〔Overnight Brief〕Marvell-Google Custom-Silicon Warrant for 58.97M Shares at $206.58 Lifts MRVL ~12% Pre-Market as Itochu Pledges 10 Japan Data Centers, Fractile Eyes $6.5B, and Nebius Files $4.5B Convertible - Capex, Capital, and Doctrine Converge in One Night

0. Overnight Arc

A capital-markets wave swept the AI tape: Marvell's binding custom-silicon deal with Google and a warrant for up to 58.97M shares at $206.58 lifted MRVL ~12% pre-market ^[1][2]; Itochu unveiled ~10 Japan data centers by 2030 at ~50MW each ^[3][4]; Nebius filed a $4.5B convertible preferred note offering ^[5]; and chip startup Fractile pursued a $6.5B round with a ~$250M Anthropic chip commitment ^[6]. The model layer kept moving in parallel - OpenAI's CFO guided to a 2027-or-earlier IPO ^[7] and shipped a teen-focused ChatGPT ^[8][9], Tesla deployed ByteDance's Doubao LLM ^[10][11], and Anthropic's Dario Amodei traded public blows with White House PCAST co-chair David Sacks over centralized vs distributed AI ^[12]. One night, three frontiers on the same tape: silicon, capital, doctrine.

1. The Compute Capex Chain

• **[NEW] Marvell-Google custom-silicon expansion:** Marvell signed a binding agreement on July 29, 2026, then on August 18 issued Google a warrant for up to 58.97M shares at $206.58/share, with vesting tied to custom-product revenue scale - the larger Google's procurement, the larger the potential equity grant ^[1][2]. Scope spans AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing products tied to Google's TPU ecosystem ^[1][2]. Pre-market: MRVL ~12% ^[1].
• **[NEW] Itochu data center buildout:** Japan's Itochu plans ~10 data centers by 2030, with total investment in the hundreds of billions of yen, ~50MW per facility, and 1-2 projects per year, sited in the Tokyo metro, Osaka metro, and Kyushu ^[3][4]. Itochu may partner with JR East, which holds land and power relationships, to ease land/power constraints ^[4]. Japan's data center services market is projected at 5.65 trillion yen by 2030, up ~30% from 2025, and Mitsubishi Corp is among the other large trading houses piling in ^[4].
• **[NEW] Fractile funding:** AI chip startup Fractile is in talks at a $6.5B valuation; Anthropic has agreed to purchase ~$250M of Fractile chips ^[6]. Single-source via Cailian Press social wire; treat the $6.5B as unconfirmed.
• **[ONGOING] NVIDIA-Mercor investment:** NVIDIA is in discussions to invest in data-labeling firm Mercor as part of a $20B-round funding, with General Catalyst in talks to lead ^[13][14]. Mercor's NVIDIA revenue is rising via the Nemotron open-model program; NVIDIA paid Mercor "tens of millions of dollars" last quarter, per The Information ^[13]. Single-source on the round size.

2. Capital, Earnings, and the OpenAI Exit Clock

• **[NEW] Nebius $4.5B convertible:** AI compute rental company Nebius plans to issue $4.5B in convertible preferred notes ^[5]. Pricing/conversion terms not disclosed in source.
• **[NEW] OpenAI IPO guidance:** OpenAI's CFO said the company will list in 2027 or earlier ^[7]. CFO named in source; no further identity detail.
• **[NEW] SpaceX-Cognition:** SpaceX reportedly attempted to acquire AI coding startup Cognition ^[15]. No terms disclosed; single-source social wire.
• **[NEW] Accelink Technologies 1H 2026:** Revenue 6.61B yuan (+26.07% y/y), net profit attributable to shareholders 582M yuan (+56.34% y/y), non-GAAP net profit 568M yuan (+57.77% y/y); basic EPS 0.73 yuan (+55.32% y/y); dividend 3.7 yuan per 10 shares ^[16]. Company frames demand as AI-infrastructure-driven ^[16].
• **[NEW] Kunlun Tech 1H 2026:** Revenue 5.359B yuan (+43.55% y/y), overseas revenue 5.203B yuan (+51.21% y/y), net profit attributable to shareholders 1.088B yuan; AI revenue scaling sequentially under the upgraded "4+3 strategy" ^[17].

3. Models, Distribution, and the Centralization Fight

• **[NEW] OpenAI teen ChatGPT:** OpenAI introduced a more age-appropriate ChatGPT version for teens ^[8][9].
• **[NEW] Tesla + ByteDance Doubao:** Tesla in-vehicle systems began rolling out ByteDance's Doubao LLM, per Volcano Engine's official account ^[10][11]. No user or market-share figures in source.
• **[NEW] Replit Free Mode:** Replit Free Mode is now powered by OpenAI's GPT-5.6 Luna, per Replit's X post ^[18]. Distribution-channel signal.
• **[NEW] Waymo + Gemini:** Google said Waymo is integrating Gemini into a custom robotaxi called "Ojai" ^[19].
• **[NEW] China-side stack:** MIIT plans to focus the robotics industry on 2 key links and 4 priority sectors; Shanghai Zhangjiang's 15th Five-Year Plan targets AI phones/computers; Beijing added 20 newly filed generative AI services; Jiangsu issued an "AI + industrial software" action plan; Momenta, XHEART, and QNX will co-build a safety-standard autonomous platform; JD.com launched a robotics strategy targeting ~10B-yuan resource commitment to ~100 brands by 2028 ^[20][21].
• **[ESCALATED] Amodei vs Sacks - centralization vs distribution:** Investor Gavin Baker on the All-in podcast claimed Anthropic CEO Dario Amodei had internally said Anthropic might become "the only private company in the world besides governments" ^[12]. Anthropic's Sasha de Marigny (Chief Brand and Communications Officer) and multiple researchers denied the framing; Amodei published a long rebuttal calling "centralized in a few hands vs distributed" a false dichotomy ^[12]. White House PCAST co-chair David Sacks (also an All-in co-host) replied August 17 with a 9-point post arguing frontier AI is too powerful to distribute ^[12]. The fault line is no longer safety-vs-acceleration; it is concentrated-vs-distributed ^[12].

4. Source Quality and What Would Falsify

• The Fractile $6.5B valuation and ~$250M Anthropic commitment are single-source via Cailian Press social wire ^[6]; the SpaceX-Cognition acquisition attempt is an unattributed social relay ^[15].
• The Itochu data center plan is well-sourced via Nikkei relays ^[3][4][22], but all three traces converge on a single primary report - confirm against an Itochu IR release or local-government filing.
• Accelink and Kunlun Tech results are company filings relayed through Chinese financial wires ^[16][17]; treat as primary.
• Falsification tests: a) does Marvell disclose the warrant vesting schedule in its next 10-Q and does Google procurement clear any threshold; b) does Itochu file land/power agreements with named prefectures or with JR East; c) does Fractile announce a lead investor with a term sheet; d) does the Marvell-Google warrant trigger enough revenue to materially dilute the equity grant, given the 58.97M-share ceiling at $206.58 ^[1][2].