ctg. morning brief, 22 Sep 2026
On 22 Sep 2026, SemiAnalysis dropped "Computation and Data Movement for Inference" (21 Sep): MoE serving is four regimes (prefill, midfill, decode attention, decode experts), not one GPU job. Their worked example projects Kimi K3 on Blackwell racks; frontier configs often stay under ~80 GB HBM per GPU once idle KV moves to cheaper DRAM/SSD; a rough 2027 stage target lands near 400-500 GB local fast memory for ~2M live tokens at ~70 kB/token KV. Decode energy dominates (1,000 output tokens means about 1,000 full passes). Numbers are SemiAnalysis projections, not measured benches.
Why it matters: Europe keeps buying fab speeches. Useful inference is memory bandwidth, KV movement, and orchestration. If that is the product, sovereignty is a token factory, not a ribbon-cutting. Edition candidate: inference as factory design vs Chips Act demand politics.
Source: https://newsletter.semianalysis.com/p/computation-and-data-movement-for
Danish defence VCs Final Frontier and Myriad merged under the Final Frontier brand and are raising Fund II at a €100 million target for pre-seed and seed defence tech across the New Nordics and Ukraine (Tech.eu, 21 Sep). First closes: Final Frontier €4.5m (May 2025, Anders Fogh Rasmussen among backers); Myriad €2m (Aug 2025). Combined portfolio already 15 companies; Fund II first close aimed by year-end.
Why it matters: Europe is spending on national security again. Early capital still arrives in crumbs. A Nordic specialist platform at €100m scale is the missing layer between garage and prime.
Source: https://tech.eu/2026/09/21/final-frontier-and-myriad-unite-behind-100m-defence-tech-fund/
Arm joined the European Chips Design Platform (EuroCDP) under the Chips Act / Chips JU as the first commercial IP provider in the program (Arm Newsroom, 21 Sep). Participants get Arm Flexible Access (IP, tools, training); Arm cites 300+ Flexible Access members and 400+ tape-outs as of June 2026. EuroCDP also stands up Design Enablement Teams, many Arm Approved Design Partners. Chips Act framing still points to more than €43 billion through 2030.
Why it matters: Wednesday's Chips Act 2.0 debate is about demand. Monday's move is about design access. Fabrication without design IP is a dead end. This is a small door into a big ecosystem.
Source: https://newsroom.arm.com/blog/arm-joins-eurocdp-europe-chip-innovation
Antwerp's TUSK IC (KU Leuven MICAS spinout) raised €15 million Series A led by Matterwave Ventures, FORWARD.one, and Flanders Future Tech Fund / PMV to scale Ka-band beamformer ICs in mainstream 300 mm bulk CMOS instead of SiGe (New-Tech Europe, 21 Sep). ConnectKa: eight-channel TX 27.5-31 GHz and RX 17.7-21.2 GHz, 360° phase control under 5.625° steps, up to 128 beam states; ESA ARTES contract earlier in 2026 via BELSPO. Money funds prototype to volume; foundry and customer still unnamed.
Why it matters: Electronically steered LEO/MEO terminals need cheap RF channels at volume. Standard CMOS is the cost bet Europe can actually manufacture. Satcom ground gear is dual-use infrastructure wearing a commercial badge.
Source: https://www.new-techeurope.com/2026/09/21/tusk-ic-raises-e15-million-to-scale-standard-cmos-ka-band-beamformers/