〔Day Digest〕$35B China Nvidia Orders and $10B Germany Bet Meet CoreWeave Funding Stress and Elice's 32% Debut Crash — AI Capex Reprices as Funding, Trust and Demand Diverge

0. Today's Arc

Three threads collided across the Asian session and European open: funding stress at the infrastructure layer (CoreWeave, Elice), a fresh wave of demand from China and Europe ($35B of Nvidia chip orders, a roughly $10B Germany package), and a governance shock (the OpenAI bugging indictment) ^{[1][2][3][4][5]}. Net: AI capex is repricing, not retreating — the question is which side of the trade bleeds first.

1. Capital Markets — Funding Stress at the Edges

• **[ESCALATED] CoreWeave:** unsecured bond yields ~15%, CDS spreads above 1000bp, with ~$1.5B of unsecured paper sitting structurally subordinated inside a ~$35B debt stack ^[4]. Barclays global macro research head Ajay Rajadhyaksha framed the move as a 2023 SVB-style localized signal rather than a 2007 New Century-style systemic event; Goldman argued the unsecured market is not CoreWeave's marginal funding cost ^[4].
• **[NEW] Elice (KOSDAQ):** largest 2026 KOSDAQ IPO at 90,500 KRW top-of-range pricing, raising 201B KRW (~$151M) with retail oversubscribed ~400x — then opened near 78,000 KRW (-14% vs issue) and fell as much as 32% to 61,400 KRW within minutes, erasing ~300B KRW of market cap ^[5]. The pattern repeats: Nvidia-backed Firmus cut its IPO price 25% and withdrew, Nscale's ~$300B funding ask was cut nearly in half, and BlackRock's AirTrunk was delayed ^[5].

2. Demand Side — Capex Is Not Slowing

• **[NEW] China hyperscaler Nvidia orders:** Reuters reports ByteDance, Alibaba and Tencent placed over $35B of Nvidia chip orders ^[2]. Tagged to the Tencent-side read from Hunyuan head Yao Shunyu — "we can't not work overtime like DeepSeek, but we can work DeepSeek into overtime" — with Hunyuan headcount nearing 2,000 a year after the 28-year-old joined the Office of the President ^[6].
• **[NEW] China compute build-out:** MIIT called for faster expansion of computing power infrastructure, with research framing domestic substitution, cluster build-out and earnings delivery as a triple resonance ^[7]. UBTech set up a 20M RMB embodied-intelligence subsidiary in Huizhou ^[8]; iSoftStone (软通动力) and partners set up a 65M RMB Tibet smart-compute company (45% Yitong Century Guangdong, 40% iSoftStone, 15% Shaanxi Xingmo) ^[9]; Alibaba Cloud entered the Gartner container management leaders quadrant for a 4th consecutive year, the only Asia-Pacific name there ^[10].
• **[NEW] Nvidia Germany:** Jensen Huang signed a roughly $10B European investment package — a cooperation letter of intent with Germany's economics ministry, four German startup investments and an AI factory deal with Siemens — layered onto the "Made for Germany" 4.1 trillion euro potential-investment pledge led by Siemens and Deutsche Bank ^[3]. Huang: "To be honest, nothing pains me more than seeing a country with world-class technology talent fall so far behind in tech" ^[3].

3. Trust and Governance

• **[NEW] OpenAI bugging case:** ex-engineer Abdulrahman Fuzan (Saudi national, US green card since 2020, hired July 2025) had bail denied on Oct 19 by a California federal court; prosecutors cite alleged ties to Saudi royals and flight risk; he is charged with using an electronic device to capture oral communications from a leadership-level meeting room, facing up to 5 years ^[1]. Fortune's first-day-in-court read sits alongside the indictment ^[11].
• **[NEW] Anthropic:** Semafor reports a leaked AI bill is drawing backlash ^[12].

4. Models, Robotics and Applications

• **[NEW] DeepSeek:** fully open-sourced the DeepInfer inference framework, already serving V4.1 Flash at high throughput and low latency, with future model inference code slated for the same channel ^[13].
• **[NEW] Dream Labs:** the ex-Nvidia GR00T team emerged from stealth with the Dream-1 model and a $90M seed round, $3/hour public API pricing, claiming zero new training data is required to control unseen dual-arm parallel grippers ^[14].
• **[NEW] Momenta:** CEO Cao Xudong forecast 2028 as the L4 scaling inflection, framing a "smart-driving Moore's Law" of safety/performance decadal gains every 2 years, with world models, data scaling and model scaling as the three drivers ^[15].
• **[ONGOING] T3 Mobility** showed Robotaxi full-stack tools — "central cloud + edge cloud" redundant architecture, virtual dispatch, in-cabin assistant, automated depot — at the 2026 Physical AI & World Model Conference in Nanjing ^[16].
• **[NEW] Clinical AI:** a model in *Lancet Digital Health* predicts cancer patient survival from a single face photo, trained on 40M+ face images and 30k+ patient photos; lead author Raymond Mak, radiation oncologist at Mass General Brigham ^[17].

5. Sectors, Policy and What Would Falsify the Read

• **[NEW] Pharma:** the National Medical Products Administration (NMPA) held an "AI + drug regulation" meeting, pledging support for key tracks to break through first ^[18].
• **[NEW] Water transport:** the Ministry of Transport released the first batch of AI application scenarios for water transport, plus ten intelligent agents and ten vertical models targeted for 2027 ^[19].
• **[NEW] Power:** the 2026 Power Embodied Intelligence Application Contest practical round opened at China Southern Power Grid's Banqiao training base in Dongguan, 15 teams, three real substation tasks (inspection, emergency outage operation, pre-meter-replacement voltage check) ^[20].
• **[NEW] Guangdong:** 7 more generative AI services registered, cumulative 108 ^[21].
• **Source quality control:** the CoreWeave SVB-not-2008 framing is relayed through a social post, with Barclays and Goldman named but not directly quoted; treat the unsecured-bond spread, not a term-loan print, as the marginal indicator ^[4]. The Elice 32% slide is one name, not a sector — Firmus, Nscale and AirTrunk are different listings in different markets, so quote the pattern, not a point ^[5]. The $35B China Nvidia order book is a Reuters exclusive with undisclosed buyer mix and delivery schedule, so flag the magnitude, not unit economics ^[2].