ctg. morning brief, 14 Sep 2026
On 14 Sep 2026, From SemiAnalysis (13 Sep): the HBM content race is breaking. Nvidia's Rubin Ultra drops to 192GB per GPU from 288GB on Rubin and B300 (12-hi standard today), with the supply chain preparing 8-hi as the new default. SemiAnalysis argues inference wants to go further: 4-hi HBM keeps the same bandwidth per cube (2048 I/Os maxed at 4 dies) at far lower $/bandwidth, and models 8-hi / 12-hi as a 12.1% / 26.3% all-in system cost premium vs a 4-hi Rubin Ultra NVL576 baseline without matching token throughput. Numbers are SemiAnalysis's.
Why it matters: Europe's AI factories are buying capacity. The frontier is optimizing tokens per scarce HBM wafer. Wrong stack height is a sovereignty tax. Edition candidate.
Source: https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi
Bruegel (10 Sep, PB 18/2026): the EU has about 2 GW of AI compute in 2026, ~5% of global capacity, vs ~35 GW in the US and ~5 GW in China. Planned projects barely move the EU share to 5.6% by 2031. Private capital is not the bottleneck: 76 of 101 Europe2031.ai projects are fully private (84% of planned capacity). Binding constraints are speed-to-operation and grid access, which CADA largely leaves out. Authors say redirect gigafactory subsidies into competitive grants that pay regions for faster permits.
Why it matters: Sovereignty papers do not plug a rack. Permits and transformers do. Europe is still arguing about who owns the logo while the queue for power decides who gets the chips.
Source: https://www.bruegel.org/policy-brief/how-can-europe-address-its-pressing-ai-compute-infrastructure-shortfall
Germany's Bundeswehr is evaluating 30 AI tools to speed battlefield decision-making, aiming for a prototype in Q2 2027 and full use in 2028, cyber command chief Vice Adm. Thomas Daum told Handelsblatt (Reuters, Sep 10). Most tools under review are German-made; French ChapsVision is among the few foreign firms. Daum has ruled out buying Palantir's Maven for now on sovereignty grounds.
Why it matters: NATO already runs Maven. Berlin is building a local shortlist instead of renting the American one. Tempo vs dependency is the whole European AI-defense argument in one briefing.
Source: https://www.reuters.com/technology/german-military-evaluating-30-ai-tools-introduction-2027-2026-09-10/
NestAI is opening a defence AI R&D centre in Latvia: ~30 scientists and engineers, ~€10m planned investment over coming years, announced around the Riga Conference on Sep 11 (Baltic News / company). Follows a June MoU with Finland's MoD and Estonia's Defence Forces, and an August Sweden launch. Focus: NestOS adaptive AI for unmanned and C2 systems in contested, disrupted-comms environments. Latvia's defence spend is approaching 5% of GDP in 2026.
Why it matters: Eastern flank ops are writing the requirements. NestAI is putting labs where the testing ranges and the threat picture already exist.
Source: https://balticnews.com/nestai-expands-operations-in-latvia-to-advance-defence-ai-on-natos-eastern-flank/
Poland's NCBR will open STEP Defence tech calls this autumn: Polish firms and science-industry consortia can seek a combined PLN 500m from FENG for critical defence technologies (gov.pl / NCBR, Sep 8). Max grant per project PLN 140m. Tracks cover breakthrough innovations (Path A) and cutting strategic third-country dependence (Path B). Catalogue includes C-UAS, air/missile defence, ammo, cyber/AI/EW, military mobility. Calls start once the Commission clears the FENG programme change.
Why it matters: European funds finally have a defence lane with real ceilings, not another workshop. The autumn call is when Polish dual-use shops stop waiting for VC alone.
Source: https://www.gov.pl/web/ncbr/technologie-krytyczne-na-rzecz-obronnosci-i-bezpieczenstwa-polski