ctg. morning brief, 11 Sep 2026
On 11 Sep 2026, Paris-based Arlequin AI raised a €28m Series A for topological neural networks, a non-LLM architecture built for relationship analysis on messy investigation data (company / Tech Funding News, Sep 10). Round co-led by redalpine and OTB Ventures with Bpifrance's Defence Innovation Fund; Xavier Niel and ZEBOX joined; Vsquared and 10x Founders upped. Total raised now €32.4m. About 50 people, 30+ clients across four European countries (security, defence, fraud, integrity). Offices in London and Berlin; Silicon Valley lab planned by end-2026. Valuation not disclosed.
Why it matters: Europe's sovereignty story usually means a smaller American model. Arlequin is betting on a different architecture for high-stakes decisions where hallucination is not a vibe problem. Edition candidate.
Source: https://arlq.ai/news/arlequin-ai-series-a-topological-neural-networks
Auterion and Ukrainian partner Airlogix have started mass production in Germany of more than 5,000 AI-guided strike drones for Ukraine under a Berlin order industry sources put near €300m, due mid-2027 (AA / Defense Post, Sep 4-10). Seth X for medium-range swarm work; Anubis up to 1,500 km with a ~50 kg warhead. Auterion OS, Airlogix airframes. First units already built near Munich; second plant eyed in eastern Germany.
Why it matters: Germany is not waiting for a European missile committee to invent stockpiles. Software from Munich, hardware from Ukraine, volume on a German line.
Source: https://thedefensepost.com/2026/09/10/germany-begins-mass-production-of-5000-plus-strike-drones-for-ukraine/
Bundeswehr cyber chief Vice Admiral Thomas Daum says Germany is evaluating about 30 AI tools for battlefield decision support, aiming for a prototype in Q2 2027 and wider use in 2028 (Handelsblatt via Reuters / TNW, Sep 10). Most candidates German; French ChapsVision still in. Palantir Maven was ruled out in April. Allies already field Maven.
Why it matters: Sovereignty buys control. It also buys an 18-month gap while other armies already run the American stack. The hard part after the bake-off is integration, not the pitch decks.
Source: https://thenextweb.com/news/bundeswehr-30-ai-tools-prototype-2027-palantir-alternative
From SemiAnalysis (10 Sep): their Energy Model now tracks 75 GW of firm, binding OEM orders for behind-the-meter AI compute, about 20 GW ordered in Q2 2026 alone. They estimate ~3 GW of US datacenter IT capacity will be BTM-powered by year-end, with multiple years of triple-digit growth. Six gates between press release and first power: bankability, permits, fuel, equipment, workforce, islanding physics. Numbers are SemiAnalysis's.
Why it matters: Europe is writing gigafactory speeches. The US is ordering turbines and reciprocating engines like the grid already lost. Attribute every GW figure as SemiAnalysis.
Source: https://newsletter.semianalysis.com/p/what-is-so-hard-about-behind-the