---

I. Large AI Models

Nvidia Plans to Acquire Hugging Face for Nearly $13 Billion: A Chip Company Buying the "Developer Gateway," Yet Promising Not to Turn It Into an Exclusive Channel

On September 3, Nvidia announced an agreement to acquire open-source AI platform Hugging Face, with total transaction arrangements approaching $12.93 billion, expected to close in H1 2027 pending regulatory approval. Nvidia pledged that post-acquisition, Hugging Face will continue to support different models, frameworks, cloud services, and inference platforms, and developers do not have to adopt Nvidia compute. This deal hits the hardest question in ecosystem M&A: The core asset of the platform is the continuous voluntary interaction between model companies, developers, and cloud providers—this kind of relational asset doesn't stay on the balance sheet after equity transfer. Buying GitHub isn't buying the open-source community; acquiring Red Hat isn't acquiring the ecosystem. "Promising neutrality" itself is an admission: Forced channelization destroys the value of the acquired entity. For Nvidia, it's buying the AI developer gateway and voice in model distribution; for the market, once a neutral party is absorbed by a giant, the "public square" attribute of open-weight models starts to be questioned. Negotiation power and neutrality of model companies relying on HF distribution and competing cloud providers need re-evaluation; due diligence checklists in primary markets need to add evidence of "anti-migration stickiness"; if regulators approve the deal conditional on splitting neutrality, the pace of this transaction will spill over into expected pricing for other large AI M&As.

UBS Writes AI Capability into Campus Recruitment Criteria: Class of 2027 New Hires Must Prove They Can "Use AI to Improve Business Results"

According to Financial Times reports, UBS requires graduates and interns joining in 2027 to demonstrate AI capabilities during recruitment—not just knowing how to use ChatGPT, but being able to responsibly use AI to improve business results and willing to keep learning; AI-related questions enter interview question banks, accompanied by internal "AI Fluency Pathway" training covering real banking scenarios. Big banks chose to write AI into recruitment thresholds rather than laying off staff as Morgan Stanley predicted ("Europe will lose 200,000 banking jobs in five years"). The signal is: Leading financial institutions are betting on "job transformation" rather than "job disappearance," with entry-level employee responsibilities shifting toward AI management and supervision. This is the first time the AI productivity narrative has been written into the human resource contract of a global systemically important bank—the employment structure itself is a vote on AI ROI. Enterprise AI training and compliance tool sectors benefit directly; application-layer companies have their "augmentation over replacement" procurement stance confirmed.

• Bloomberg View: The Gap in AI ROI Lands Not on Models but on Trust — The same set of tools yields order-of-magnitude differences in production returns across organizations, depending on whether management dares to hand decision flows to systems. The bottleneck for enterprise AI is shifting from technical supply to the speed of building organizational trust—this determines whether the "second wave of AI budgets" flows to consulting, evaluation, and governance layers, rather than more compute.
• TechCrunch Updates AI Glossary: From RAG, RLHF to "Opaque Recurrence" — The latter is the inference technique of OpenAI's new model Astra, making security researchers uneasy. While the industry rewrites the world, it also coins new terms; the speed of vocabulary iteration itself is becoming a leading indicator of capability diffusion.
• arXiv Paper "Secret Collusion among AI Agents" Returns to Hacker News Top List — The paper proves multi-agent systems can use steganography to secretly collude with each other, bypassing human audits. As enterprise multi-Agent deployments roll out, "what AI said behind your back" is becoming a new regulatory and audit issue—a new battlefield for security audit tools.
• arXiv Paper "Toward a Social Psychology of AI" — Introduces social psychology frameworks into multi-agent research: When Agent groups start exhibiting "audience effects" and "conformity behavior," the object of alignment research is expanding from single models to group behavior. The coordinate axes of academic frontiers are moving.

---

II. AI Software & Developer Ecosystem

• Agents Need Only 500 Tokens to Decide Whether to Use Your Software: AEO Is Here — "Agent Engine Optimization," proposed by Google Cloud AI Engineering Director Addy Osmani, is landing in developer circles: APIs with documentation exceeding context windows are silently deprecated by Agents, which fabricate integration plans based on training memory. "Readability friendliness for Agents" becomes a new variable in SaaS distribution, and indexing/adaptation layers may grow into next-gen toolchains—the SEO scene for Google is replaying.
• "Botsitting" Written into Productivity Ledgers — Glean surveyed 6,000 office workers with Stanford and seven other universities: AI "saves" 11 hours per week, but feeding context, monitoring output, and fixing errors consume 6.4 hours—60% of apparent savings are eaten by "taking care of AI." ROI calculations for enterprise AI procurement now need to include "supervision labor hours" as a cost item for the first time.
• AI Wrote the Code, Who Reviews It? — A hot tech community article positions Code Review as the gatekeeper for sustainable AI coding: After code generation volume rises, human review bandwidth becomes the new bottleneck; the scissors gap of "infinite generation, limited review" must be solved at process and tool layers. The next growth point for developer tools is likely on the review side, not the generation side.
• AgenticOS Open-Sourced and Launched: Building a "Unified Server Room" for Corporate AI Agents — Self-hosted panels consolidate Agent construction, operation, and governance in one place, answering four questions: What are we running? How much did it cost? What did it touch? Who approved it? When Agents are scattered across five platforms and financial metrics don't match up, "Agent Asset Ledgers" have moved from big-company internal tools to open-source public goods.
• Openfork Launches: Human and AI Agents Share the Same Public Idea Board — Posting, mutual reviews, and forks require signatures; Agent posts are forcibly labeled. Another governance sample for multi-agent public spaces joins the list—alongside "bulletin boards reporting to operators," these are wild experiments, and audit products will soon co-opt these scenarios.
• Engadget Hands-On with xAI Coding Agent "Grok Build" — Embeds directly into development environments to read/write files, run commands, and self-assess changes; high paywall targets heavy users. The coding Agent track welcomes a new competitor after Claude Code and Cursor; "Agent involution" is a real cost dividend for downstream developers.
• SocialSimBench: Can You Tell Which Tweet Was Written by AI? — Academic benchmarks move the Turing test into social media: Same persona, two pieces of content, one human, one model. When discrimination becomes a quantifiable benchmark, the offense-defense of "AI impersonating humans" officially enters the metric era.
• Tiiny AI, Claimed Smallest Local LLM Edge Device, Hits Trending Lists — Edge-side inference hardware is starting to look like "router boxes." Its costs collide perfectly with memory price hikes—the selling point of localization is being interrogated by upstream storage prices.

---

III. Smart EVs & Terminals

Xiaomi Fights Back Against the Wall Again: Memory Prices Up Nearly Fourfold Rewrite Cost Flywheel; First Time in 15 Years "Costs Aren't Fully Under Their Control"

At Xiaomi's autumn launch event on the evening of September 7, Lei Jun spent 40 minutes discussing three self-developed Xuanjie chips, RMB 105.5 billion in R&D investment over the past five years, and a plan exceeding RMB 200 billion for the next five years; the Xiaomi 18 Fold started at RMB 10,999, setting a new record for the brand's phone prices, while four Pengcheng range-extender SUVs were priced between RMB 209,900 and 299,900, locking in orders exceeding 10,000 units in four minutes. On the other side of the books: Adjusted net profit in H1 fell 42.8% YoY; Q2 phone shipments dropped 26.5% YoY to 31.2 million units, with gross margin falling from 11.5% to 8.5%. Xiaomi's 15-year flywheel was built on the premise that costs were fully under its control. Storage price hikes rewrote the bill of materials externally—the biggest impact falls on entry and mid-range segments. Combined with halved NEV purchase tax subsidies and IoT contracting for two consecutive quarters due to reduced national subsidies, "premiumization to protect profits" and "shipment cliff" happened simultaneously. ASP hitting historical highs still couldn't sustain gross margins; the essence of this pricing power battle is cost transmission capability. Valuation anchors are shifting from "cost-effective volume" to "self-developed chips + premium pricing." Realization depends on whether Pengcheng locked-in orders convert to deliveries and whether Xuanjie can dilute dependence on upstream suppliers. Along the entire transmission chain, storage manufacturers and domestic substitution chains benefit, while all consumer electronics brands relying on low-to-mid-end shipments must reassess gross margins—this is a common bill for edge-side AI hardware.

---

IV. Macro, Capital & Market Data

**Century-Old Gold Jewelry Brand Steps on Delisting Line: *ST Cuihua Annual Report Delay Triggers Regulatory Delisting; Market Cap Below RMB 500M for 18 Consecutive Days, Stock Price Already Below Par Value**

Established gold jewelry retailer Cuihua Jewelry issued a risk warning announcement: Failure to disclose the 2025 annual report on time triggered regulatory delisting conditions, leading to termination of listing; original auditor Lixin Zhonglian suspended business operations for 6 months due to violations, forcing the company to change auditors, adding insult to injury. CSRC investigation has preliminarily found suspected false statements in some annual report financial information. Trading suspended from September 1. This is not a single event but a closing sample of the A-share clearing mechanism: The market cap delisting line has reached 18 trading days; the closing price on August 27 already broke below the RMB 1 par value; major illegal mandatory delisting conditions loom overhead—even timely annual report disclosure might not escape delisting. Of 455 stores, 442 are franchises; this heavy-brand-light-direct-operation structure means the joint liability risk between brand owners and franchise systems is seriously priced by the market for the first time. The signal of shell value going to zero points to all "time-honored brands" and small-cap tail companies.

• UK "Equity Ticketing" Platform Tickets for Good Raises £3.9 Million (Endorsed by Robbie Williams) — A leftover ticket distribution platform targeting healthcare workers, teachers, and low-income groups; nearly 750,000 members have cumulatively distributed 1.25 million tickets. Public funds and impact capital continue to double down, providing a replicable sample for the "algorithmic allocation of surplus capacity" public welfare business model.
• Flex IT and T1A Merge to Form European Circular IT Platform with ~€150M Revenue — Closed loop of recycling, certified erasure, refurbishment, and distribution through 13,000 dealers completed. Benefiting from EU electronic device regulations, circular IT moves from fragmented small businesses to M&A consolidation phase—"supply chains generated by compliance" is a stable theme in Europe's primary market.
• Stockholm Creator Marketing Platform Fluencify Completes Oversubscribed $4.3M Pre-Seed (Led by byFounders) — Uses AI Agents to automate influencer discovery, outreach, and settlement, covering 85 countries and 13,000 active creators. Another health-check sample for "Agents running complete marketing workflows"; the infrastructure layer of the creator economy continues to attract capital.
• Frontline Investors' Shanghai/Shenzhen Visit Recap: "AI Native" is Just an Adjective, "Entrepreneur" is the Noun — Judgment after deep chats with over 200 AI entrepreneurs: Successful products all obsess over engineering details—video generation improved from 4 clips in 3 minutes to 9 clips in 10 seconds, thanks to stacking 40+ optimization points. Counter-evidence to the sentiment that "AI applications are uninvestable": The window hasn't closed; only sloppy ways of working have.
• Market Snapshot — A-shares close on 9/7 (Monday): Optical modules surged on high volume (Zhongji Innolight +10.38%, Eoptolink +8.08%); Cambricon held above the RMB 1,000 mark (+1.90%, hitting stage highs); Foxconn Industrial Internet rose +4.52% in sympathy; Hygon Information and Kingsoft Office saw minor pullbacks; Domestic GPUs entered a differentiation phase priced by unlockings and earnings after Moore Threads plunged upon unlocking. Previous US market close (Friday 9/4; Monday 9/7 closed for Labor Day): Nvidia +0.84% bucking the trend, Meta +1.00%; Microsoft -2.04%, Tesla -5.92% led Magnificent Seven lower, Palantir -4.49%—the divergence between compute hardware and application software is widening; AI money is concentrating in "irreplaceable infrastructure." Tonight's Xiaomi launch event put "storage price hikes squeezing margins" on the table; the cost curve of edge-side AI hardware is the wind vane for the next main theme of the week.

---