01MODEL RELEASES3 stories
OpenAI launches GPT-Live-1 full-duplex voice API: tiered pricing for voice agents
OpenAI官方+多媒体Confidence High
Front-end voice layer $0.05/minute, Tau3 voice intelligence benchmark 86.2% (previous generation Realtime-2.1 was 45.7%), answer latency 0.798 seconds (previous generation 1.41 seconds), can connect to backend inference models such as GPT-6 Astra.
Take · Voice agents' 'ears and mouth' and 'brain' are now officially priced in tiers—independent pricing for the voice layer means call-chain costs can be split apart and optimized.
Embodied open-source double release: Unitree humanoid foundation model + Ant world model open-sourced same day
宇树 / 蚂蚁灵波多源Confidence High
Unitree open-sources humanoid robot foundation model UnifoLM-WLA-1.0: a single model coordinates 64 tasks (54 desktop + 10 whole-body mobility), with weights/code/datasets fully open; it leads open-source models on 7 embodied benchmarks; Ant Lingbo open-sources three LingBot-World 2.0 world models, including 1.3B Small (targeting real-time generation on a consumer-grade single GPU).
Take · Open-source focus is shifting from language models to two new niches: 'embodied foundation + world models'.
NASA × IBM open-source lunar foundation model
NASA × IBM官方Confidence High
Ice-zone recognition error down 22% vs. SwinV2-B; about 2 million tiles dataset now on Hugging Face.
Take · Open-sourcing research foundation models is another public good for 'AI for Science'.
02PRODUCT3 stories
OpenAI opens Agents API public beta: hosted Codex harness, no API fees.
OpenAI官方Confidence High
On the same day, Agents API public beta opens with hosted Codex harness and no API fees.
Take · Turning the Agent runtime framework into a platform entry point further concentrates distribution power in leading platforms.
Customer service Agent pricing shifts to 'per verified result'.
Salesforce / Intercom / Zendesk多源Confidence Mid
Salesforce Agentforce $2/successful resolution (launching in June-July), Intercom Fin $0.99/result, Zendesk billed per verified resolution, echoing the domestic 9/6 'token pay-for-performance' framing.
Take · Billing units shift from seats/Token to outcomes, making outcome measurement and settlement audit tools a recognized gap.
ChatGPT Images 2.5's two versions sweep top two spots on LMArena image leaderboard
OpenAI / Google榜单Confidence Mid
9/8 image leaderboard: ChatGPT Images 2.5's two versions, sunburst 1421 / flare 1399, take the top two spots; Google Nano Banana 2.5 is in anonymous blind-test stage (rumor-level, not officially announced).
Take · Image-generation competition is entering a niche track where multiple versions of the same model are ranked separately.
03INDUSTRY9 stories
Oracle FY27Q1: cloud infrastructure revenue +121%, RPO reaches $664 billion
甲骨文官方财报Confidence High
Cloud infrastructure revenue +121% to $7.4 billion (expected $7.19 billion), RPO $664 billion (expected ~$639.9 billion), single-quarter new AI cloud contracts over $30 billion, quarterly capex $28.5 billion, full-year guidance $70 billion, after-hours at one point +8%.
Take · AI infrastructure enters an 'orders are king' validation phase: RPO is firmer than current-period revenue.
HBM shortage spills over into domestic compute price hikes: Ascend 950DT indicative price +20–50%
华为 / 寒武纪路透（匿名供应链）Confidence Mid
Reuters 9/10, citing anonymous supply chain sources: Huawei Ascend 950DT indicative price +20–50% vs. two months ago (over 250,000 yuan/card), Cambricon 690 up 20–30%, 950PR from about 60,000 to 80,000+, MetaX and Iluvatar CoreX follow (Huawei/Cambricon did not respond).
Take · HBM export controls and gray-market premiums are transmitting along the domestic supply chain, with domestic compute power taking over the price hikes.
PPI +5.4% beats expectations, rate-hike pricing rises to about 73%, 30Y breaks 5.36%
市场CME / 市场数据Confidence High
August PPI +0.4% MoM, +5.4% YoY, above expectations; CME FedWatch pricing for a 25bp hike on 9/16 rose from about 60% to 71.8–73%; 10Y U.S. Treasury closed at 4.96%, 30Y rose above 5.36% (highest since June 2007); the ECB raised rates by 25bp for the second time this year to 2.5%; Philadelphia Semiconductor Index -2.66%, NVIDIA -2.26%, AMD -3.36%, Micron -4.9%. August CPI will be released tonight at 20:30 (expected headline 3.4% / core 2.4%).
Take · Macro tightening and compute shortage are reinforcing each other, shifting AI pricing logic to a dual track of 'cash-flow validation + compute scarcity'.
Capital flows: Harvey raises $550 million at a $15.5 billion post-money valuation; Alibaba plans to lead a round in evaluation startup UniPat AI.
Harvey / 阿里多源Confidence Mid
Legal AI Harvey confirmed $550 million in funding at a $15.5 billion post-money valuation (a 41% premium over $11 billion in March), with ARR over $400 million and 80% of the Am Law 100 as its clients, moved from single-source pending on 9/9 to confirmed; Bloomberg (people familiar with the matter): Alibaba plans to lead a $300 million round in AI training/evaluation startup UniPat AI at about $2.5 billion post-money, with Tencent/HongShan China participating (not officially announced).
Take · Capital is flowing to the "evaluation / data" layer—referees are worth more than players.
DOJ reviews whether NVIDIA's $17 billion Groq deal is a de facto acquisition.
英伟达 / 美国司法部NYTConfidence High
NYT reports the DOJ sent NVIDIA a formal document request, reviewing whether last December's $17 billion Groq "non-exclusive license + talent acquisition" (founder Jonathan Ross and others have joined) was a workaround to avoid merger review.
Take · If a violation is found, the Q4 AI M&A wave and exit channels will systematically cool.
Enflame Technology listed on the STAR Market today (688801.SH).
燧原科技发行公告Confidence High
Issue price 142.18 yuan, subscription on 9/6 about 4073 times; Moore Threads/MetaX/Biren and other domestic AI chip companies have all completed capitalization.
Take · The "four little dragons" of domestic computing power have all gone public, with procurement and ecosystem cooperation channels opening simultaneously.
SpaceX compute hosting revenue about $1.1 billion per month
SpaceX高盛会议（未经审计）Confidence Mid
CFO disclosed at Goldman Sachs conference: a newly signed single compute hosting deal to bring about $1.1 billion per month from December (about $13.3 billion annualized); previously signed Anthropic at about $1.25 billion/month and Google at about $920 million/month; ground compute deployment to exceed 2GW by year-end.
Take · Compute leases are becoming a new long-term cash-flow asset class.
Shanghai Public Security issues 12.125 million yuan tender for an agent construction and evaluation platform
上海市公安局中国政府采购网Confidence High
9/9 announcement: tender for 'Agent Construction and Evaluation Tool Subsystem', budget 12.125 million yuan (bid opening 9/29), comprising a three-piece set of construction platform/development toolbox/evaluation tools; requires low-code orchestration, multi-tenant isolation, A2A protocol, and DeepSeek integration.
Take · First public large-value tender for a police-grade Agent platform, with 'evaluation tools' listed as an independent category.
South Korea's 10% penalty rule takes effect today; California signs SB 813
韩国 / 加州官方Confidence High
South Korea's revised Personal Information Protection Act takes effect 9/11: intentional or gross negligence repeat within 3 years, a single incident affecting 10 million+ people, or leaking again after failing to comply with a corrective order can incur a fine of 10% of global total revenue; suspected leaks must be notified within 72 hours; preventive investment can reduce penalties by up to 40%; California signed SB 813 on 9/9—the first US law for an independent third-party AI safety assessment framework (paired with AB 1405 auditor registration system).
Take · Global compliance enters the era of heavy penalties + independent assessment.
04RESEARCH3 stories
Terminal-Bench 4.0 leaderboard switch: old benchmark's inflation squeezed out
Artificial Analysis官方榜单Confidence HighProgress · Day 1
AA v4.3 leaderboard switch: Gemini 3.8 Flash on TB 4.0 fell from 87.6% to 19.7% (-67.9 points), Muse Spark 1.3 from 84.3% to 33.3%; only GPT-6 Astra (59.6%) and Fable 5.1 (55.1%) held the top (official leaderboard figures: Astra 58.18% / Fable 44.55%; some sources' 19.1% is a harness difference).
Take · Citing leaderboard scores must include version numbers; leaderboard switches directly hit narratives that rely on public scores for selection/procurement/financing.
Morgan Stanley AI Guidebook: Big Four cloud capex 2026→2028 three-step leap
摩根士丹利AI GuidebookConfidence High
Big Four cloud capex: 2026 917 billion → 2027 1.47 trillion → 2028 1.64 trillion USD; compute capacity demand 36 → 144GW (4x).
Take · Quantitative constraints on the compute supply side corroborate the HBM shortage and rising prices of domestic cards.
Jobs narrative diverges from evidence: AI has caused about 200,000 Americans to be laid off, may create about 1 million new jobs.
《经济学人》9/4 刊复盘Confidence High
The Economist analysis: AI has caused about 200,000 people in the US to be laid off, but may create about 1 million new jobs (data center electricians, cooling technicians, etc.); annual data center construction investment exceeds $75 billion, up +60% YoY.
Take · Layoff narrative diverges significantly from evidence on net job creation, a key focus for the next 3 months.
05INSIGHTS5 stories
Today's main theme: pricing logic shifts from 'narrative-driven' to 'cash flow validation + compute scarcity'.
前瞻深度版研判Confidence Mid
Macro tightening (August PPI +5.4% YoY, above expectations; FOMC rate-hike pricing at about 73%) resonates with a severe compute shortage (HBM scarcity pushes domestic-card indicative prices up +20–50%); Oracle proves orders are king with $664 billion in RPO, while AI hardware is first to squeeze out valuation froth amid a high-rate pullback.
Take · The same-day occurrence of “order realization” and “valuation squeeze” is a classic sign of a regime-shift period.
Performance-based billing settlement audit tools jockey for position
行动深度版建议Confidence Mid
Build Agent middleware for “outcome metering + settlement audit” (success-rate verification, dispute arbitration, reconciliation SDK), benchmarked against Salesforce $2/resolution and domestic token pay-for-performance framing; path: first connect 1-2 customer-service Agent scenarios to build a metering prototype → abstract into a general billing protocol → seek platform-side certification.
Take · Time window 9/12–11/30 (positioning period as the industry switches billing models); metering tools naturally double as “outcome notarization” and can be bundled with evaluation capabilities.
Evaluation-layer assetization: vertical leaderboards + data moat
行动深度版建议Confidence Mid
Leverage Alibaba’s 300 million investment in UniPat and TB 4.0’s leaderboard change exposing old leaderboard froth to build industry vertical evaluation sets (government/enterprise, customer service, code) and continuous leaderboard-refresh services; path: first publish 1 credible vertical leaderboard to build a brand → monetize via “evaluation dataset subscription + enterprise private evaluation.”
Take · Time window 9/12–10/15 (valuation narrative strongest within the capital-revaluation window); Shanghai Public Security’s tender listing “evaluation tools” separately is demand evidence.
Hedging compute price hikes: price locking and inference cost reduction
行动深度版建议Confidence Mid
With domestic cards +20–50% and NVIDIA +15% in 2027, immediately assess price locks on existing contracts, lease-to-own, and inference-side quantization/distillation compression needs; path: inventory Q4 compute gap → negotiate price locks within September → hedge with inference optimization projects.
Take · Time window 9/12–9/30; a one-month delay in requesting quotes may cost an extra quarter's budget.
Verification criteria note: 28 items verified, 15 accepted / 8 corrected / 3 excluded / 2 deferred
提示深度版核验Confidence Mid
Shujianzhi today verified 28 items: 15 accepted; 8 corrected (NASA×IBM ice-zone error is 22%, not 23%; 30Y is highest since 2007, not 2004; DOJ review targets and onboarding personnel scope; Agents API 'restricting competitor ads' removed for lack of source; Morgan Stanley capex follows the four major cloud providers' scope, etc.); 3 excluded; 2 deferred (China Southern Power Grid 728.6 万 split has no public disclosure page; Musk G20 timetable single-source).
Take · When citing today's data, use the corrected criteria; single-source/unofficial items are flagged in confidence.
Ecosystem Pulseself-updating
News278Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Performance-based billing settlement audit tools jockey for positionBuild Agent 'outcome metering + settlement audit' middleware (success-rate verification, dispute arbitration, reconciliation SDK); first integrate 1-2 customer-service Agent scenarios for a metering prototype, then abstract into a general billing protocol and seek platform-side certification.
window 9/12-11/30 2. Evaluation-layer assetization: vertical leaderboards + data moatPublish 1 trusted industry vertical evaluation set (government/enterprise/customer service/code) to build brand, then monetize via 'evaluation dataset subscription + enterprise private evaluation'; combine with metering tools into a 'metering + evaluation' dual engine.
window 9/12-10/15 3. Hedging compute price hikes: price locking and inference cost reductionTake stock of Q4 compute gap; negotiate price locks and rent-to-own by September; simultaneously add inference-side quantization/distillation compression requirements to hedge price hikes for domestic chips and NVIDIA.
window 9/12-9/30