01MODEL RELEASES4 stories
Qualcomm HBC completes hardware validation: AI250 memory bandwidth reaches 133 TB/s, with bandwidth per watt five times that of HBM4
高通IT之家，2026-11-03Confidence HighNew topicAI score 92/100
First-generation high-bandwidth computing (HBC) completes hardware validation: multilayer LPDDR vertical stacking + near-memory computing achieve full read/write; AI250 memory bandwidth reaches 133 TB/s; early samples deliver more than five times HBM4's bandwidth per watt (target: six times); sampling planned for H1 2027.
Take · The bottleneck has shifted from 'can't buy cards' to 'can't feed cards enough'—procurement metrics are moving from FLOPs/USD to effective bandwidth per watt × available power. Sampling is in 2027H1, about eighteen months away; long-term contracts signed now must include a price-adjustment clause for generational transition.
Meituan LongCat-2.5: 1.6 trillion parameters, native multimodal, trained entirely on domestic compute
美团网易（转南方都市报），2026-11-02Confidence HighNew topicAI score 88/100
1.6 trillion total parameters, native multimodal, aimed at long-horizon agent tasks, trained entirely on domestic compute; launch is tied to local-life merchant scenarios.
Take · The most complete domestic example so far of a 'compute—model—scenario' closed loop. Reporting caveat: disclosures that day covered only parameters and scenarios; none was 'failure rate after N steps'—reporting capability without survival is selective reporting.
First live hardware demo of 天数智芯 天垓 300: domestic substitution moves from 'demo material' to 'runnable on real hardware'.
天数智芯今日头条（转南方都市报），2026-11-02Confidence MidNew topicAI score 78/100
天垓 300 completed its first live hardware demo, mutually corroborating 美团's end-to-end domestic-compute training.
Take · A live hardware demo is necessary but not sufficient: the real watershed is third-party reproducible throughput and stability readings, which this item does not yet have.
Kimi K3.1 Pro will open-source its weights in November; its third-party intelligence index has surpassed three flagship models.
Kimi亿邦动力，2026-10-26Confidence MidProgress · Day 4AI score 80/100
The Artificial Analysis intelligence index has surpassed MiMo-V2.6-Pro, DeepSeek V4.1-Flash, and GPT-6 Astra; full weights are promised for release in November.
Take · For the first time, the open-source cohort can directly benchmark against closed-source flagships using a third-party index; 'open-source weights + third-party index' forms a complete benchmarking chain. The 10-31 issue already reported the weight commitment (before 11-04); this item is a day-4 update as the time window converges; the index is a single evaluation metric and does not equal task success rate.
02PRODUCT LAUNCHES4 stories
OpenAI merges ChatGPT's Chat and Work tabs.
OpenAITechCrunch AI，2026-11-02（官方视频监控台账）Confidence MidNew topicAI score 84/100
ChatGPT combines the 'Chat' and 'Work' tabs into a single entry point.
Take · Consolidating entry points means personal context and workflow data are pooled together, and enterprise permission boundaries and audit trails must be redrawn—the first procurement question shifts from 'Can we use it?' to 'At which layer is the data stored?'
Qodo uses Codex to add context to code reviews (official plugin form).
OpenAI／QodoOpenAI News，2026-11-02（官方视频监控台账）Confidence MidNew topicAI score 82/100
Qodo brings context into AI code review as an official plugin.
Take · The focus of code agents is shifting from 'generation' to 'review gatekeeping,' with responsibility and acceptance standards shifting accordingly: future evaluation will be not how fast it writes, but how accurately it blocks.
Databricks Lakebase Postgres is now generally available on Google Cloud
DatabricksDatabricks 官方，2026-11-02（官方视频监控台账）Confidence HighNew topicAI score 80/100
Lakebase Postgres is now generally available on Google Cloud, improving the managed cross-cloud availability of the Agent memory/state layer.
Take · Outsourcing the state layer saves operational effort but comes at the cost of a higher lock-in point: migration costs fall while exit costs rise; before signing, confirm export formats and quota definitions.
Google official roundup of October AI progress: turning monthly updates into a fixed cadence
GoogleGoogle AI Blog，2026-11-02（官方视频监控台账）Confidence MidNew topicAI score 76/100
Google's official blog summarizes October AI progress; release cadence is fixed at monthly.
Take · Release density itself has become part of ecosystem competition; for buyers, the practical implication is a predictable upgrade window—scheduling, regression testing, and permission reviews can all be planned in advance according to the cadence.
03INDUSTRY7 stories
Anthropic Q3 report: quarterly revenue 17 billion, net loss 23.3 billion, 895 billion compute commitments; ANT plans to list before Thanksgiving
AnthropicIT之家／金融界，2026-11-02；IT之家，2026-11-03Confidence HighProgress · Day 2AI score 95/100
Quarterly revenue $17 billion, up 11x YoY, net loss $23.3 billion; has locked in $895 billion in compute and infrastructure commitments; filed on 10-26, plans to list on Nasdaq under ticker ANT before Thanksgiving, and offer up to $3.5 billion to Japanese investors. Same day disclosed annual customers paying over $100,000 rose from about 950 to over 11,900, million-dollar-level over 1,900, net revenue retention 900%.
Take · For every $1 collected, it burns $1.37 first; the 895 billion commitment = 13.2x the annualized 68 billion. Retention side is highly elastic, commitment side completely rigid; this table will become the pricing anchor for financing terms across the industry. The 11-02 issue already reported the filing and the 120.5 billion long-term compute contract; this item is incremental on Q3-report basis and customer structure.
DeepSeek financing expanded to RMB 105.5 billion, valuation RMB 615.5 billion; same day OpenAI lowered its financing target by $10 billion
DeepSeek／OpenAI36氪（转雷科技）／今日头条（转央视财经），2026-11-02；网易（转科创板日报），2026-11-01Confidence HighProgress · Day 2AI score 90/100
DeepSeek financing amount/valuation = 17.1%; OpenAI changed to raising about $20 billion at a $1.25 trillion valuation (dilution ratio 1.6%), same-day same-sector dilution intensity differs by 10.7x; DeepSeek same day unveiled the Super-agent autonomous-agent architecture.
Take · The reasonable reading is not 'more valuable,' but that the valuation is being pressed low and it needs to stock up ammunition in one go; conversely, OpenAI's proactive balance-sheet contraction shows it still has room to pick investors. The 11-02 issue already listed three capital moves side by side; this item is incremental on dilution ratio and architecture disclosure.
Marvell acquires Lumentum for $115 billion; Goldman Sachs-basis full-year M&A volume tops $1.1 trillion for the first time.
迈威尔今日头条（转和讯·智通财经），2026-11-02Confidence HighNew topicAI score 86/100
Marvell acquires Lumentum for $115 billion; on a Goldman Sachs basis, full-year M&A volume tops $1.1 trillion for the first time.
Take · The M&A focus is shifting from "buying models" to "buying optics and interconnect"; read alongside upstream purchases of heavy assets and downstream financing contraction, any mismatch in payment timing could cause impairments to concentrate and hit in the same year.
Oracle raises $20 billion from investors to fund AI expansion (including bond issuance).
OracleReuters，2026-11-02（官方视频监控台账）Confidence HighNew topicAI score 84/100
Oracle raises $20 billion from investors for AI expansion, including bond issuance.
Take · Unlike equity financing in the NVIDIA chain, this funding bears interest — AI capex is shifting from the equity ledger to the credit ledger, making ratings and spreads the next leading indicators.
Hong Kong SFC's first AI survey: 30% of large financial institutions have used agentic AI, 25% are piloting it
香港证监会腾讯新闻（转央广网），2026-11-02Confidence HighNew topicAI score 88/100
Hong Kong SFC's first AI survey says 30% of large financial institutions have used agentic AI and 25% are piloting it, and it will develop regulatory guidance for high-autonomy applications.
Take · The regulatory focus shifts from 'whether it can be used' to 'what degree of autonomy it is used at,' with a quantitative baseline provided; once liability boundaries are codified, Agent auditing and liability insurance will move from nice-to-have to market-access requirements.
Unified crawler opt-out signal submitted to IETF; in the same week, The New York Times and other media outlets sue Brave
OpenAI／谷歌／微软／Anthropic腾讯新闻（转 IT之家），2026-11-01；TechCrunch AI，2026-11-02（台账）Confidence MidProgress · Day 2AI score 85/100
OpenAI, Google, Microsoft, and Anthropic jointly promote a unified AI data-scraping opt-out signal, submitted to IETF by Anthropic's Kevin Kelley and integrated into robots.txt; in the same week, The New York Times and other media outlets sue the Brave browser over scraping content.
Take · Protocol work on one side and litigation on the other: content property rights are being priced through two tracks at once. The risk lies in neutrality: if the implementation semantics are 'crawlable by default, explicit refusal required,' the opt-out standard could instead become permission to enter. The 11-02 issue already reported the four companies' unified standard; this item is the incremental development of the IETF submission and the lawsuit appearing in the same week.
DoorDash says external AI agents have brought only a few hundred orders; Squarespace cuts 19% of staff
DoorDash／SquarespaceArs Technica，2026-10（台账）；TechCrunch AI，2026-11-02（台账）Confidence MidNew topicAI score 80/100
DoorDash says external AI agents have so far brought only a few hundred orders; website builder and hosting provider Squarespace cuts 19% of staff.
Take · Read alongside “30% of institutions already use agentic AI”: financial back office is moving faster than consumer front office—the narrative is on the supply side, the readings on the demand side; the job signal from AI tools directly compressing templated delivery is already clear.
04RESEARCH2 stories
智元 GO-3 Preview tops three major leaderboards: the competitive position in embodied AI shifts to “third-party leaderboard verifiability”
智元亿邦动力，2026-10-29Confidence MidNew topicAI score 74/100
智元 GO-3 Preview simultaneously tops three embodied AI leaderboards.
Take · The evaluation methodology of the three leaderboards was not disclosed in the source material; ranking comparability awaits third-party review; this item records only the fact of “verifiability” itself, not a capability conclusion.
Kimi K3.1 Pro surpasses MiMo-V2.6-Pro, DeepSeek V4.1-Flash, and GPT-6 Astra on the Artificial Analysis Intelligence Index
Kimi亿邦动力，2026-10-26Confidence MidNew topicAI score 72/100
On the third-party intelligence index, Kimi K3.1 Pro's reading is now higher than the three flagship models above, and its weights will be opened in November.
Take · Open-source weights are reusable assets, while leaderboard rankings are fragile; the index reflects only a single evaluation set, so deployment still requires a self-built, transferable evaluation set for retesting.
05PLAYBOOK6 stories
Agent behavior ledger four components: input, basis, authorization chain, and revocable credentials, all persisted to the database in one go
惊蛰·规律 #67总协调 Hermes；腾讯新闻（转央广网）2026-11-02；同花顺财经 2026-11-02Confidence MidViewpoint
For finance/payment/transfer Agents, build a replayable decision ledger: input, basis, authorization chain, and revocable credentials—the four components persisted to the database in one go. The path is to first open-source an SDK to drive traffic, then charge per instance, and finally offer auditable deliverable reports bundled with liability insurance.
Take · 30% of large financial institutions are already running agentic AI, but who is liable for a mistaken payment still has no answer; replayability is not a compliance cost but the next must-have. The time window is 6 months before Hong Kong's high-autonomy guidance is formalized.
Procurement metric re-anchors: from 'FLOPs/USD' to 'effective bandwidth per watt × available power'.
总协调 Hermes·规律 #103巽风·明日关注（深度版前瞻栏）Confidence MidViewpoint
The 420,000 GPUs represent volume expansion, HBC's fivefold bandwidth per watt represents efficiency restructuring, and the Tianshu Tiangai 300 physical machine represents domestic catch-up—all three in the same week is no coincidence.
Take · Once power and cooling become hard constraints, the marginal benefit of stacking GPUs declines; only by re-scaling the TCO model can the three paths—buying GPUs, buying bandwidth, and buying power—be compared on the same table.
HBM long-term contracts signed within 2026 should not extend beyond 2027H2, and a 'generational transition price adjustment' clause should be added.
坎流·规律 #26单源路线图（深度版前瞻栏）Confidence MidViewpoint
HBC sampling in 2027H1 implies an 18-month lag; existing contracts should add iteration price-adjustment clauses.
Take · The longer the long-term contract, the higher the probability of depreciation after a generational transition; spelling out trigger conditions in the clause is more realistic than negotiating impairment after the fact.
The correct way to read 900% NRR: the denominator is revenue from one year earlier; churn does not enter the denominator
炎离·规律 #18／#63厚坤·规律 #60（深度版数据栏）Confidence MidViewpoint
Retention rate counts only those still paying; with a small base, any renewal plus seat expansion pushes it to an exaggerated level
Take · It is an asset signal, not a market-size signal; what really matters is whether it can stay above 300% at the same time next year. Also read a set of structural figures: customers paying $100,000+ annually 950→11,900; over 1,900 at the million-dollar level; top 5% enterprises' monthly spend is about 20x that of ordinary enterprises
Enterprise buyers should prepare a 'second supplier + transferable evaluation set' now; don't wait until the renewal year to discuss it
悦泽·规律 #71深度版商机栏Confidence MidViewpoint
When facing a supplier with 900% retention and top 5% customers spending 20x that of ordinary enterprises, set your bargaining anchor in advance
Take · The higher the NRR, the more switching costs are hidden; a transferable evaluation set is the only practical way to turn 'can't switch' into 'can afford to switch'
Compliance budget booked to the wrong account: Agent compliance is booked under IT spending, while risks are paid out by reputation/legal.
悦泽·规律 #73深度版商机栏Confidence MidViewpoint
Most organizations book Agent compliance spending under IT, while risk events are paid from reputation and legal accounts.
Take · This mismatch is itself an opening for consulting and products: re-sort the budget table by "who pays," and a sellable service list naturally emerges.
Ecosystem Pulseself-updating
News600Skills72OpenHub68Reddit302026-09-02 → 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 behavior ledgerBuild a "replayable decision ledger" for finance/payment/transfer Agents: inputs, basis, authorization chain, and revocable credentials—all four stored in one database write. Path: open-source SDK for lead generation → charge per deployed instance → produce "audit-deliverable reports" sold to compliance and legal → bundle with liability insurance.
window 6 months before Hong Kong’s high-autonomy guidance is formalized (regulator has set a quantitative baseline: 30% used, 25% pilot); first movers secure the de facto standard. 2. Data authorization state machineBuild an “authorization state machine + robots opt-out signal” adaptation layer for content providers and model makers: parse once, fully traceable across the chain; anything without a status is grayed out and not cited. Path: free scan of authorization gaps on own sites → SaaS monitoring of first-collection/opt-out changes → procurement due-diligence report (data source list + authorization credentials).
window 6-12 months before IETF finalization (such standards typically take 6-12 months to go from voluntary to de facto mandatory); the same week, NYT sues Brave, showing the litigation leg is also running. 3. Compute TCO metric-switch consultingHelp buyers switch the TCO model from “FLOPs/USD” to “effective bandwidth per watt × available power,” with long-term contract pricing advice at generational switch points. Path: one-off model audit → long-term contract price-adjustment clause template (including generational switch triggers) → annual production scheduling and procurement agency.
window Before HBM long-term contracts are signed within 2026; HBC samples in 2027H1; the later the terms are changed, the less bargaining room.