ctg. morning brief, 15 Sep 2026
On 15 Sep 2026, From SemiAnalysis (14 Sep): Vera Rubin NVL72 agentic inference on their AgentX benchmark. At 170 TPS on apples-to-apples TRTLLM NVFP4 dense, they measure ~67x total throughput per TCO vs GB300 Dynamo TRTLLM under owning-cost assumptions; at 60-100 TPS (where most providers actually serve) the gap is 1.4x-3x. On power, they already see up to 7x token throughput per MW vs Jensen's GTC claim of ~3x vs Blackwell, and model ~2x more annual profit per GW than the strongest GB300 config they show. Numbers are SemiAnalysis's, early pre-release software.
Why it matters: Europe is buying AI factories in megawatts and purchase orders. The frontier is tokens per scarce watt and dollar. Wrong generation is a sovereignty tax. Edition candidate.
Source: https://newsletter.semianalysis.com/p/vera-rubin-nvl72-agentic-inference
From SemiAnalysis (14 Sep): where the robot brain lives. Frontier robot models already outgrow the chassis (π0.7 on an off-robot H100; DreamZero needs two GB200s). SemiAnalysis models silicon efficiency crossing ~7 robots per shared GPU and DRAM around ~5; one B300 sustains 12 robots within a 500 ms chunk deadline vs Jetson Thor on every unit. Boston Dynamics runs System 2 offboard on Google TPUs; Sunday Robotics fled home WiFi jitter back onboard. Numbers are SemiAnalysis's.
Why it matters: European factories want humanoids without rewriting every plant network. The fight is not only arms and batteries. It is jitter, uplink MIMO, and who owns the planning layer.
Source: https://newsletter.semianalysis.com/p/a-brain-too-big-to-carry-on-device
ASML's $200m EUV tools are all but sold out through 2027, and customers are locking $400m High NA: TSMC from 2030, Samsung and SK Hynix memory High NA from 2028, Intel already past a million High NA wafers (Reuters, Sep 14). JPMorgan put ASML at ~94% of global lithography in 2025; EUV remains a monopoly. Eindhoven broke ground on a plant sized for up to 20,000 workers.
Why it matters: Europe's choke point still prints the world's AI chips. The question is whether the continent keeps the toolmaker without keeping the foundries that eat the output.
Source: https://www.reuters.com/world/asia-pacific/asml-extends-chipmaking-dominance-customers-embrace-high-na-2026-09-14/
France's DGA notified MBDA and Safran for Thundart, the guided rocket winner of FLP-T to replace M270-derived LRU launchers (AeroTime, Sep 11). Reach jumps from ~80 km to 150 km; LPM annex targets at least 13 systems by 2030 and 26 by 2035. 50/50 JV in the Paris region due January 2027; designed, developed, produced in France for full autonomy of use by 2030. Beat ArianeGroup-Thales FLP-t 150.
Why it matters: Paris is buying sovereign deep fires after donating LRUs to Ukraine. Range and jam-resistant guidance matter more than another allied lease.
Source: https://www.aerotime.aero/articles/france-orders-thundart-rocket-system-replace-lru
Leonardo's 30 mm X-Gun won its first export via Kongsberg for Poland's SAN counter-UAS programme (EDR / company, MSPO week). PGZ frames SAN as SAFE-financed (~€3bn cited by PGZ's Grabowski): Kongsberg MCT30 turret, Leonardo gun, Weibel radar, plus Polish private shops. Kongsberg's earlier MCT30 award: €1.67bn for 18 C-UAS batteries.
Why it matters: SAFE money only counts when it becomes a layered battery on a Polish battery park. This is European kit stacking, not another study.
Source: https://www.edrmagazine.eu/mspo-2026-leonardo-x-gun-selected-for-the-polish-san-c-uas-programme