ctg. morning brief, 19 Sep 2026
On 19 Sep 2026, SemiAnalysis (18 Sep) on DeepSeek Engram: learned multi-token embeddings that prefetch from host DRAM, so less HBM is needed for the same quality. Their DeepSeek-V4.1-Flash config uses roughly 189 GiB of Engram memory; offloading to DRAM (and even SSD in experiments) frees HBM for weights and KV cache. They flag NVIDIA's Rubin Ultra HBM despec from 1024GB down to ~200GB per chip as the constraint architecture is codesigning around. On InferenceX agentic serving, MI355X is still 2-4x worse performance per dollar than B200 after TCO normalisation (SemiAnalysis numbers).
Why it matters: Europe's compute gap is usually told as capacity. SemiAnalysis says bandwidth and architecture may matter more: if Engrams keep winning, the race shifts from max HBM stacks to memory hierarchy and day-0 CUDA software. Edition candidate: what that does to Europe's fab and cloud bets.
Source: https://newsletter.semianalysis.com/p/engrams-embedding-entendre-codesign
Thales launched HexaForce (17 Sep): AI-enhanced multi-domain C2 spanning land, sea, air, cyber, space, EM, and information, tested at NATO CWIX 2026. Built with cortAIx (~800 engineers). Thales pitches sovereign data control and ITAR independence; CEO Patrice Caine frames it as covering more of NATO's needs than Palantir's Maven slice (Reuters / EDR / Defence Matters). Airforce Technology cites a target-processing ambition from about 100 to 1,000 per day.
Why it matters: The sovereignty fight is not only chips. It is who runs the command stack and where the data sits. Europe finally has a named prime product in the ring with US battlefield AI.
Source: https://www.reuters.com/business/aerospace-defense/frances-thales-urges-international-framework-ai-development-2026-09-17/
Indra unveiled DRIZZLE at UNVEX (17 Sep): expendable autonomous interceptor drones for mass UAV attacks. More than 400 km/h, about 80 km range alone, about 130 km with a mothership that flies the first 80 km at 30,000 feet then releases interceptors for another ~50 km. Hard-kill C-UAS effector or onboard self-protection for vehicles and helicopters; motherships can swarm and then act as attack vectors (Indra PR).
Why it matters: Cheap attack drones need cheap interceptors with economies of scale. Spain just put a fungible hard-kill layer on the table with real range numbers.
Source: https://www.indragroup.com/en/news/indra-unveils-drizzle-its-new-family-of-autonomous-effectors-to-counter-mass-drone-attacks
Quantum Systems took MOSAIC UXS to sea at a NATO exercise in Sesimbra, Portugal: Seavex (long-endurance unmanned surface) and Subvex (AI underwater vehicle), linking maritime assets with air and ground fleets in one architecture (company / ENGtechnica, mid-Sep coverage). Co-CEO Florian Seibel: the step is connecting domains, not adding another vehicle.
Why it matters: Air and land neo-primes are racing. Maritime is where European coastlines and undersea cables actually sit. Same software spine across domains is the product thesis.
Source: https://engtechnica.com/quantum-systems-adds-seavex-subvex-to-mosaic-uxs/