---

Today's three main threads. First, the staggering scale of giants' compute bills: Meta once estimated spending $10 billion on Anthropic services this year. Second, regulators starting to reach inside models: NSA officials stated hope to access all AI models. Third, Chinese large model vendors collectively turning to free strategies: Alibaba and Moonshot AI are using scale to buy time. In US stocks, AI leaders pulled back collectively after earnings. Nvidia fell over 4.5% in a single day; Philadelphia Semiconductor Index plunged 3.5%. A-share compute chain cooled simultaneously. Changes in compute, regulation, and business models will continue to ferment in coming weeks.

I. Large AI Models

Alibaba and Moonshot AI Turn to Free Model Strategies, Grabbing Scale Before Discussing Charges

According to Bloomberg, Alibaba and Moonshot AI recently changed large model services to free, using low thresholds to exchange for user scale and real usage data, pushing commercialization to the backend. For developers, free means deployment costs instantly zero, opening windows for testing and migration. Domestic top model vendors generally shifted focus this year from competing on parameters to competing on implementation scenarios. Free is the fastest way to grab scenarios. Early players opened markets with open-source and low-price models; now it's turn for head giants to use free as customer acquisition tools. Amid cooling price wars, raising the free flag again indicates traffic competition remains the current focus. Whether this path works depends on backend service and ecosystem monetization efficiency. Another side of free strategy is pressure of compute costs transferred directly to resource utilization rates. Whoever has lowest idle rate survives longest.

Anthropic Researcher Demonstrates Model Self-Improvement Direction, Using AI to Train AI

An Anthropic researcher publicly shared research ideas on self-improving models, using other models to train models. This is a direction bet on by several frontier labs. Such "model critiquing model" loops have scattered internal practices across various companies. Difficulty lies in making the training loop truly self-rotate, not staying at evaluation stage. Disclosed content remains at lab stage, quite distant from product implementation. If it works, R&D efficiency could jump, and issues of credibility and alignment will be pushed to forefront. Head labs' allocation of compute and manpower to this direction this year already indicates its priority. Partners and regulators are watching such research. Once self-rotating loops establish, dependence on human data for model iteration will drop significantly.

• New Memory-Based AI Architecture MIRaS Published This Week. Paper published in academic journal Nucleic Acids Research proposes writing memory mechanisms directly into model structure, challenging "forgetful agents." Unlike common external memory plugins, this direction treats memory as native model capability. Once agents possess decent long-term memory, reliability of multi-turn tasks and cross-session collaboration improves significantly, addressing industry anxiety about agent implementation. Published in authoritative molecular biology journal, discussion of memory-based architectures is spreading interdisciplinary, itself a signal. Combination of memory and reasoning is viewed by many teams as watershed for next-gen agents.
• Meta Once Estimated Spending $10 Billion on Anthropic Services This Year. According to New York Times, Meta set annual Anthropic service budget at $10 billion (~67.3 billion RMB) in April this year, adjusting arrangements later due to external AI service price changes. Converted, this equals paying ~$2.5 billion per quarter for external model services. This figure explains source of Anthropic's high revenue growth and reveals true cost structure of parallel self-research and external procurement for giants. For Meta, such external compute bill only accelerates pace of self-research and open-source models. Any self-research progress cashes in directly on books.

II. AI Software

AMD Releases ROCm 10.0.0, Shifting Focus to AI Inference

AMD released GPU accelerated computing software stack ROCm 10.0.0 this week, focusing on inference performance, developer tools, and profiling on Instinct, Radeon, and Ryzen AI platforms. Update also expanded support for GPU virtualization. Software stack is key link in AMD's struggle with CUDA ecosystem, previously criticized as "usable but not easy to use." Accelerated versioning means it starts seriously courting developers in inference market, which happens to be fastest growing compute demand currently. Optimization of software stacks for inference scenarios directly determines performance per watt and unit cost, unavoidable street fight for AMD. Coverage of Radeon and Ryzen AI extends its software ambition from data center to consumer end.

• Self-Media Assistant Real-World Test: Produces 17 Finished Products During One Meal. Live materials from a press conference generated in real-time by AI. Press releases, long images, posters, highlight videos popping out one after another. While speakers talk on stage, photos start appearing off stage. Pace of content production rewritten entirely. Humans only responsible for final check and signature. Pipelines connected by such tools are transforming content middle offices from labor-intensive to compute-intensive. Last round was text; this round is video. Unit cost of AI-generated content continues dropping monthly.
• Minecraft Founder Persson Changes Mind. Notch, who publicly opposed AI programming, admits early judgment might be wrong. In January this year he wrote AI programming was terrible idea. He is industry veteran, watching tools iterate several versions yearly, finally changing stance. Such shifts not uncommon among senior engineers. This billionaire's attitude shift reflects collective turn of developer group this year: from resistance to adoption to dependency, process faster than most expected.
• Cisco President Defends AI in Classrooms. Patel stated in interview that AI tools can alleviate uneven educational resource allocation; key is solving usage thresholds for teachers and students first. From equipment manufacturer perspective, selling in classrooms isn't just devices, but entire AI teaching solution and underlying network bandwidth business. Controversy in education scenarios lies in data and screen time. Vendors need balance between commercialization and public responsibility. With extra teaching assistant in class, teacher focus shifts from knowledge transmission to guiding thinking.

III. Humanoid Robots

WRC2026 Observation: Power-Efficient Nervous Systems More Urgent Than Stronger Algorithms

Technical discussions at this conference shifted focus from brain to nervous system. Analysis points out if 1 to 10 billion robots deployed in future, energy consumption per unit and supply chain power capacity become bottlenecks. Compute can be stacked, but electricity may not suffice. Power consumption ratios of joint motors, sensors, and main control chips are being redistributed. Entire supply chain reprioritizing for energy optimization. Power-saving body design thus becomes more urgent topic than stronger algorithms. For startups, this is relatively fair starting line; giants' advantage in models dilutes fastest here. Measuring maturity of robot system, energy density becoming harder metric than spec sheets.

IV. Autonomous Driving

• Ireland Studies Using AI to Govern Wrong-Way Driving on Highways. Irish transport department evaluating AI vision systems to identify wrong-way vehicles and provide instant warnings, reducing severe accidents. Implementation pace of such scenarios often earlier than flashy urban driverless cars; regulatory boundaries clearer. Identification accuracy, false alarm handling, and liability boundaries are three hurdles before project landing. For suppliers, this is rare certain order for AI vision in transport infrastructure. Successful run provides replicable samples for other member states.
• BYD H1 Net Profit Attributable to Parent 12.325 Billion Yuan, Down 20.54% YoY. Revenue 344.815 billion yuan; gross profit 64.989 billion yuan. While sales stabilized, price war erodes profit margins; industry clearance continues. After electrification bonus peaks, intelligence becomes main bet for next leg. Pace of H2 smart driving solutions is observation focus. Scale advantage remains, but per-unit profit thinned by price war. Whether overseas and smart driving contribute incremental gains worth tracking.
• Nearly Half of UK Drivers Willing to Switch Next Car to Chinese Brand. Carwow survey shows ratio rose from 35% in H1 to 49%. Brand awareness climbs with sales. Intelligent configurations and cost-performance listed as most important reasons by respondents. Channel, insurance, and after-sales chains in European market adapting simultaneously. Acceptance of Chinese cars in Europe entering new phase.

V. Physical AI

• Viewpoint: Actual Lifespan of AI GPUs May Exceed Three Years. Market often judges compute bubble by "three-year depreciation." Practitioners argue with operational data that actual lifespan of inference GPUs longer than imagined. Depreciation assumptions directly relate to cloud provider pricing and financing capabilities. If premise loosens, valuation of existing compute assets and pace of new capacity construction need recalibration. In compute arms race, mispricing asset lifespan may be most expensive variable.

VI. Macro and Market Data

US Stocks Three Major Indices Close Slightly Lower; Chip Stocks Lead Decline

Dow 53559.99 (-0.02%), S&P 7711.76 (-0.25%), Nasdaq 26402.42 (-0.52%). Philadelphia Semiconductor Index plunged 3.47%. Nvidia fell sharply 4.57% to $217.55. Amazon bucked trend up 3.97%. Google, Microsoft, Meta closed red. Hawkish stance released again at Jackson Hole symposium; September rate hike probability rose to ~57%. Tech stock valuations under pressure.

A-Share Compute Chain Cools Simultaneously; Foxconn Industrial Internet Buckes Trend Red

Influenced by sentiment transmission from US chip stock pullback, A-share compute chain fell collectively. Optical module duo Innolight closed down 0.90%, Eoptolink down 2.47%. Cambricon dipped slightly 0.13%, oscillating narrowly above 1000 yuan. Hygon Information down 0.20%. Foxconn Industrial Internet bucked trend up 0.28%. Kingsoft Office up 1.29%. Funds shifted to segment leaders with performance support. Institutional views generally believe short-term pullback after earnings doesn't change medium-term direction of compute prosperity; divergence only in pace. Next phase focus on resonance between domestic chips and compute operations.

• NSA Officials State Hope to Access All AI Models. US National Security Agency deeply involved in government's new voluntary model testing program. Top officials stated hope to access all models. Data boundary between intelligence agencies and AI companies becomes focus of next round of game. Voluntary program starts; mandatory clauses may follow. Compliance costs for head model vendors