Strategic PartnersIssue 655 records
NVIDIA
A robotics, sovereign-infrastructure and power-constraints edition with no new SLED deployment evidence. At CoRL, NVIDIA introduced GR00T 2 and Isaac Lab 3.0 with vendor-reported results (40 of 50 tasks completed at least once; 71% to 95% average success with one to five demonstrations); an independent Nov 2 preprint on the prior GR00T N1.7 finds all policies struggle in open-world tasks and rankings shift after fine-tuning. NVIDIA, Rebellions and SK Telecom announced an NVLink Fusion collaboration with R200 systems expected in 2028, Deutsche Telekom said its Munich AI factory's first phase is fully booked and it will add B300 systems to reach 20 MW, and MIT Technology Review argues electricity is the binding constraint on data-center buildouts. All NVIDIA and operator figures are unverified claims.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 vendor claim1 independent research1 independent reporting
- Outcomes
- 3 emerging2 cautionary
- Source freshness
- 5 new this fortnight
- Research completed
- 2026-11-10
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
NVIDIA-platform capacity plans are conditional on 2028 availability, financing decisions and electricity, not just chip supply
Three sources describe growth that depends on conditions outside GPU supply. The Rebellions and SK Telecom release says R200 systems are expected ready for deployment in 2028 and uses 'plans' language. Deutsche Telekom's Munich expansion to 20 MW is announced, but its further Bavaria expansion has no final investment decision and depends on financing. MIT Technology Review reports that developers see electricity, more than public opposition, as the constraint on AI buildout. For SLED planners this supports treating announced capacity as plans to verify, with power and financing gates, rather than deliverable capacity. None of the sources is a SLED deployment.
Operating questionWhich availability dates, financing decisions and utility confirmations must be in writing before a public institution plans around announced AI capacity?
Supporting evidenceNVIDIA, Rebellions, SK TelecomDeutsche TelekomMIT Technology Review
Vendor-reported robot-policy results need independent, local evaluation before adoption
NVIDIA reports GR00T 2 completed 40 of 50 tasks at least once and reached 71% to 95% average success with one to five demonstrations, with partner uses described as testing. An independent multi-site preprint on the prior GR00T N1.7 found all systems struggle on open-world tasks, rankings shift after fine-tuning, and no policy finished any G1 humanoid task. The independent study does not evaluate GR00T 2, so the two do not contradict each other; together they show that headline and leaderboard figures may not transfer to a specific robot and task. Labs should measure on their own hardware with stated trial counts.
Operating questionWhat local, blinded trial protocol and trial count will a lab require before relying on a vendor's robot-policy success rates?
Supporting evidenceNVIDIATan Sang Nguyen et al. (26 authors incl. Siddhartha Srinivasa, Dieter Fox, David Hsu)
Evidence ledger
Full record · every source keeps its link and limitations
Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation (GR00T 2, Isaac Lab 3.0, new benchmarks)
NVIDIA introduced GR00T 2, an open world action model built on Cosmos, plus Isaac Lab 3.0 (Newton physics integration) and Isaac Lab-Arena 0.4 developer preview with new DexBench and VABAR benchmarks. NVIDIA reports GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations and reached 71% average success across 50 tasks with one demonstration per task, 95% with five. It also reports up to 2.5x training throughput for Isaac Lab 3.0 with Newton versus Isaac Lab 2.3 with PhysX in a dexterous-manipulation benchmark. Named partner uses (Nexuni, Foxconn, Siemens, MicroAGI) are testing or evaluation work; no production rollout is stated.
Why it matters, evidence and limitations
- Why it matters
- Limited and indirect. Robot learning is not a core SLED workload; relevance is mainly to university robotics labs and state or regional workforce and advanced-manufacturing programs that may adopt open NVIDIA simulation tooling. No state, local, or K-12 deployment is described.
- Evidence and measured results
- Vendor blog read in full. All headline numbers (40/50 tasks, 71%/95% success, 2.5x throughput, top-three ranking on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces) are NVIDIA-reported; the blog gives no sample sizes, trial counts, baselines or confidence intervals for the success rates. Partner deployments are described as testing, evaluating or exploring.
- Limitations and uncertainty
- Vendor announcement; numbers not independently verified in this edition (see the independent RoboColosseum evaluation of the prior-generation GR00T N1.7). The 2.5x throughput is for one benchmark against one earlier release. Conference-week marketing context.
RoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild
A multi-site preprint evaluates five vision-language-action policies and three frontier AI agents on four embodiments across 5,699 blinded pairwise Open Track comparisons at 124 sites. All systems struggle to complete open-world tasks reliably; Open Track rankings poorly predict post-fine-tuning performance, with rank reversals on all three shared embodiments; rankings agree weakly across embodiment suites (Kendall W = 0.24). After fine-tuning on 100 demonstrations per task, GR00T N1.7 leads on SO-101 and pi0.5 leads on Franka and Bimanual-YAM. On the G1 humanoid no policy completes any of five tasks (average progress 13.3% to 23.2%).
Why it matters, evidence and limitations
- Why it matters
- Limited. Relevant as independent context for any university robotics lab or workforce program evaluating NVIDIA's GR00T family; no SLED deployment is evaluated.
- Evidence and measured results
- Abstract page read on arXiv (v1, 2 Nov 2026) and the PDF text (via pdftotext) checked for its limitations statement. Covers GR00T N1.7, the prior generation, not GR00T 2, and the study states it is an initial snapshot with uneven model and embodiment coverage and five fine-tune tasks per embodiment.
- Limitations and uncertainty
- Preprint (not peer reviewed). Evaluates GR00T N1.7, not GR00T 2; results may not carry over. Coverage differs across models and embodiments, and the study does not claim exhaustive comparison.
NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion
NVIDIA, Rebellions and SK Telecom announced a collaboration, supported by Korea's Ministry of Science and ICT, in which Rebellions will connect its R200 NPUs to NVIDIA's platform through NVLink Fusion and the MGX rack architecture, and SK Telecom plans to deploy the systems in its AI data center for sovereign AI services including the national AI for All project. The companies plan a disaggregated-inference approach using NVIDIA GPUs for prefill and Rebellions NPUs for decode. R200 systems are expected to be ready for deployment in 2028.
Why it matters, evidence and limitations
- Why it matters
- Limited. It shows how a national government sponsors a sovereign-AI stack on NVIDIA's open-platform interconnect. For U.S. SLED the transferable point is the heterogeneous-accelerator and procurement-lock-in question, not any deployment, since nothing is deployed until 2028.
- Evidence and measured results
- Press release read in full. It is a plan and a collaboration announcement: the language is 'plans to', 'aims to', and 'expected'. No performance, power, cost or efficiency measurements are given; the release's efficiency claims are goals.
- Limitations and uncertainty
- Vendor press release with executive quotes; no independent verification; deployment is three years away and all performance statements are aspirational.
AI demand rises: Deutsche Telekom expands AI factory in Munich
Deutsche Telekom says the first stage of its Munich Industrial AI Cloud, live since February 2026 and described as adding 50 percent to German AI capacity, is fully booked seven months later. It will add NVIDIA Blackwell Ultra B300 systems alongside B200 and RTX PRO 6000 Blackwell Server Edition GPUs, nearly another quarter of a billion in investment, and raise capacity to 20 megawatts. Users named include Siemens, Daimler Truck and research organizations; the company says the public sector can use it. A further Bavaria expansion is under discussion but no final investment decision has been made.
Why it matters, evidence and limitations
- Why it matters
- Limited and indirect: a European operator-run AI cloud, not a SLED deployment. It is a data point for state and campus leaders weighing buy-versus-rent GPU capacity and shared sovereign or regional compute models.
- Evidence and measured results
- Operator press release read in full. Demand and booking claims, the 20 MW and 'doubling to 20,000 GPUs' figure (a federal minister's quote) are company or government statements without independent audit; utilization, pricing and power-source details are not provided.
- Limitations and uncertainty
- Operator marketing with political endorsements. 'Fully booked' is not quantified; the minister's quote also says 'after one year', conflicting with the seven-month figure, which shows imprecision in the quoted claims. Not independently verified.
The backlash about data centers might really just be about electricity
Reporting from Virginia's Data Center Alley argues that opposition to data centers is driven largely by electricity: grid-connected facilities (an average 54 MW in Loudoun County) have led Virginia to require data centers to pay the full cost of new transmission lines, and developers expect rules that could cut them off from grid power at peak. Larger AI sites (300 MW at minimum, ideally about 6,000 MW) tend to use on-site gas turbines, which bring air-pollution and neighbor concerns. A developer says industry ambitions are constrained less by public opposition than by a lack of electricity.
Why it matters, evidence and limitations
- Why it matters
- Moderate for infrastructure planning: state and local governments, utilities and campuses face siting, permitting and cost-allocation decisions as AI facilities grow. The article is not about NVIDIA specifically; it is general context for the power constraints on NVIDIA-based capacity.
- Evidence and measured results
- Newsletter reporting read in full: one reporter's visit and interviews with an activist and developers, with figures attributed to sources and some linked reports not opened here. It contains no NVIDIA-specific data and no systematic measurement.
- Limitations and uncertainty
- Commentary and reporting by one author; not an audit or study. Contains no NVIDIA-specific information. Some claims (for example the Anthropic turbine purchase) are described as reported. Relevance to SLED is by inference.
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 research
- Research conducted outside the implementing organization.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.