Strategic PartnersIssue 495 records
NVIDIA
A small edition led by NVIDIA's sovereign-AI push in Kazakhstan and the scrutiny it drew. Firebird says it holds a U.S. export licence for 45,000 NVIDIA Grace Blackwell GPUs for Phase 1 of a Kazakh AI factory (commercial launch targeted for Q3 2027) and pledges at least 10% of capacity to Kazakh companies and government agencies. Independent coverage points out that 880 NVIDIA GPUs are running in the country today, that the 1 GW Ekibastuz figure is a long-term target and that reported deal values differ. In Germany, a Friedrich Naumann Foundation commentary says the Oct 19 government-NVIDIA declaration lacks binding milestones and puts the power-cost gap with the U.S. at roughly 2.6x for a 100 MW site. On software, NVIDIA's TensorRT 11.1 Global Performance Tuner reports 1.14x-1.63x inference gains on ten undisclosed workloads, a vendor figure with limited SLED relevance. No new U.S. state, K-12 or university NVIDIA deployment was found for Oct 23-24. Lowndes County (Alabama) has no ruling before its Oct 27 hearing, and San José's standards return to the Planning Commission on Nov 18.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 independent reporting2 vendor claim
- Outcomes
- 2 emerging2 cautionary1 mixed
- Source freshness
- 5 new this fortnight
- Research completed
- 2026-10-25
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Sovereign-AI announcements with NVIDIA run well ahead of binding commitments and installed capacity
In Kazakhstan, Firebird's release cites a 45,000-GPU export licence and about 500 MW of 'total potential' capacity, with Phase 1 targeting Q3 2027, while Aivio notes that the NVIDIA GPUs running in the country today are Kazakhtelecom's 880 and that NVIDIA's post gives no dollar figure for the investment agreement signed at the presidential palace; AI Affairs reports that the 1 GW Ekibastuz figure is a long-term target and that accounts of the June $10 billion package differ. In Germany, the Friedrich Naumann Foundation finds that the government-NVIDIA declaration leaves binding milestones, investment commitments and responsibilities open. Across both, the headline figures are permissions, targets or intentions rather than contracted, delivered capacity.
Operating questionBefore a public body plans programs around a government-NVIDIA or partner AI-factory announcement, which figures are signed, funded and installed, and which milestones and owners are binding?
Supporting evidenceFirebird Inc.; Kazakhtelecom; Government of Kazakhstan (Ministry of Artificial Intelligence and Digital Development); NVIDIAAivio News (GrowQ AB); subjects: NVIDIA, Firebird, Kazakhtelecom, Government of KazakhstanAI Affairs; subjects: NVIDIA, Government of Kazakhstan, Kazakhtelecom, Firebird, Freedom Holding Corp.Friedrich-Naumann-Stiftung für die Freiheit (author Steffen Saebisch); subjects: German federal government, NVIDIA, Siemens, Audi
Power price and grid readiness, not GPU supply, are the stated gate on where sovereign AI compute gets built
Kazakhstan's AI minister, quoted in Firebird's release, calls reliable, competitively priced energy the foundation of affordable intelligence, and AI Affairs reports that the Ekibastuz build depends on reliable electricity, cooling and connectivity, with Carnegie warning (via SCMP) about unreliable grids. In Germany, the Friedrich Naumann Foundation's model puts annual power for a 100 MW IT-load site at about 136 million euros at German industrial prices versus about 52 million at U.S. prices, and says investors struggle to find grid-connected sites. Both governments' AI-compute ambitions are framed as energy and siting problems first.
Operating questionFor a proposed public or partner AI factory, what delivered power price, grid-connection date and cooling capacity are confirmed for the first phase, and how sensitive is the business case to them?
Supporting evidenceFirebird Inc.; Kazakhtelecom; Government of Kazakhstan (Ministry of Artificial Intelligence and Digital Development); NVIDIAAI Affairs; subjects: NVIDIA, Government of Kazakhstan, Kazakhtelecom, Firebird, Freedom Holding Corp.Friedrich-Naumann-Stiftung für die Freiheit (author Steffen Saebisch); subjects: German federal government, NVIDIA, Siemens, Audi
Evidence ledger
Full record · every source keeps its link and limitations
Firebird (Oct 23): US export licence for 45,000 NVIDIA Grace Blackwell GPUs in Kazakhstan; at least 10% of capacity pledged to Kazakh companies and government agencies
Firebird says it has secured a U.S. export licence for 45,000 NVIDIA Grace Blackwell GPUs for Phase 1 of its Kazakhstan AI Factory, under construction with Kazakhtelecom at Data Center Valley and targeting commercial launch in Q3 2027; a planned Phase 2 would take total potential capacity to about 500 MW. The release also describes an estimated $5 billion investment and a cluster of up to 109 MW with more than 45,000 GB300 GPUs and over 130 permanent jobs. Firebird commits at least 10% of total GPU capacity to Kazakh companies and government agencies, plans Firebird Labs Kazakhstan (equity, compute access, scholarships, research partnerships with alem.ai) and Kazakh-language AI work. The expansion builds on cooperation agreements signed with the government in June 2026.
Why it matters, evidence and limitations
- Why it matters
- International public-sector signal, not a U.S. SLED case: it is a concrete example of a government securing a reserved share of a privately operated NVIDIA AI factory in return for state cooperation. The 10% public/local allocation is the transferable element for U.S. states weighing shared or reserved capacity in partner-built facilities; export-licensing, energy and governance context does not transfer directly.
- Evidence and measured results
- Company press release (classed vendor-claim) quoting Kazakhstan's Deputy Prime Minister and Minister of AI and Digital Development, NVIDIA's CEO and Firebird's CEO. No contract text, allocation mechanism, pricing or delivery schedule beyond the Q3 2027 target is published.
- Limitations and uncertainty
- Self-reported announcement before any Phase 1 operation. Internal figures differ (45,000 GPUs and up to 109 MW and $5 billion for Phase 1; ~500 MW total potential with Phase 2); the June 2026 package has been reported elsewhere as a $10 billion, 100,000-GPU plan. 'Total potential' capacity is not committed capacity.
Aivio News (Oct 23): Firebird's 45,000-GPU licence is 'a permission, not a delivery'; Kazakhtelecom runs 880 NVIDIA GPUs today
Aivio compares the Astana announcements with installed capacity using NVIDIA's own figures: Kazakhtelecom has deployed 400 H200 and 480 B300 GPUs (880 accelerators) in Astana and Almaty, and plans more than 1,000 GPUs at Data Center Valley plus more than 1,000 Vera Rubin GPUs in Ekibastuz next year, which it calculates as under 5% of the licensed Firebird volume and in a different programme. It stresses that the 45,000-GPU export licence is a permission rather than a delivery, that 500 MW is 'total potential' across a phase under construction and a planned phase, that NVIDIA's post attaches no dollar figure to the investment agreement signed at the presidential palace, and that NVIDIA's investment in Higgsfield AI is undisclosed. It treats the inDrive/Pony.ai robotaxi item as a feasibility assessment, not a launch, and names the Q3 2027 date as the test.
Why it matters, evidence and limitations
- Why it matters
- Useful scrutiny for any public-sector reader of sovereign-AI announcements: it shows how to separate licences, potential capacity and reserved shares from GPUs actually running. The method (compare headline commitments to installed counts from the same announcement) transfers directly to U.S. state and university AI-factory announcements.
- Evidence and measured results
- Independent news analysis of NVIDIA's blog post and the Astana visit; arithmetic based on NVIDIA's published GPU counts. Single-author, short article; no independent site visit or official documents beyond NVIDIA's post.
- Limitations and uncertainty
- Relies on NVIDIA's blog for figures; the under-5% comparison spans different programmes (Kazakhtelecom versus Firebird). The publisher is a small Sweden-based outlet; no comment from Firebird or the government is included.
AI Affairs (Oct 23): Huang pitches Kazakhstan as Europe's AI-factory hub; Ekibastuz 1 GW is a long-term target, June deal figures differ and grid limits are flagged
Drawing on a CBN press release, The Astana Times, the South China Morning Post and DKNews, AI Affairs reports that Huang urged Kazakhstan to operate national AI infrastructure and keep control of its data, calling the capacity to build its own LLMs 'a national security concern', and cast the country as an AI-factory hub for Europe. It reports that Data Center Valley in Ekibastuz is designed for phased expansion to as much as 1 GW, with the initial 215 MW phase described as fully reserved and Kazakhtelecom's first 50 MW facility due in June 2027, but that DKNews called 1 GW a long-term target rather than operational capacity and identified electricity, cooling and connectivity as requirements. Accounts of the June agreements differ ('US$10 billion plan' versus 'worth up to $10 billion'). Separately, Freedom Holding is building a $2 billion NVIDIA-based Sovereign AI Hub in Astana (Akashi site, up to 100 MW, first B300 cluster in 2027). It sets Huang's demand case against Ken Griffin's 21 October bubble warning and an April Carnegie warning about grids, connectivity and regulation.
Why it matters, evidence and limitations
- Why it matters
- International sovereign-AI context with limited direct U.S. transfer. Relevant to U.S. public-sector readers as a worked example of how a national government frames AI compute as infrastructure alongside grids and telecom, and how independent reporting distinguishes targets, reservations and operating capacity.
- Evidence and measured results
- Independent news report aggregating named secondary sources (CBN press release, The Astana Times, SCMP, DKNews, Carnegie via SCMP); no primary documents are linked for the 215 MW reservation or the deal values.
- Limitations and uncertainty
- Secondary aggregation: key figures (215 MW fully reserved, deal values, 1,000+ Vera Rubin GPUs) are attributed to a CBN press release and other outlets rather than verified by the publication. The bubble discussion is context, not evidence about this project.
Friedrich Naumann Foundation (Oct 23): Germany-NVIDIA declaration leaves milestones, investment commitments and responsibilities open; power price gap estimated at about 136 million versus 52 million a year for 100 MW
Writing after the Invest in Germany summit, the foundation argues that the federal government and NVIDIA (with Siemens, Audi and others) plan to advance industrial AI models, but the published declaration leaves concrete, binding milestones, investment commitments and responsibilities open. It says investors struggle to find sites with grid connections and competitive power prices, and gives a model calculation based on Epoch AI data: a 100 MW IT-load data center uses about 709 GWh a year, costing about 136 million euros at Germany's new-contract industrial price of 19.2 cents/kWh (including full electricity tax) versus about 52 million at the U.S. price. It cites Bitkom (57% of German companies use AI, nine in ten see themselves at the start), legal uncertainty over data protection and the AI Act, and a foundation study finding 61% of German token consumption on OpenRouter went to Chinese models in the summer.
Why it matters, evidence and limitations
- Why it matters
- International government signal on the Germany-NVIDIA pact already in the archive (signed Oct 19). New since then: a structured critique that the declaration is not yet a project and that power cost and siting, not announcements, set where AI compute is built. The power-cost framing transfers to U.S. state AI-infrastructure incentives; the German regulatory points do not.
- Evidence and measured results
- Commentary by a liberal party-affiliated political foundation; the power figures are the author's model calculation on Epoch AI data and stated tariffs, not a measured study; survey figures are attributed to Bitkom and the foundation's own Taipei hub.
- Limitations and uncertainty
- Opinion piece with a policy agenda (deregulation, lower power costs); the cost comparison uses a single tariff and ignores other site costs; it does not examine the declaration text in detail. The 61% token-share figure refers to one platform.
NVIDIA Technical Blog (Oct 23): TensorRT 11.1 Global Performance Tuner searches 200+ compiler knobs with accuracy filtering; 1.14x-1.63x over default across ten unnamed workloads
NVIDIA describes the Global Performance Tuner in TensorRT 11.1: build routes are explicit combinations of more than 200 compiler knobs; trtexec expands a route expression, builds and measures each candidate, discards routes that crash or fail an accuracy threshold, and selects the fastest eligible engine. Across ten evaluated workloads, the best eligible route delivered 1.14x to 1.63x end-to-end performance versus the default route (normalised to 1.0x). Best practices include fast/mixed/full search strategies, --dryRun, resumable tuning with --tuningCacheFile and --continue, representative inputs and reference outputs, and re-tuning after any change to model, shapes, GPU, CUDA/driver or TensorRT version. A Python implementation is to follow in Torch-TensorRT 2.15.
Why it matters, evidence and limitations
- Why it matters
- Limited direct SLED relevance. Applies to public universities, research centres and agencies that self-host inference on NVIDIA GPUs (for example campus AI factories) and want more throughput from existing hardware; the accuracy-threshold design is relevant where public services need documented model quality.
- Evidence and measured results
- Vendor technical blog with a single relative-performance figure; the ten workloads, GPUs, batch sizes, accuracy thresholds and tuning time are not disclosed. NVIDIA's separately opened tuning documentation lists caveats: gains are model-dependent, routes are version- and hardware-dependent, builds are not bit-deterministic, and non-default routes have no performance guarantee across major releases.
- Limitations and uncertainty
- Vendor-reported relative gains on undisclosed workloads; no absolute latency, cost or energy figures; build-time cost of tuning not quantified; no independent replication found.
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.