Strategic PartnersIssue 469 records
NVIDIA
At its GTC Berlin keynote on Oct 21, NVIDIA introduced DSX POD, factory-built containerised Vera Rubin AI factories it says can be deployed in as little as 14 days at a prepared site, plus DSX designs it says cut cooling water to near zero and power by up to 30%. Same-day reviews note the clock starts only after site preparation, so grid connection and permits, which public bodies control, are unchanged, and the figures are vendor maxima with no disclosed deployment. NVIDIA also launched the OpenShell Reference Architecture for containing AI agents; an independent test of the underlying runtime found defaults held but auto-approval and audit-mode settings let data out. On public research compute, Bristol and France's GENCI plan federated training across national supercomputers so sensitive data stays in-country, HPI is adding GB300 NVL72 for researchers and students, and TIM is building an NVIDIA-based sovereign AI platform for Italian public administration. An analyst cautions that data sovereignty is not technology independence.
- Evidence records
- 9
- Cross-source patterns
- 3
- Evidence classes
- 6 vendor claim2 independent reporting1 independent research
- Outcomes
- 6 emerging2 cautionary1 mixed
- Source freshness
- 8 new this fortnight1 recent
- Research completed
- 2026-10-22
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
NVIDIA's AI factory speed and efficiency claims are conditional on site readiness, climate and grid access, which public bodies control
DSX POD's six-months-to-14-days figure applies only at a prepared site. The companion DSX release's water and power savings are 'up to' maxima that depend on dry coolers and climate and rest on partner tests and a simulated grid trial. Independent and analyst reviews the same day note that grid queues, permits and substations, which counties, utilities and regulators control, are untouched, and that no commercial deployment has been disclosed.
Operating questionWhen an AI facility is proposed or procured, what is the timeline from site start to first production workload, what cooling method and supplemental-cooling hours are expected, and which water and power figures will be metered and reported?
Supporting evidenceNVIDIA; 17 named delivery partners including Bull, Dell Technologies, HPE, Lenovo, Schneider Electric, Vertiv, Anduril and ArmadaNVIDIA; Nscale; Lambda; Emerald AI; EPRI; Nebius; U.K. National Grid; Schneider Electric; Siemens and other ecosystem partnersalfw.ai editorialNAND Research (Steve McDowell)
Agent containment holds by default but depends on how operators configure approvals, enforcement and logging
NVIDIA's OpenShell Reference Architecture adds DPU-based monitoring and quarantine on top of the OpenShell runtime, with major security and software vendors integrating. An independent pre-registered test of the runtime found default policies blocked every exfiltration path tried, but auto-approval, audit mode, read-write rules and parameters on read-only rules let data out, and the policy prover does not cover MCP or GraphQL. The hardware layer has not been independently tested and applies only where BlueField is installed.
Operating questionBefore agents use staff credentials on public systems, who approves access changes, are L7 rules enforced rather than audited, is auto-approval disabled, and have canary exfiltration tests been run on the actual runtime and driver?
Supporting evidenceNVIDIA; partners including CrowdStrike, Palo Alto Networks, Cisco, Microsoft, SAP, ServiceNow, Palantir, BBVA, SantanderSorami (Australian cyber security consultancy)
Sovereign public AI compute is being defined as in-country data and operational control; technology independence varies
The UK-France federated learning plan keeps sensitive data in each country and spans NVIDIA and non-NVIDIA supercomputers through open-source software. TIM's platform for Italian public administrations and research centres offers in-country hosting and residency controls on the NVIDIA stack. An analyst argues that data and deployment sovereignty does not guarantee independence at the technology layer, and that buyers should keep choice and negotiating leverage.
Operating questionFor a public AI compute arrangement described as sovereign, which layers are actually controlled locally (data, operations, technology), what would exit or migration cost, and can workloads federate across different hardware?
Supporting evidenceBristol Centre for Supercomputing (University of Bristol); GENCI (France); IDRIS; CINES; University of Cambridge; NVIDIATIM Group / TIM Enterprise (Telecom Italia); NVIDIANAND Research (Steve McDowell)
Evidence ledger
Full record · every source keeps its link and limitations
NVIDIA introduces DSX POD: partner-built, containerised Vera Rubin AI factories it says cut on-site deployment at a prepared site from six months to as little as 14 days
At the GTC Berlin keynote on Oct 21, NVIDIA introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers and integrate and validate before delivery. Computing, networking, cabling, rack-level power and liquid cooling are assembled in the factory; partners complete rack-to-rack cabling, wiring, piping and functional validation before shipment, separately from site preparation. NVIDIA says units connect to a prepared site's power, cooling and networking and reduce deployment time from six months to as little as 14 days. Seventeen partners, including Bull, Dell, HPE, Lenovo, Schneider Electric, Vertiv and defence-oriented Anduril and Armada, are to deliver DSX POD worldwide in early 2027. The release frames the benefit as earlier time to first production token and revenue.
Why it matters, evidence and limitations
- Why it matters
- Limited and indirect for U.S. SLED today. The design targets large greenfield builds, so the near-term relevance is for state research-computing consortia, national labs and university systems weighing new capacity, and for local governments asked to approve sites. Factory-built containers shift work off site but not the zoning, grid connection or permitting that counties and utilities control.
- Evidence and measured results
- Vendor press release with no named customer, price, pod capacity or power draw. The six-month baseline and 14-day figure are NVIDIA's own and are conditional on a prepared site. Delivery is a 2027 plan. NVIDIA's product FAQ, also opened, states customers still need connections to site power, facility water, cooling and networking.
- Limitations and uncertainty
- Claims are forward-looking and undemonstrated in a disclosed commercial deployment; no capacity, cost or energy figures; availability early 2027.
NVIDIA says DSX reference designs can cut facility cooling water to near zero and use up to 30% less power for the same output; partner tests and a simulated UK National Grid trial cited
NVIDIA says its DSX reference designs pair 100% liquid-cooled Vera Rubin infrastructure (coolant up to 45 degrees Celsius) with dry coolers in a closed loop, which can reduce facility cooling water to near zero; it estimates 54 million litres a day saved at 2-gigawatt scale versus cooling towers, and notes local climate determines whether supplemental mechanical cooling is needed. DSX MaxLPS dynamically allocates power within a provisioned budget; NVIDIA says it can produce the same tokens with up to 30% less power. Cited partner tests: Lambda, 24% more cluster token throughput and 19% more GPU capacity on HGX B200 within a fixed budget; Nscale in Iceland, 49% higher inference throughput and 37% more managed GPU capacity across four GB300 NVL72 racks within 264 kW. In an early DSX Flex trial with Emerald AI, EPRI, Nebius and the U.K. National Grid, a Blackwell Ultra cluster cut power draw by up to 40% in response to simulated grid events while critical workloads continued. New open-source DSX OS components entered developer preview.
Why it matters, evidence and limitations
- Why it matters
- Relevant to SLED as context for local and state decisions on AI data centres, where water and grid impact dominate public debate (as in the archived Lowndes County thread). It gives officials vendor-stated design options to ask about, not evidence that any specific facility will use less water or power. A U.K. grid operator's participation in a trial is the most public-sector-relevant element.
- Evidence and measured results
- Vendor release. Efficiency figures are maxima ('up to') under unstated conditions; the 54-million-litre figure is a modelled comparison, not a site measurement. Partner tests are described without methods, baselines or independent review. The grid trial used simulated events.
- Limitations and uncertainty
- All figures from NVIDIA or partners; no independent verification; climate-dependent supplemental cooling; dry coolers' effect on power, noise and land not described.
Independent analysis: DSX POD's 14-day figure starts only after site preparation; grid queues, permits and substations are untouched and the claim is undemonstrated
alfw.ai's same-day analysis says the 14-day clock begins only when a container reaches a site whose power, cooling and networking are in place, quoting NVIDIA's own wording that factory work 'takes place separately from site preparation'. It describes 'two clocks': the integration clock NVIDIA compresses against its own six-month baseline, and the pre-connection clock of grid queues, permits and substation work, which can run for years and is unaffected. It notes no named customer, price or pod capacity, that the timeline has not been demonstrated in a disclosed commercial deployment, and that the 'up to 100%' water and 'up to 30%' power figures are platform maxima under unspecified conditions, not POD measurements. It cites Microsoft's termination of its anchor agreement at Nscale's Loughton project, where grid power may not arrive until the 2030s, as context, while stating the reason for that exit is not public. It proposes early-2027 deliveries with a named customer's full timeline as the test.
Why it matters, evidence and limitations
- Why it matters
- Useful for SLED decision-makers who hear 'weeks not months' claims: local governments, utility regulators and university capital planners control or depend on the slow clock (permits, grid connection), so vendor speed claims do not shorten their part of the process.
- Evidence and measured results
- Editorial analysis based on NVIDIA's release, the keynote live blog and other coverage, with sources listed. No original data; it does not test the product. Its Loughton example is explicitly caveated.
- Limitations and uncertainty
- Commentary, not measurement; partly based on wire copies of the release; Loughton example is context, not causal evidence.
NVIDIA launches OpenShell Reference Architecture: agent sandboxing on Vera CPUs with BlueField-4 'in-silicon' enforcement via NVIDIA Sentry; security vendors, SAP and ServiceNow integrate
NVIDIA announced the OpenShell Reference Architecture, available now, for building, containing, observing and enforcing agents. It combines Vera Rubin infrastructure for model inference, Vera CPUs running the open-source OpenShell runtime for agent harnesses and tool use, and BlueField-4 DPUs running NVIDIA Sentry for continuous out-of-band monitoring; NVIDIA says if an agent moves outside its software boundary Sentry quarantines and stops it in milliseconds, and that the offload boosts agent capacity by up to 20%. It extends to RTX Spark, DGX Spark and DGX Station at the edge and runs on Canonical, SUSE and Red Hat Linux, Kubernetes, Windows via Microsoft Execution Containers, and AWS, Google Cloud and Azure. Security vendors (1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI, Zscaler) are integrating support; BBVA and Santander are exploring it; ServiceNow says its AI Passport layer lets enterprises revoke an agent's access across software and hardware within seconds.
Why it matters, evidence and limitations
- Why it matters
- Relevant to SLED CIOs and CISOs drafting agent policies: it is a concrete architecture for containing agents that act with staff credentials, and several partners (ServiceNow, Microsoft, SAP, CrowdStrike, Palo Alto) are common in state, local and campus estates. No public-sector deployment is named, and the hardware enforcement layer requires BlueField DPUs that most public data centres lack.
- Evidence and measured results
- Vendor release with partner list and performance claims ('milliseconds', 'up to 20%', '30x higher throughput per megawatt') without methods. Partners are mostly integrating or exploring, not reporting production results.
- Limitations and uncertainty
- No independent test of Sentry or the full architecture; partner commitments are integrations or explorations; hardware enforcement limited to BlueField-equipped systems.
Independent test of NVIDIA OpenShell v0.1.2: default policy blocked every exfiltration path tried and cut an agent-driven secret leak from 10/10 to 0/10, but operator settings such as auto-approval and audit mode let data out
Sorami tested OpenShell v0.1.2 (released Sep 28, 2026) under a pre-registered plan: 35 deterministic test IDs over 123 trials plus agent tests at N=10 per condition, with public logs. Default deny held for HTTP, HTTPS, raw TCP, direct IP, external DNS and unlisted binaries (3/3 each); a malicious setup script run by a local agent leaked a canary secret 10/10 without OpenShell and 0/10 under the default policy. Operator-enabled settings changed the result: read-write rules leaked 10/10; a GET-only rule still let data out via query string or header; leaving enforcement unset (audit mode, the default for L7 rules) let a forbidden POST through while logging it as allowed; automatic approval granted new public hosts with no human in 12/12 trials, including drafts from blocked connections. The policy prover reported GraphQL, MCP, WebSocket and JSON-RPC rules as unsupported while the loader accepted them. No bypass of a documented control was found; three logging gaps appeared on the VM driver. Newly relevant because NVIDIA's Oct 21 OpenShell Reference Architecture builds on this runtime; the page metadata also shows a modification dated Oct 22 (UTC).
Why it matters, evidence and limitations
- Why it matters
- One of the few independent, reproducible tests of an agent sandbox. For SLED security teams evaluating NVIDIA's new reference architecture, it shows the runtime's defaults hold but that common convenience settings (auto-approval, audit mode, broad rules) defeat containment, and that MCP traffic is not covered by the policy prover. It also maps controls to a government framework (Australia's ISM).
- Evidence and measured results
- Pre-registered test plan with deviation log and public evidence repository; Wilson 95% intervals for agent tests. Limits stated by the authors: one release, one VM driver on one Apple Silicon laptop without an NVIDIA GPU; Docker, Podman and Kubernetes drivers untested; a small local model (qwen3:8b), not frontier agents; AI assistants helped build the harness; no independent specialist review; repeated runs are not independent replication. The author is a consultancy that sells agent security reviews.
- Limitations and uncertainty
- Tested v0.1.2, which may differ from the runtime shipped with the Oct 21 reference architecture; single host and driver; small agent; consultancy author with commercial interest.
UK and France plan federated learning across national AI supercomputers (Isambard-AI, Jean Zay, Adastra) so sensitive data stays in-country; announced at GTC Berlin with NVIDIA technical support
BriCS and GENCI announced a federated learning plan under the UK-France ConcordAI partnership to connect Isambard-AI (NVIDIA Grace Hopper, part of the government-funded AI Research Resource) with France's Jean Zay (NVIDIA) and Adastra (other hardware), so researchers in both countries can train larger models while each keeps control of its own machines and data. Each site trains locally and shares only model updates, so sensitive data in fields such as healthcare, defence or aerospace need not leave its country. Teams at Bristol, Cambridge, IDRIS and CINES, with NVIDIA's support using the open-source NVIDIA FLARE, have been testing with increasing model sizes and checking results match single-machine training; direct connection is targeted by the end of 2026. The release also notes Juelich and GENCI interest in linking JUPITER AI Factory and AI France Factory infrastructure, and GENCI plans to connect its supercomputers with French hyperscalers for startups and industry.
Why it matters, evidence and limitations
- Why it matters
- Directly relevant to public research computing: two publicly funded national systems are sharing capacity across borders without moving data, including across mixed vendor hardware. The model transfers to U.S. state or university consortia (for example, multi-institution health research) that face data-residency or data-use restrictions, with the caveat that it is a plan with testing under way, not a production service.
- Evidence and measured results
- Joint announcement by the public operators, framed as a plan. It reports testing activity but no results, model sizes, performance overheads or governance agreements. NVIDIA's role is technical support and software.
- Limitations and uncertainty
- Announcement of a plan; no published test results, overhead figures or allocation rules; completion target end of 2026.
Hasso Plattner Institute adds NVIDIA GB300 NVL72 to its AI data centre for researchers and students, announced at GTC Berlin
HPI announced at GTC Berlin a major investment in partnership with NVIDIA to expand its AI data centre with GB300 NVL72 technology. Managing Director Daniela Gerd tom Markotten said HPI will be among the first research institutes in Europe to integrate the system, which will be used to analyse large scientific datasets, accelerate simulations and develop new models; researchers and students will have access for training and running large AI models. HPI hosted a keynote watch party where students discussed developments with an NVIDIA team on campus.
Why it matters, evidence and limitations
- Why it matters
- A university-level example of rack-scale AI hardware being opened to students as well as researchers, a pattern seen in prior editions (NJIT). Transferability to U.S. public institutions is limited: HPI is privately funded, and the release gives no size, cost, allocation or facility details.
- Evidence and measured results
- Institute's own short release; no figures on investment size, number of racks, power, cooling, allocation policy or timeline.
- Limitations and uncertainty
- No quantitative detail; 'partnership with NVIDIA' terms undisclosed; privately funded institution.
TIM Enterprise to build NVIDIA-based sovereign AI platform hosted in Italy for businesses, the Public Administration, startups and research centres, with edge nodes under IPCEI-CIS
TIM said it is expanding sovereign AI infrastructure with NVIDIA so Italian businesses, Public Administrations, startups and research centres can develop and deploy AI while keeping control over data, models and operations within Italy. The TIM Enterprise platform, hosted in Italy, will integrate TIM's central cloud and distributed edge with NVIDIA full-stack infrastructure, with controls supporting data residency, access governance, security and regulatory compliance. It will offer managed services for open models including NVIDIA Nemotron (host, fine-tune, evaluate, operate), NeMo Guardrails for programmable policies, and hosting of models developed by Italian universities and research centres. As part of IPCEI CIS (an EU Important Project of Common European Interest on cloud), TIM is developing edge cloud nodes across Italy equipped with NVIDIA GPUs.
Why it matters, evidence and limitations
- Why it matters
- An international example of a national telecom positioning sovereign AI cloud explicitly for public administration and universities, analogous to U.S. state cooperative contracts or government cloud offerings. Transferability is limited: the release has no customers, capacity, prices, certifications or timeline, and 'sovereign' here means in-country hosting and control, not vendor independence.
- Evidence and measured results
- Operator press release; no capacity, customers, pricing, certification (for example national cloud qualification) or delivery dates.
- Limitations and uncertainty
- Announcement only; no figures; EU-funded edge programme context not quantified.
NAND Research on GTC Berlin: factory-built AI factories and DPU-based agent enforcement are promising, but sovereignty at the data layer deepens dependence on NVIDIA's stack
NAND Research's same-day analysis summarises the keynote (DSX POD, supply-chain capacity agreements, Google Cloud A5X Vera Rubin instances expected in November 2026, OpenShell Reference Architecture, extended Confidential Computing, Agent Knowledge Foundation, robotics and Hyperion) and gives a view: factory integration improves time to production 'only if power, site readiness, and partner execution keep pace'; DPU-based agent enforcement addresses a legitimate problem; but power, site readiness, supply-chain capacity, security effectiveness and partner execution will determine results. On sovereignty it argues that while NVIDIA offers more regional capacity and customisation, it 'deepens dependence on its hardware and software ecosystem', that open models do not ensure infrastructure independence when tools are tightly coupled to NVIDIA, and that buyers should measure promises against deployment speed, utilisation, reliability, cost and architectural flexibility while preserving choice and negotiating leverage.
Why it matters, evidence and limitations
- Why it matters
- Relevant to public buyers in Europe and to U.S. states adopting 'sovereign AI' language: it separates data and deployment control from technology independence, which matters for procurement strategy, exit planning and competition.
- Evidence and measured results
- Analyst opinion with disclosure that the firm may have advised companies mentioned; no new data. Summaries of announcements are consistent with NVIDIA's releases.
- Limitations and uncertainty
- Opinion; possible advisory relationships disclosed generically; no independent measurements.
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.