6 fault isolation narrative.
- Limitations and uncertainty
- Trade press based on NVIDIA/partner statements; methodologies differ; overlaps prior MaxLPS/Flex coverage.
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.
- Independent research
- Research conducted outside the implementing organization.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.