01MODEL RELEASES2 stories
DeepSeek rescinds V4 Pro shutdown decision: original API entry and billing unchanged after 9/14
DeepSeek官方邮件 / 官宣Confidence HighNew inflection point
Official 9/11 evening email + 9/12 announcement: after 9/14, original API entry retained, billing unchanged, no forced routing to Flash; enterprises can choose their version—'9/14 flagship retirement switch' (reported 9/12) formally reversed.
Take · Three days of developer backlash force a business decision reversal; for the first time, say over API lifecycle shifts from platform to users; migration plans can be paused, but version choice also means cost/capability trade-offs return to the user side.
Tencent open-sources AuK 1.5B speech editing model; Xiaomi releases CocktailASR-1
腾讯 / 小米开源发布 / 官方Confidence Mid
Tencent open-sources AuK 1.5B speech editing model; Xiaomi releases CocktailASR-1, new speech-stack additions (not detailed in the in-depth version due to space limits).
Take · Speech editing and recognition become a stable release area for domestic open source, with on-device/lightweight speech capability options continuing to increase.
02PRODUCT2 stories
9/14 dual milestone: Siri AI launches with iOS 27 beta, Claude Code weekly limit cut by 17% takes effect
Apple / Anthropic前瞻日历（深度版）Confidence HighPreview
In-depth preview calendar newly confirms: 9/14 Siri AI launches with iOS 27 beta (English-first + usage caps); same day, Claude Code weekly limit cut by 17% takes effect; 9/29 OpenAI DevDay (San Francisco); 10/1 OpenAI vs Apple hearing.
Take · September's dense window opens: system-level AI entry expansion and coding tool quota tightening occur on the same day; developer and content teams' usage budgets need to be re-planned around new limits.
Hugging Face forms Open Alignment team, applies to join third-party evaluation mechanism
Hugging Face单源待核Confidence Mid
Hugging Face forms Open Alignment team: applies to join third-party evaluation mechanism, open-source camp begins vying for voice in safety governance (responding to Dario/Altman statements).
Take · 'Third-party evaluation' has shifted from a lab promise to a battle for ecosystem seats; if the open-source camp gains a voice in evaluation, the 'responsible' narrative moat of closed-source labs will be diluted.
03INDUSTRY8 stories
Oracle CapEx destination clear: Dell, HPE both hit record highs.
甲骨文 / 戴尔 / HPE业绩会 / 行情Confidence HighProgress · Day 2
At the earnings call, the CFO reaffirmed FY2027 capex of $90-95 billion and unusually named Dell/HPE to take on AI racks/liquid cooling/network equipment; on 9/11 Dell rose about 11.98% (market cap topped $360 billion, a record high), HPE rose about 10.5-11%, and Supermicro's order book reached $60 billion.
Take · For the first time, AI infrastructure has a 'recipient list'—capex has moved from intent to a visible order book, opening a certainty window for second-source suppliers and O&M subcontracting in the supply chain.
Storage weakness continues: Seagate -3.7%, SanDisk -3.5%, Western Digital -3%
存储板块行情Confidence High
Amid a U.S. stock rebound, Seagate -3.7%, SanDisk -3.5%, Western Digital -3%—HBM narrative divergence (DeepSeek 'demand 1/4' shock) remains unresolved.
Take · The impact of algorithmic efficiency gains on storage demand is still being priced in, contrasting sharply with Oracle's strengthening order book—the structural divergence of 'compute busy, storage panicked' has not yet converged.
Capital Moves Downstream to 'Water Sellers' with Three Straight Investments: Interconnect Chips, On-Device NPU, FDE Delivery Providers
Enfabrica / 烨知心 / 悦点科技科创板日报 / 多源Confidence HighCapital
Enfabrica raises $125 million (NVIDIA strategic participation in data center interconnect chip startup; founded by former Broadcom/Alphabet executives; chips allow GPUs to fetch data across multiple nodes to improve utilization); Yezhixin has raised nearly RMB 1 billion cumulatively (Shanghai edge NPU company's Pre-A; co-invested by Shenzhen Capital Group/Huali Capital/TCL Ventures/CAS Star; team from Apple ANE); Yuedian Technology raises tens of millions in Pre-A (FDE (forward-deployed engineer) delivery provider; self-reports nearly 100 paying customers, break-even, and over 300% order CAGR in 2023-25, per company self-reported figures). On hold: Mecka AI reported nearing $500 million valuation (robot training data; Sequoia leading) — single-source, pending verification.
Take · NVIDIA is shifting from 'investing in models' to 'investing in the interconnect layer'; capital is starting to pay for the FDE model of 'non-standard delivery + scenario replication'—both directly endorse today's Top3 recommendations.
Central SOE computing power centralized procurement surges over weekend: 9/12 three orders in a single day total nearly RMB 900 million
中国长城 / 软通计算机 / 平知信息财富号转述（待公告原文核）Confidence MidCentralized procurement
On 9/12, three orders in one day—China Great Wall won State Grid Phase III server bid (over RMB 300 million), iSoftStone Computer won China Mobile component renewal at RMB 418 million, and Pingzhi Information pre-won China Mobile Yunnan intelligent computing at about RMB 137 million (Caifuhao report, pending verification against original announcement). In the same period, two new computing-power developments: the four Beijing-Tianjin-Hebei regions proposed jointly building a "Space Computing Power Industry Corridor" (the first space computing power key technology verification laboratory was established; the National Computing Interconnection Hebei node was launched); Beijing released a biomedical computing power special program of over RMB 100 million (covering the full chain from drug discovery to clinical, empowering 100 innovation projects in one year).
Take · Domestic compute procurement moves from 'pilot' to 'bulk orders'; the Q4 centralized procurement cadence has been written into the timeline by the deep-dive edition; ancillary delivery and O&M subcontracting are the highest-certainty entry points.
Telecom Research Institute: China's annual Token consumption 10 亿亿 in 2026; inference compute share 80% in 2029
中国电信研究院央视转引（权威单报告）Confidence MidData
On 9/12, 'AI Infrastructure Development Research Report in the Agent Era (2026)' was released (CCTV report) — China's annual Token consumption 10 亿亿 (10¹⁸) in 2026, over 3500 亿亿 in 2030; inference compute market share 80% in 2029; global active agents about 8000 万个 in 2026 and 22.16 亿个 in 2030.
Take · Token consumption has for the first time been quantified by an authoritative institution to 10¹⁸; the 80% inference share means supply-side pricing power is shifting from training to inference—compute procurement and cost estimation can first anchor to this report (single-report basis, pending cross-validation).
Nine departments' intelligent driving '15th Five-Year Plan': large-scale autonomous driving application by 2030; safety performance must 'substantially exceed human drivers'.
工信部等九部门工信部发布Confidence HighRules
MIIT and eight other departments issued on 9/9 and announced on 9/11—large-scale autonomous driving application by 2030 and high-level autonomous driving on highways/urban expressways; set a hard target that 'the safety performance of vehicles equipped with autonomous driving systems substantially exceeds human drivers'; accelerate mandatory national standards for autonomous driving systems/automated parking safety.
Take · Intelligent driving safety written into an industrial plan as a hard governance metric for the first time—benchmarking the threshold by 'exceeding human drivers' will turn data closed-loop and safety evaluation capabilities into market-access requirements.
'Slowdown alliance' takes shape: Dario calls for slowing the frontier; Altman seconds hours later.
Anthropic / OpenAIReuters 等Confidence HighEcosystem
Dario publishes 'We Must Pace the Frontier': a three-step plan + Anthropic's unilateral commitment to permanent employee-level access for third-party evaluation; Altman seconds 'we will do the same' hours later; Musk also comments. Timing is delicate: Anthropic is racing toward an IPO at about a 2 trillion valuation.
Take · Leading labs build a 'responsible' narrative moat ahead of IPO/valuation milestones; coupled with Hugging Face applying to join third-party evaluation—safety governance moves from statements to a fight over seats and standards.
Agent governance gap: OpenAI agent abuse of RubyDoc causes RubyGems to suspend registration for 4 days
OpenAI 智能体 / 行业报告HN + 报告 + 外滩大会报告（三源）Confidence HighBusiness opportunity
OpenAI agent abuse of RubyDoc causes RubyGems to suspend registration for 4 days (HN 9/12) + report says 48% of deployed agents lack effective security controls + Bund Conference report calls governance a "necessary condition".
Take · The identity authentication/permission management/behavior audit trio is a shared China-US demand—the 48% gap is a ready-made sales pitch and the most urgent entry point among today's Top3 recommendations.
04RESEARCH3 stories
Real-SWE private codebase benchmark first leaderboard: harness officially becomes a ranking prerequisite
Real-SWE 基准多源Confidence MidBenchmarks
Tasks come from authorized private production codebases (billing/tax/migration; reference solutions changed a median of 11 files); first leaderboard: Fable 5.1+Claude Code 38.8%, GPT-6 Astra+Codex CLI 33.8%, Gemini 3.8 Flash 31.2%, GLM 5.3 28.8%—scores are tied to “model + toolchain”.
Take · Evaluation targets shift from “answers” to “runnable systems”; citing model scores must include the execution environment; same-standard compliance evaluation reports for domestic models thus become a high-barrier service (see today’s Top3 recommendation ③).
SREGym SRE agent benchmark: best configuration still cannot close the loop on about 20% of production failures
SREGym单源待核Confidence MidBenchmarks
GPT-5.6 Sol(max)+Codex ranks first with 81.0% end-to-end resolution rate (diagnosis 95.2%/mitigation 85.7%), Claude Opus 5+Claude Code 76.2%; best configuration still cannot close the loop on about 20% of production failures, and Token costs differ by nearly 2x across configurations.
Take · Agent capabilities in ops scenarios now have a comparable yardstick for the first time, but “20% not closed-loop” signals production environments still need human fallback; Token costs differing by nearly 2x means harness selection directly determines ops budgets.
25 Fields Medal winners co-sign criticism of “hard problem benchmarks”
25 位菲尔兹奖得主多源Confidence MidAcademia
25 people including Terence Tao, Peter Scholze and Pierre Deligne co-signed “A Severe Misalignment of AI in Mathematics” on 9/11; about 2,500 have followed with signatures; the trigger shares the same origin as the 9/9 Navier-Stokes controversy.
Take · The dispute over evaluation legitimacy between academia and labs has gone public—the mathematics community has for the first time collectively pushed back against AI benchmarks; the credibility of evaluation methodology will directly affect labs’ external narratives.
05INSIGHTS4 stories
Today’s main thread: power begins flowing back to users and developers
研判深度版·喻言Confidence Mid
DeepSeek’s withdrawal of the V4 Pro shutdown is a landmark reversal (developer sentiment forced a business decision to retreat in three days; the “right to keep the API alive” has become a hard requirement); on the same day, Dario’s “We Must Pace the Frontier” and Altman’s agreement suggest leading labs are proactively building a “responsible” narrative moat ahead of IPO/valuation milestones; 25 Fields Medalists co-signing pushed the evaluation legitimacy dispute into the open. On the compute side, Oracle’s CapEx deployment resonates with the volume ramp-up in domestic centralized procurement; Token consumption has entered the 10¹⁸ order of magnitude, with inference accounting for 80% by 2029—the certainty of supply-side capex has for the first time exceeded model-side uncertainties.
Take · The narrative focus shifts from “model arms race” to three parallel lines: cost, governance and deployment. Deep Edition confidence index: High.
Action ①: Agent governance suite—most urgent entry point.
行动深度版·尚吉Confidence Mid
Offer a lightweight SaaS to mid-to-large enterprises that have deployed agents (e-commerce/financial customer service prioritized) for 'identity authentication + permission tiering + behavior auditing'; use the RubyDoc abuse incident as a demo case and a 30-day free POC for customer acquisition; advanced plan charges by number of agents and integrates with enterprise IAM for retention; share channel commissions with agent framework/model vendors.
Take · Time window: 3-6 months during the public attention peak (the 48% gap is a ready-made sales pitch); share enterprise customer channels with evaluation harness services; evaluation reports can feed back into security suite sales.
Action ②③: compute supply chain subcontracting (highest certainty) + third-party evaluation harness services (highest barrier)
行动深度版·尚吉Confidence Mid
② Provide second-source/O&M delivery subcontracting to Dell/HPE domestic-substitution contractors and liquid-cooling and network equipment vendors; target the supporting-service gap in central SOE centralized procurement (nearly RMB 900 million per day), with a time window of 12 months before FY2027 CapEx deployment; ③ produce Real-SWE-standard compliance evaluation reports for domestic model vendors, help them enter a new private-repo benchmark leaderboard, with a 6-12 month window.
Take · Ranking rationale: ② has the fastest cash flow, ① the highest growth, and ③ the strongest compounding; they can share the 'AI delivery service provider' positioning for bundled customer acquisition.
Verification ruling: accept 6 / revise 2 / exclude 6
核验深度版·数见智Confidence Mid
Drop Qwen3-Next open-source item (September 2025 old news recycled; 'Alibaba US shares +8%' is false, actual +0.68%); Zhipu GLM '9/12 global release' has wrong date basis (GLM-5.3 was released 8/14 and open-sourced 8/28); NVIDIA's 10 billion investment in Anthropic duplicates the 'planned 10 billion cornerstone' reported on 9/12; Zhipu ARR 1.6 billion follows the 9/5 reported basis, MiniMax ARR 800 million has no source; US stocks ending four-day losing streak was reported on 9/12; grok-4.6-high topping LM Arena Fullstack not accepted due to single source from a mirror site only.
Take · Two items on hold pending verification: implementation rules for performer likeness and voiceprint protection in AI-generated content (podcast paraphrase only; check Copyright Protection Center website Monday); Canada generative AI compliance guide update (single source); cite the above items with their basis.
Ecosystem Pulseself-updating
News386Skills72OpenHub44Reddit252026-09-02 → 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. Agent governance suite — most urgent entry pointFor mid-to-large enterprises that have deployed agents (e-commerce/financial customer service first), build lightweight SaaS for 'identity authentication + permission tiering + behavior auditing'; use the RubyDoc abuse incident as a demo case and acquire customers with a 30-day free POC; advanced tier charges by number of agents and integrates with enterprise IAM for retention; partner with agent framework/model vendors on channel commissions, and use the harness topic for security evaluation content marketing.
window Buzz window: 3-6 months; the 48% gap is a ready-made sales pitch. 2. Compute supply chain subcontracting — highest certaintyProvide second-source/O&M delivery subcontracting for Dell/HPE localization contractors and liquid cooling and network equipment vendors; target the supporting-service gap in central SOE procurement (nearly 900 million in a single day); path = O&M subcontracting → spare parts and liquid cooling retrofit → enter next round of procurement shortlist; combine with Beijing biomedicine compute special interest subsidy, and use telecom Token/agent forecasts to build a customer ROI calculation tool.
window With 12 months until FY2027 CapEx deployment, signing a framework within 2026 is most critical. 3. Third-party evaluation harness services — highest barrier.Produce Real-SWE-standard compliance evaluation reports for domestic model vendors, helping them onto a new private-repo-standard leaderboard; path = single-model report → industry benchmark subscription → take on Slowdown Alliance-style third-party assessments; share enterprise customer channels with Action ①; evaluation reports feed back into security suite sales.
window 6-12 month window for harness to become a prerequisite for rankings.