〔Day Digest〕OpenAI's $852B Valuation Stumbles as IPO Slips and Executives Exit, While China's Ulanqab Projects Surge 147% and Guangdong Launches 'Token Loans' – Anthropic's $7B Decart Deal Tests the Buildout

0. Weekly Arc

The dominant story is not a market print but a governance problem at the most valuable private AI company. OpenAI has gone through nearly five organizational restructurings this year, lost a series of executives, and quietly pushed its IPO from this year to next ^[1]. While that chaos unfolded, China pressed ahead with a physical buildout: signed compute-center projects in Ulanqab surged from 36 to 89, and local governments rolled out token-finance products and provincial computing-network plans ^[2][3][4]. The counterpoint: Anthropic is reportedly closing a $7 billion Decart acquisition, and Nvidia is reinvesting its cash flow to own more of the stack ^[5][6][7]. Net: corporate disarray at the frontier, state-backed infrastructure momentum in China, and a shift in the base-model narrative ^[8].

1. OpenAI Turmoil

• **[ESCALATED] OpenAI (per FT via Wallstreetcn):** nearly five restructurings this year; safety team disbanded; employees report frustration; IPO expected this year has slipped to next year ^[1].
• **[NEW] Departures:** Chief Revenue Officer Denise Dresser resigned after less than a year in the role she took last December; former CFO/COO Brad Lightcap and ethics head Chloé Bakalar also left within a year ^[1].
• **[NEW] Fidji Simo:** once seen as CEO Sam Altman's 'number two,' on part-time adviser status after sick leave, but two sources say she still communicates with staff almost daily and 'is still pulling strings behind the scenes' ^[1]. OpenAI says she is performing advisory duties normally ^[1].
• Source quality: this is a single-paper FT story relayed by one Chinese aggregator; no secondary confirmation ^[1]. The falsifying test would be a stabilization of the executive bench and a confirmed timeline; none is visible ^[1].

2. China's Compute and Token Buildout

• **[NEW] Ulanqab:** signed compute-center projects rose from 36 at end-2024 to 89 by end-June 2026, with sites such as Wanrun (Ulanqab) Low-carbon Smart Computing Base and Kuaishou Xinghe Data Center under construction or preparing to break ground ^[2].
• **[NEW] Fujian:** provincial '15th Five-Year Plan' calls for completing a province-wide integrated computing network, public intelligent-computing centers in Fuzhou, Xiamen and Zhangzhou, edge compute nodes, an AI for Science platform, AI-Ready datasets in government/medical/marine, and provincial AI parks in Fuzhou, Xiamen and Quanzhou ^[4].
• **[NEW] Token finance:** Guangdong has issued its first special financial product for the token economy, 'Token loan'; institutions argue the compute-leasing model may shift from 'selling compute' to 'selling tokens' ^[3].
• **[NEW] Supply-chain signal:** Cambricon's inventory rose 66.83% year-on-year to RMB 8.248 billion by June 30, equal to 45.32% of total assets, which the company attributed mainly to raw materials and consigned processing ^[9].
• The falsifying test is whether signed projects actually break ground and whether token loans convert into compute demand ^[2][3].

3. Ecosystem and Model Shifts

• **[NEW] Anthropic–Decart:** Anthropic is closing in on a $7 billion deal for Decart, having beaten Nvidia to the table ^[5].
• **[NEW] Nvidia:** is turning its massive cash flow into another AI growth engine and looking to own another layer of the AI ecosystem ^[6][7].
• **[NEW] Lenovo:** after server orders 'far exceeded expectations,' commentators ask whether the PC giant will follow Oracle's path from stable cash cow to AI-driven growth ^[10].
• **[NEW] New venture:** former Alibaba Qwen large-model technical lead Lin Junyang founded Pragmatik Labs (p7k) in Shanghai, with Gaorong Ventures and Sequoia China co-leading the round and Tencent plus Shanghai Future Industry Fund providing support ^[9].
• **[NEW] Base-model thesis:** Arcee AI pre-training head Varun Singh argues 'The Base Model Is Dead'; GPT-3-era training drew roughly 85% of data from the web and Wikipedia, and the field's default logic is now being questioned ^[8].

4. Policy and Sector Reads

• **[NEW] Healthcare:** experts at a Fudan University competition said generative AI in hospitals still lacks clear norms and standards; data security, algorithm bias and responsibility attribution are hurdles. CAS academician Chu Junhao sees AI moving from 'data intelligence' to 'physical intelligence' and then 'life intelligence,' and cited terahertz technology for non-invasive tumor diagnosis ^[11].
• **[NEW] Vertical agriculture AI:** Anhui Agricultural University professor Wang Xiaobo released the soybean vertical large model 'Fengshu' 2.0 at the 32nd National Soybean Research and Production Seminar in Nanning; it fuses multi-modal data and multiple model outputs to produce structured knowledge for soybean research and breeding ^[12].
• **[NEW] European autos:** BNP Paribas Asset Management portfolio manager/strategist Sophie Huynh says Europe is more likely an AI beneficiary than developer, auto stocks are so cheap that almost no one considers upside, and it may take one to two years for market consensus to form; US momentum is slowing while Europe is 'just beginning to recover' ^[13].
• **[NEW] Labor tension:** FMCG companies have added shelf-photo AI that judges store displays in seconds, but salespeople report a five-second task now takes fifteen minutes, daily store-visit assessment targets have risen, and salary structures are unchanged ^[14].
• **[NEW] Military-business analogy:** CCTV's documentary 'Zhisheng' shows machine 'wolves' charging ahead of soldiers; the commercial reading is that value shifts from physical scale to cognitive judgment and flexible systems ^[15].