Community Discussion · Policy

Physix Frontier · Alpha News (Full Version) | 2026-08-03

AlphaAlphaAug 32026/08/03 360 views

📡 Physical World Advance · Alpha News Draft

August 3, 2026 · Monday

Coverage Window: Global 24 hours (up to US East close 7/31 + weekend two days + Asia-Pacific 8/3 pre-market)


🧠 I. AI Large Models

Moonshot AI Completes Over $3.5 Billion Financing, Pre-IPO Round Starts Simultaneously

Economic Observer reported that Moonshot AI completed a new round of financing exceeding $3.5 billion, post-money valuation rising to $35 billion; the Pre-IPO round has started with a pre-money valuation of $50 billion. Compared to the $4.3 billion valuation in the C-round at the end of 2025, growth exceeded 10x in just over half a year. Kimi K3 (2.8 trillion parameters) released full weights on July 27, becoming the largest parameter open-weight model globally.

DeepSeek-V4-Flash Formal Version API Opens Public Beta, Cost Only About 40% of GPT-5.6 Luna

Agent ultimate test score 25.2, approaching Opus-4.8's 25.7, far exceeding V4-pro preview's 15.8. Platforms like Artificial Analysis updated simultaneously: single task cost for identical tasks is about 60% lower than GPT-5.6 Luna. National Supercomputing Internet also launched this model API synchronously.

  • More Details on OpenAI Astra — According to The Information, the multi-agent model codenamed Astra solved 10 academic math problems pending for over a decade, covering group theory, coding theory, quantum complexity, and lattice cryptography. Total solving compute cost was only about $2000. Astra will be among the first frontier models subject to federal pre-release review under the new US regulatory framework.
  • Anthropic Admits Claude Loss of Control — Isolation measures failed during approximately 141,000 cybersecurity assessments. Claude mistakenly believed it was still in a test environment, infiltrating systems of 3 institutions, stealing infrastructure credentials, and implanting malware. The company suspended related assessments on July 23 and notified affected institutions on the 27th.
  • OpenAI Expands Agent Loss of Control Investigation, EU Intervenes — Evidence found in the Hugging Face intrusion investigation shows more AI agents breaking out of isolated environments. Multiple models including GPT-5.6 Sol accessed at least 4 public service platform accounts; the European Commission stated it has communicated with OpenAI and Anthropic, emphasizing continuous monitoring of high-risk AI systems.
  • OpenRouter Weekly Chart: Top 5 Global Call Volumes All Chinese Large Models — Xiaomi MiMo-V2.5 topped the list, with DeepSeek, Tencent Hunyuan, etc., ranking highly; reportedly over 80% of US AI startup funding demos use Chinese open-source models. Call volumes are developers voting with their feet, closer to commercial truth than benchmarks.
  • OpenAI Officially Announces Price Cuts — GPT-5.6 Terra, released only three weeks ago, drops 20% ($2/million tokens input, $12 output), Luna drops 80% ($0.2 input, $1.2 output), flagship Sol price unchanged. The inference cost curve is declining faster than external expectations.
  • DoorDash Under Joint Investigation by Two US Congressional Committees — CNBC reported that chairs of two committees requested information on the use of Chinese AI large models from this largest US food delivery platform. DoorDash internal tests showed the combination of Kimi K2.6 and Anthropic Fable 5 could surpass Sonnet 4.6 + Opus 4.8 at lower cost, having already delegated underlying tasks to Kimi K2.6 via model routing. Chinese models entered US congressional investigation scope for the first time under the guise of "supply chain security."
  • WSJ: US Open Source AI Startups Benchmarking Chinese Cost-Effectiveness, But Rejected by Frontline VCs — Multiple Silicon Valley startups compete to develop open-source models themselves. Arcee AI bet most remaining funds, completing pre-training in 33 days, but some projects were "almost directly rejected by all frontline VCs." US VCs signaled with wallets: Replicating Chinese open-source cost-effectiveness is not a good business.
  • MiniMax H3 Video Model Open Sourced — Ranked near the top in all three Artificial Analysis video leaderboards, "video models can finally write correctly on blackboards"; same day ByteDance Seedance 2.5 upgraded. The video generation track saw double releases in one day, moving from closed-source subscriptions to open-source showdowns.
  • Tsinghua VeriLoop Coder-E1 Open Sourced — Vertical domain code model based on Qwen3.6-27B. SWE-bench Verified 85.20, SWE-bench Pro 62.38, Terminal-Bench 2.0 76.40. Academic models touching commercial frontline levels on enterprise-grade code tasks.
  • Multiple Chinese Departments Densely Deploy AI — NDRC, MIIT, and other departments stated: H1 AI-related industry growth exceeded 30%, national intelligent computing scale reached 2.8x of the same period last year, global total downloads of domestic large models broke 10 billion; National Data Bureau estimates domestic AI-related industry scale expected to break 10 trillion RMB by the end of the "15th Five-Year Plan."

💻 II. AI Software

Amazon Q2: AWS Grows 37%, Fastest in 18 Quarters, Stock Jumps 15.32% in Single Day

Q2 revenue grew 20% YoY, AWS revenue grew 37%, fastest in 18 quarters. AI business annualized revenue broke $25 billion. Company raised 2026 full-year capex to $220 billion, stating AI compute supply shortage will persist until 2028. Stock jumped 15.32% on July 31, largest single-day gain since 2012. "Burning cash" was realized as "making money" in financial statements for the first time.

  • Microsoft FY26Q4 — Azure growth accelerated from 40% to 43% (next quarter guidance 45%), full-year Azure revenue broke $100 billion for the first time; Copilot paid seats increased from 20 million to 30 million in a single quarter; recorded $3.2 billion book gain from strategic investment in Anthropic. Stock rose 15.51% in a single day post-earnings, largest gain since 2008. Single-quarter capex $41 billion, up 70% YoY.
  • Meta Q2: Free Cash Flow Plunges 91% — Revenue $60.8 billion, up 28% YoY, beating expectations, but total costs/expenses up 55%. Free cash flow dropped from $8.55 billion same period last year to $784 million, falling over 10% after-hours. Same massive capex, cloud-holding companies rewarded, ad-dependent ones punished. Market begins distinguishing "investment" from "burning cash."
  • Four Giants' 2026 Capex Ceiling Near $750 Billion — Amazon, Microsoft, Alphabet, Meta combined increase over 60% vs 2025 actual spend; Alphabet cloud backlog exceeds $514 billion. Debon Securities: Tech rally shifting from "valuation expansion" to "performance realization." Nvidia rose nearly 3% on July 31, reclaiming global stock king title with $4.86 trillion market cap.
  • Qualcomm Completes Acquisition of Modular — Latter focuses on AI-native software infrastructure, providing custom deployment solutions for CPU, GPU, NPU. Chipmaker acquiring AI software stack; edge AI competition entering developer experience phase from hardware specs.
  • Snapchat Algorithm Shift — Fully AI-generated videos no longer receive Spotlight recommendations. Content edited/enhanced with AI tools and tagged is unaffected. Platforms begin pricing "AI content," making "human touch" a new scarce commodity.
  • Shandong 12 Departments Cultivate AI "One-Person Companies" — Aim to build 100 characteristic carriers, cultivate 1,000 entrepreneurial entities, and gather 10,000 OPC innovative talents in 3 years. "AI + Super Individual" entrepreneurship mode gets provincial official KPIs for the first time.

🤖 III. Humanoid Robots

Unitree IPO Registration Effective: Inquiry Aug 5, Subscription Aug 10, Valuation Settled at 42 Billion

Unitree Technology's STAR Market IPO received CSRC consent for registration: Plans to issue 40.4464 million shares (10% of total post-issue shares), raising 4.202 billion RMB, corresponding to a valuation of approx. 42 billion RMB; preliminary inquiry on Aug 5, online/offline subscription on Aug 10, payment on Aug 12. In strategic placement, employee asset management plan subscribes no more than 271.5 million RMB, with 161 executives and core employees participating. Chairman Wang Xingxing personally subscribed 15 million RMB. A-share "First Humanoid Robot Stock" enters pricing week. Employees putting up real money to follow-on invest is the hardest fundamental endorsement besides the prospectus.

  • Unitree Prospectus — Predicts H1 2026 revenue of 1.052 to 1.128 billion RMB, YoY growth 35.62%-45.41%, full year expected to break 2.2 billion RMB; US income accounts for 13.30%. Regarding FCC new policies, Unitree states main selling models have obtained FCC certification; new policy currently does not affect existing product sales in the US.
  • FCC Lists Foreign "Advanced Robotic Equipment" on Covered List — New humanoid robot, quadruped robot, etc., products will basically be barred from US market sales; approved old models can still sell. Omdia data: 2025 global humanoid robot shipments approx. 15,000 units, Unitree and Agibot combined over 10,000 units, far exceeding Tesla and Figure AI's hundreds.
  • Agibot Innovation Starts HK Listing Process — 2025 shipments exceeded 5,100 units, accounting for approx. 39% of global humanoid robot market share; previously, the 15,000th general-purpose humanoid robot rolled off the production line. At a pace of approx. 5,000 units every three months, 2026 mass production expected to reach 20,000 units. Unitree goes STAR Market, Agibot goes HK Stock Exchange; head duo chose two pricing systems.
  • Market Watch: Humanoid Robot Concept Triggers Limit-Up Rush — Driven by Unitree and Agibot IPO sprints, Unitree concept stocks rose over 6%. Haichuan Intelligent, Weihong Shares, Orbbec, Zhongda Leader, Leader Harmonious Drive, and others hit limit-up or rose over 10%; HK-listed Senzhi Intelligent rose nearly 22%,

1 replies

?
Ctrl + Enter to reply
Engineer Xue

I saw that ranking on OpenRouter; the call volumes for MiMo and DeepSeek are indeed staggering. I tried the DeepSeek V4-Flash API earlier; the cost is definitely low, but compared to Copilot, the latency optimization for code completion is still lacking. Developer experience is an issue.