ctg. morning brief, 29 Sep 2026
On 29 Sep 2026, Kongsberg signed a US Marine Corps framework worth up to $404 million for Naval Strike Missile launcher units, weapon control systems, and sustainment through 2034, with work mostly in Norway (Naval News / company track, 28 Sep). Call-offs land as firm orders; the frame backs NMESIS land-based anti-ship batteries.
Why it matters: Norway keeps selling serial missile infrastructure into the US force design. That is export production capacity, not a concept slide.
Source: https://www.navalnews.com/naval-news/2026/09/kongsberg-signs-404-million-framework-agreement-for-usmc-nsm-equipment/
Rolls-Royce will invest £300 million (~$398 million) to build and upgrade UK aerospace and defence manufacturing and engineering sites, including more than £150 million across Bristol, Inchinnan, Rotherham, and Ansty for defence-related capacity, plus more than £140 million in Derby (Reuters, 28 Sep).
Why it matters: Sovereign engines and defence kit need floorspace and tooling. London is putting real money into the factories that make them.
Source: https://www.reuters.com/business/aerospace-defense/rolls-royce-invest-398-million-uk-manufacturing-2026-09-28/
EU member states approved the first five European Defence Projects of Common Interest under EDIP: DECODER (drones/counter-drones), integrated maritime and seabed defence, SPACE, federated air and missile defence with early warning, and Eastern Flank Watch. Commission fact sheets cite indicative envelopes from roughly €3.5-5bn (DECODER to 2033) up to €60-100bn (Eastern Flank Watch to 2036), with ~€325m EDIP seed for EDPCI setup and Ukraine in four of five projects (Council / Defense News, 28 Sep).
Why it matters: Brussels finally named the joint industrial bets. The numbers are aspirational envelopes, not cash in the bank, but the menu is now public and fundable.
Source: https://www.defensenews.com/global/europe/2026/09/28/eu-nails-down-five-defense-priority-areas-to-channel-common-spending/
SemiAnalysis walked GLM-5.3 sparse attention and HiSparse KV offload: sparse attention cuts SDPA memory/bandwidth but still needs full context in HBM for top-k selection unless hierarchical DRAM offload is used. On their InferenceX board they model GB300 as lowest serving cost at 150 tok/s (~$0.044/M total tokens vs $0.238 for MI355X ATOM in one comparison; GB200 ~82% cheaper than ATOM at that target). They also flag B200 Dynamo-SGLang at ~$0.0666/M total tokens vs $0.1265 for MI355X when holding p90 TTFT under 2s (SemiAnalysis, 28 Sep). Keep these as SemiAnalysis modeled figures.
Why it matters: Europe's compute buildout lives or dies on tokens per watt and HBM. Sparse attention is not a free lunch on capacity unless the serving stack offloads KV the hard way.
Source: https://newsletter.semianalysis.com/p/sparse-savings-persistent-demand-inside-glm53
Mistral opened a Munich hub for Physics AI and industrial AI, bringing the Emmi AI team onboard and naming work with BMW on crash simulations and Siemens Energy on industrial AI, plus a TUM wind-tunnel partnership on automotive aerodynamics digital twins. The company restates a goal to build 1 GW of European compute capacity by 2030 and stresses open-weight models running on customer infrastructure under European law (company, 28 Sep).
Why it matters: Open-weight European AI is moving onto the factory floor where the proprietary data and the money sit. Munich is a product bet, not a flag-planting tour.
Source: https://mistral.ai/news/hallo-deutschland/