〔Day Digest〕Nvidia AI Server Hikes Top 15% as NAND Shifts to Molybdenum; Tiangong's 38.15s Beats Human 43.03s - Anthropic's $7B Decart Deal Tests the Largest-IPO Path as Voter Net Support Sits at -17%

0. Daily Arc

Three arcs pull the tape. Compute is the binding constraint: Nvidia AI server prices rise more than 15% ^[1] and SemiAnalysis calls the shift from tungsten to molybdenum in 3D NAND a "must-have" above 300 layers, not an optimization ^[2]. Embodied AI resets the bar — Tiangong's 38.15s in the humanoid 400m beat the human world record of 43.03s ^[3], with the World Humanoid Robot Games field expanding 138% year-on-year ^[4]. The third arc is the capital-versus-populist split: Anthropic's $7 billion Decart deal tests the largest-IPO path ^[5][6], yet US voter net support for data centers has slid to -17% in April 2026, with 833 anti-data-center groups formed in Q1 ^[7].

1. Compute Supply and Memory Materials

• **[ESCALATED] Nvidia AI servers:** prices to rise more than 15% on tight capacity; a supplier flagged forward placement in liquid cooling and AI-server test equipment already shipped to Nvidia and Alibaba customers ^[1].
• **[NEW] NAND material shift:** SemiAnalysis states "molybdenum is not an optimization, it is a must" above 300 layers because tungsten's high resistance, fluorine-based leakage, and deep-via fill limit are blocking further stacking ^[2]. Layer leaders: SK Hynix 321, Samsung 286, Micron 276, Kioxia/SanDisk 218 ^[2].
• **[NEW] China Tungsten High Tech subsidiary (Jinzhou):** 190 million yuan capex for high-precision micro-tool capacity of 140 million units per year, one-year build, second-year ramp, 14.55% post-tax IRR, funded by 178.5 million yuan of equity and 11.5 million yuan of bank loans ^[8].
• **[NEW] Changxin Storage (CXMT) STAR Market IPO** accepted for a large fundraise targeting production-line upgrades ^[9].
• **[NEW] Cloud capex 2026:** global top nine hyperscalers expected to spend more than $886.7 billion, with the top five US firms accounting for nearly 90% ^[10].

2. Embodied AI

• **[NEW] Tiangong 38.15s beats human 43.03s:** Tiangong won the 400m large-class final at the Second World Humanoid Robot Games; runner-up and third went to Glory robots at 39.45s and 39.66s, versus 1:28.03 (88.03s) for last year's champion ^[3].
• **[NEW] Field size:** World Humanoid Robot Games participants up 138% year-on-year ^[4].
• **[NEW] Swancor Advanced Materials' Qiyuan Robotics:** pre-orders open for Qiyuan Q1 and T1, first shipments scheduled September 2026, 210 million yuan in pre-payments already booked; a 10,000-unit production line has been signed with Lens Intelligent Robotics and a core-components line with Wolong Electric Drive ^[11].
• **[ONGOING] World Robot Conference 2026:** open question over who pays for humanoid robots and how the ecosystem is co-developing ^[12][13].

3. Model Layer: Incumbents vs Upstarts

• **[ONGOING] Anthropic's flagship model** continues to struggle for users as cheaper tools thrive ^[14], even as the company's annualized revenue run-rate has crossed $65 billion ^[9] and Decart's $7 billion deal frames the IPO path ^[5][6].
• **[NEW] Mystery model "Ox Alpha":** 1.048 million token context, 131,000 token max output, text/image/video input, free during preview; OpenCode has reserved 100 trillion tokens per day of capacity ^[15]. On a 10-task DeepSWE subset, independent researcher Ben Davis reported an ~80% pass rate versus Claude Fable 5 at 65%, GLM-5.3 and Grok 4.6 at 62%, and GPT-5.6 Sol at 52% ^[15]. **Source caveat:** single-source personal test, not the official DeepSWE leaderboard; also covered by Bloomberg ^[16].
• **[NEW] Token-spend inflection:** the LLM token-spend index posted its first rebound since its May peak while OpenRouter weekly token usage hit a new high — quantity growth has reasserted over price compression ^[9].

4. IPO, Capital and the Chip-Company Land Grab

• **[ESCALATED] Anthropic IPO narrative:** Decart's $7 billion Anthropic deal could appreciate after a public listing, with The Week asking whether Anthropic could be the largest IPO ever ^[5][6], and the Motley Fool flagging parallel IPO fever for both OpenAI and Anthropic ^[17].
• **[NEW] Unitree:** shares closed up more than 6x the IPO price on debut ^[9].
• **[ESCALATED] Chip-company M&A:** Nvidia signed a non-exclusive inference-license with Groq to ship Groq 3 LPU, invested in optical-interconnect plays Ayar Labs and Hark, and took a strategic stake in SSI with access to its Vera Rubin platform ^[18]. Fractile, after landing Anthropic's inference-chip order, saw its valuation jump more than 6x in months ^[9].

5. Policy, Public Backlash and Tail Risks

• **[ESCALATED] US voter backlash:** net support for data centers fell 68 percentage points year-on-year to -17% by April 2026; opposition groups numbered 833 by Q1 2026; the core grievance is local electricity bills rising about 18% against thin local job creation ^[7]. **Contradiction:** AI adoption inside firms is still accelerating — any-business-function use rose from 17.3% in November 2025 to 21.8% in August 2026, and paid AI-product penetration climbed from 7.5% in January 2023 to 54.2% in May 2026, up roughly 47 percentage points ^[7].
• **[NEW] Chongqing's 15th Five-Year AI plan** targets a 250 billion yuan smart-economy core industry by 2030 ^[19].
• **[NEW] Fiscal plumbing for compute:** ultra-long special government bonds and central-budget investment now cover AI and compute-infrastructure projects; in a sample of ten projects, 3.36 billion yuan of special bonds (~54% of sample capex) funds machine rooms, racks, GPU servers, and compute platforms ^[10]. The 2026 Green Compute (AI) Conference in Inner Mongolia and the province's broader green-compute rollout frame the regional play ^[20][21].
• **[NEW] Thematic notes:** AI video models split into T2V, I2V, and Ti2V tracks, with architectures evolving from GAN to diffusion, DiT, autoregressive and multimodal designs aimed at "world simulation" ^[22]. Airbus now identifies AI hardware as the cargo-increment driver, displacing cross-border e-commerce ^[23]. Platforms are training detection AIs against their own generative AIs to throttle the content flood ^[24], and enterprise security has begun codifying a new term — "Authority" — to govern AI Agent action scope ^[25].