01MODEL RELEASES3 stories
Kimi K3.1 Pro open weights: 2.8 trillion total parameters, 104B activated; Ascend and Azure Foundry onboarded on the same day
月之暗面 Kimi深度版·Top5③（EastFrontier/AIHOT、Superpower Daily、腾讯新闻引华为计算）Confidence HighProgress · Day 8AI score 96/100
2.8 trillion total parameters, 104B activated (about 3.7%), Stable LatentMoE architecture with 896 routed experts, 1M context; on 11-05 Huawei Ascend 950 and Atlas A3 announced 0 Day support (vLLM Ascend, SGLang); on 11-06 Microsoft Azure Foundry onboarded on the same day, offering a million-token window. The 11-04 issue previously reported a “commitment to release full weights.”
Take · A Chinese open-weight model for the first time gained two distribution channels on the same day—“domestic compute + US hyperscale cloud”; distribution is worth more than parameters. On the model side today there is only this one main thread, which itself shows the narrative focus has shifted from capability to bills.
Inkling 2 Small open-sourced: 594B total parameters, 17B activated, completed 12T token training in 68 hours
Thinking MachinesMarkTechPost（11-05）、Latent Space AINews（11-06）Confidence MidNew topicAI score 88/100
Released weights under Apache 2.0, multimodal + variable reasoning effort, completed 12T token training in 68 hours; Artificial Analysis score 31, foreign media says it is in the same tier as Kimi K3 and GLM 5.3. Official announcement date unclear; parameters and score treated as single-source.
Take · “68 hours/12T token” is the most specific publicly visible training-efficiency reading at present and can be used directly as a reference for unit-cost negotiations; but it is single-source + the announcement date is unclear, so the official repository and license need to be checked again before writing it into procurement materials.
OpenAI Images 2.6: +50% image generation speed, adds GIF; API offered in Flare and Sunburst tiers
OpenAI36氪转界面新闻（11-04）Confidence HighNew topicAI score 85/100
Official says image generation speed is up 50%, adds GIF; API split into Flare (fastest 6 seconds) and Sunburst tiers, with focus on strengthening character, product, and brand consistency in multi-turn editing.
Take · Competition shifts from “image quality” to “consistency + speed tiers”—for content pipelines, reusable consistency saves more rework than per-image quality.
02PRODUCT LAUNCHES5 stories
Cerebras × Kimi官方视频台账（YouTube·Cerebras，2026-11-07）Confidence MidNew topicAI score 84/100
Cerebras released a demo claiming Kimi K3.1 runs at 2,000 tokens/second on its platform. It is an official demo, with no third-party reproduction provided.
Take · Inference speed is being productized separately: with the same open weights, whoever’s hardware runs faster can win latency-sensitive scenarios—a direct lever for “billing per completed task.”
Original ↗Perplexity官方·Perplexity 博客（台账 2026-11-07）Confidence MidNew topicAI score 83/100
The official blog's wording is 'teammates are agents for complex ongoing work,' productizing colleague-style Agents that handle long-term complex tasks.
Take · Moving from 'one Q&A' to 'a period of ongoing responsibility' means billing and SLA must be designed by period rather than per use—if the middle layer still resells by token, gross margin will be directly squeezed out.
Original ↗Meta官方·TechCrunch AI（台账 2026-11-06）Confidence MidProgress · Day 2AI score 82/100
TechCrunch reports that Meta launched Spaces as a new way to collaborate with Muse agents; the 11-06 edition already reported that Meta opened Muse avatar video calls to US users.
Take · The battle for entry points comes before the model battle: whoever defines the 'collaboration space' by default holds the first hop of task distribution.
Original ↗阶跃星辰量子位（2026-11-07 台账）Confidence MidNew topicAI score 78/100
StepFun Terminal STEPX Neo agent phone appears at Qualcomm booth at CIIE, featuring a 'software-hardware-model integration' agent experience
Take · Device makers are moving agents from the App layer to the system layer—viewed alongside Honor writing 'trustworthiness' into product language, Chinese vendors are vying for the right to define 'system-level Agent'
Original ↗Stripe官方·Hacker News AI（stripe.dev，台账 2026-11-07）Confidence MidNew topicAI score 81/100
The Stripe developer blog published 'Your AI agent can move money in Stripe', meaning an Agent can directly initiate fund operations within the payment platform
Take · This is the real-world foundation for all 'authorization chain' discussions today: money can already move, so traceable operation records and authorization proofs become the only security boundary
Original ↗03INDUSTRY9 stories
OpenAI funding cools as three signals align: lead investors cut contributions, secondary-market stampede, White House acknowledges 'AI bubble worth watching'
OpenAI / 软银网易转华尔街见闻（11-07）、36氪转财联社（11-05）Confidence HighProgress · Day 6AI score 95/100
NetEase citing Wallstreetcn 11-07: multiple lead investors in this round cut contributions, SoftBank plans to invest only several hundred million dollars; at the end of October there were already reports of a $10 billion target cut (the 11-02 and 11-03 issues already reported the target cut); on 11-05 the Nasdaq fell as much as 5.35% intraday, SanDisk fell over 22%, and Micron fell over 17%; on 11-06 White House economic adviser Hassett said 'the AI bubble is indeed worth watching,' and SoftBank fell 10.8% in a single day.
Take · Primary-market cooling + secondary-market stampede + official acknowledgment align for the first time. Previous days were only a 'target cut'; today marks the first appearance of the harder primary-market signal, 'lead investors are no longer adding capital.'
Microsoft begins phasing out AI compute rented from Oracle: over $80 billion partnership revalued, Oracle FCF −$23.7 billion
微软 / 甲骨文IT之家引彭博社（11-07）、路透（11-06）Confidence HighProgress · Day 2AI score 93/100
Microsoft has begun phasing out some AI compute rented from Oracle and is revaluing their over $80 billion cloud partnership, because Oracle's servers use an earlier generation of NVIDIA GPUs; Oracle's free cash flow for the quarter was −$23.7 billion, and its stock fell as much as 3%. The 11-06 issue already reported that Oracle–Tencent's roughly $7 billion chip lease was canceled.
Take · The counterparty risk in the business of 'renting compute to hyperscalers' is becoming explicit: old cards are shifting from assets to liabilities, and the −$23.7 billion cash flow is the hardest data point.
The selloff's decline structure is worth more than the index: memory and optical communications fell far more than compute chips, followed by a V-shaped rebound the next day
市场结构36氪转财联社（11-05）、IT之家（11-06）Confidence HighNew topicAI score 90/100
11-05: Dow/S&P/Nasdaq intraday maximum declines 2.38%/3.57%/5.35%; SanDisk fell over 22%, Micron fell over 17%, NVIDIA fell as much as 9%; Musk's net worth shrank by $91.2 billion in a single day, Jensen Huang's by $15.1 billion. 11-06: Philadelphia Semiconductor Index rebounded as much as 6.57%, AMD rose 13.2%, TSMC rose 7.55%.
Take · The market is cutting the 'capacity narrative,' not the 'demand narrative' (memory and optical communications fell far more than compute chips); the next-day rebound is later and more verifiable evidence, so this report classifies 11-05 as a liquidity stampede plus sentiment turning point, not a trend re-rating—if a second bearish candle appears, the classification will be raised.
Bottleneck shifts tracks: Google bets $17 billion on 702MW in Wyoming; NVIDIA converts more than 130 old telecom buildings into about 1GW of compute
谷歌 / 英伟达腾讯新闻（11-06）、The Information（11-06）Confidence HighNew topicAI score 89/100
Google bets $17 billion on Wyoming; the first phase's 702MW is powered by natural gas fuel cells, with permanent emissions sequestration, air cooling, and fully bears its own electricity costs; it has signed a 20-year power purchase agreement (PPA date unknown) with American Electric Power (AEP); NVIDIA, using Apollo-owned Brightspeed assets, converts more than 130 old telecom buildings into about 1GW of distributed compute, able to accommodate more than 2200 Vera Rubin NVL72 racks.
Take · The constraint shifts from chips to 'power and permits': new builds require self-built power plus carbon sequestration, while retrofits of existing assets face fewer approvals and lower community opposition—existing distribution assets are the most undervalued scarce resource this cycle.
Alphabet opened a position in Qualcomm of about $548 million for 2.977 million shares, ranking sixth in its portfolio
Alphabet / 高通IT之家引路透社（11-07）Confidence HighNew topicAI score 86/100
Alphabet opened a position in Qualcomm of about $548 million for 2.977 million shares, about 0.5% of Qualcomm's total shares, ranking sixth in its portfolio; its total holdings market value for the quarter was $88.3 billion, down 10.9% quarter-on-quarter, and Qualcomm briefly rose about 5% after disclosure.
Take · Cutting positions while opening new single-stock positions is 'using equity to lock in a second supplier'—leaving itself a path in the compute supply chain that does not depend on a single vendor.
Alibaba led a $300 million investment in UniPat AI, at a pre-money valuation of $2.5 billion
阿里 / UniPat AI网易订阅转潮涌AI（11-06）Confidence HighNew topicAI score 85/100
Completed on 11-06, with Tencent and Sequoia China participating; founder Li Kuan was still an intern at Alibaba's Tongyi Lab in August 2025.
Take · Overseas retrenchment, domestic ramp-up, with money shifting from 'building models' to 'training and evaluation'—validation capability is beginning to be priced separately; this is today's most certain business opportunity opening.
Apple sets daily AI quota: engineers cut from $300 to $200, with plans to cut further to $100
苹果36氪转界面新闻（11-04）、东方财富（11-06）Confidence HighNew topicAI score 92/100
Some engineers’ daily cap has already been cut from $300 to $200, with plans to cut further to $100; non-engineering teams are already at $100; some teams have been asked to switch to lower-cost or local models. Another account says some teams may be cut to $25 (East Money 11-06), not cross-verified.
Take · $100/engineer/day is currently the clearest publicly visible “Token budget cap” and can directly serve as a ceiling reference for product pricing; the $25 figure is marked as unverified and not cited.
Social license for data centers becomes a hard constraint: voter opposition, OpenAI pledges to fully cover power supply, Microsoft tightens new cloud oversight
数据中心 / 政策网易转澎湃新闻（11-06）、TechCrunch（11-07 台账）Confidence HighProgress · Day 3AI score 88/100
In the midterm elections, voters in multiple places opposed data center projects (the 11-05 issue already reported 45 ballot measures “fully rejected”); OpenAI pledged to stop using confidentiality agreements with the government and to fully cover power costs, Microsoft tightened oversight of new cloud service providers and exited projects it deemed risky, and Google turned overseas (Malaysia, USD 15 billion in 2027—2028).
Take · Social license is being written into contract terms—essentially turning “power price + siting” into actionable commitments, making delivery delay risk in 2027 quantifiable.
Li Auto's chip subsidiary raises RMB 1.625 billion (16.25亿元) in first round, at a post-money valuation of RMB 17.1 billion (171亿元)
理想汽车网易转中国经营报/新浪财经（11-07）Confidence MidNew topicAI score 79/100
Li Auto's chip subsidiary completes a first round of RMB 1.625 billion (16.25亿元) financing at a post-money valuation of RMB 17.1 billion (171亿元); automakers continue to step up chipmaking (NIO Shenji has completed two rounds).
Take · Automakers turning smart-driving chips into independent financing vehicles turns compute self-sufficiency from a cost center into an asset that can be valued—but whether the RMB 17.1 billion (171亿元) valuation is reasonable depends on external customers, not internal-use orders.
04RESEARCH5 stories
Visa benchmark: 42.8% of shopping Agents fail, with only a 57.2% pass rate
Visa腾讯网 Agentic Commerce 日报（11-06）、Visa 投资者关系页Confidence HighNew topicAI score 94/100
Released on 11-05; 12 labs and 22 models executed 15 shopping tasks, each three times, with 42.8% run failures (57.2% success rate); the bottlenecks are not intelligence but selecting the right product, proving authorization, and leaving a traceable record of operations.
Take · Agent commercialization has a reproducible failure rate for the first time, and it is also the only leaderboard today usable as an acceptance standard—the three bottlenecks are all process and credential issues, exactly the opening for consulting and products.
METR measured AI’s acceleration of AI R&D at 1.9x, range 1.1x–2.9x
METR / Anthropictheainewsreport（2026-11-06 台账）Confidence MidNew topicAI score 84/100
METR measured Anthropic’s internal AI accelerating R&D by 1.9x, range 1.1x–2.9x.
Take · A single-point measurement with an extremely wide range can only indicate direction, not a promise—any long-term plan premised on “self-acceleration” should stress-test at the lower bound of this range.
OpenAI官方视频台账（YouTube·OpenAI，2026-11-07）Confidence MidNew topicAI score 80/100
OpenAI’s official video introduces LifeVal, used to measure AI progress on everyday real-life tasks.
Take · Alongside METR’s “AI doing AI research,” it forms a two-way stretch: one measures everyday real-world tasks, the other measures self-acceleration—only the two yardsticks together show where the capability boundary lies.
Original ↗Microsoft changes its tune: recommends evaluating models by 'completed tasks' rather than context length
微软Superpower Daily（11-06）Confidence MidNew topicAI score 82/100
On the same day, Microsoft recommended using 'completed tasks' rather than context length to evaluate models.
Take · Once the definition changes, the 1M-context premium cannot hold—evaluation methods often change before pricing systems do; this is a row to update in the selection table in advance.
Inkling 2 Small's public score is 31, said to be in the same tier as Kimi K3 and GLM 5.3
Artificial AnalysisMarkTechPost（11-05）Confidence MidNew topicAI score 74/100
Artificial Analysis gives Inkling 2 Small a score of 31; foreign media cited this to say it is in the same tier as Kimi K3 and GLM 5.3.
Take · Single-source score; not a basis for horizontal ranking; it can only serve as weak evidence of 'same tier'; real differences require looking at unit task cost.
05INSIGHTS6 stories
Pricing units are shifting from Token to 'completed tasks': three entities, one set of definitions
口径深度版·前瞻维度（微软、Visa、苹果三处独立来源）Confidence HighViewpointAI score 93/100
Microsoft changes course the same day, recommending evaluating models by 'completed tasks'; Visa prices Agents by task success rate; Apple manages budgets by 'dollars per person per day'.
Take · Three unrelated parties converge on the same framing within the same 48 hours—not a coincidence but an inflection point: whoever first converts their cost sheet into 'cost per task' gains pricing power first.
Three bills come due at once: tighter capital cuts the capacity narrative, tighter power raises unit costs, tighter trust raises the delivery threshold
结构深度版·前瞻维度Confidence MidViewpointAI score 90/100
Capital bill (cold financing + hardware selloff + Microsoft lease exit), power bill (702MW self-built power + retrofitting 130 old buildings), trust bill (42.8% failure rate + task-based evaluation + TQ trust level) all come due in the same 48-hour window.
Take · Inference: with all three overlapping, the most dangerous position is the 'middle layer'—no proprietary model, no proprietary power, no auditable capability. Such companies' moats will be squeezed from both ends at once.
'Accountable workflows' shift from slogan to procurement item and align fully with Visa's three thresholds
商机公众号侧清单（AI智能体落地局 11-04、跃澜 11-05）＋ Visa 基准Confidence HighViewpointAI score 91/100
An existing article argues that for enterprises building AI agents in 2026, the first gap to close is not models but 'accountable workflows'; another discusses 'who manages your digital employees'; this aligns fully with Visa's three thresholds: choosing the right product, proving authorization, and traceable records.
Take · This is the most certain entry point right now: authorization chain + operation logs + traceability; packaged as a health-check product it can drive leads, packaged as a subscription it can generate recurring revenue.
Retrofit arbitrage in existing data centers and power-distribution assets: 130 old telecom buildings vs. 702MW new-build
商机The Information（11-06）、腾讯新闻（11-06）Confidence MidViewpointAI score 86/100
Over 130 old telecom buildings, about 1GW, fewer retrofit approvals, and low community opposition; by contrast, Google's new 702MW build requires self-built power and carbon storage.
Take · Existing space is the most undervalued scarce asset this cycle: first build a regional power-distribution asset map (ownership, grid-connection capacity, community sensitivity), which is much cheaper than buying assets first.
Training efficiency has become a new comparison axis: training 12T token in 68 hours, with a 3.7% activation rate traded for unit cost
技巧深度版·数据维度（Inkling 2 Small 与 Kimi K3.1 两组读数）Confidence MidViewpointAI score 83/100
Inkling 2 Small finished training on 12T tokens in 68 hours (594B total parameters/17B active), while Kimi K3.1 trades a 3.7% activation rate for unit cost.
Take · Both numbers point to the same thing—parameter scale is no longer the main battleground. When choosing models, putting 'cost per task' and 'training/inference efficiency' side by side in two columns is more useful than looking at total parameters.
Tokens sold like data bundles: Honor × China Mobile debut Token packages on Magic9 customized phones
渠道36氪（2026-11-07）Confidence MidViewpointAI score 80/100
Honor's Li Jian proposes TQ (trust) beyond IQ and EQ, and says Honor will launch Token packages first with China Mobile on Magic9 customized phones and explore an 'intent economy' monetized by outcomes.
Take · A telecom operator becomes a Token distribution channel for the first time—this is the billing experiment closest to China's channel reality, and the first consumer-side attempt to implement 'pay by results'.
Ecosystem Pulseself-updating
News600Skills72OpenHub68Reddit252026-09-02 → 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 accountability workflow health checkEntry point: run an “authorization chain + operation logs + traceability” health check for enterprises that have deployed Agents, directly aligning with Visa’s three thresholds (selecting the right product, proving authorization, traceable records). Path: public 15-item task self-check list for lead generation → per-use audit (including failure-rate reproduction) → subscription-based continuous monitoring and quarterly re-testing.
window Next 3 months from now: 42.8% has already been made public; buyers will certainly ask next, “How do you prove your Agent won’t go off the rails?” 2. AI-ization intermediary for existing power distribution assetsEntry point: provide due diligence and retrofit plans for “old telecom buildings/old substations → small machine rooms that can be powered immediately,” benchmarking NVIDIA’s 130 buildings/1GW financial engineering path. Path: first build a regional map of existing distribution assets (ownership, grid-connected capacity, community sensitivity) → issue retrofit feasibility reports → match with compute-demand parties and share revenue by capacity.
window 6–12 months: social license and grid connection are slow variables; securing information first is cheaper than securing assets first. 3. “Per completed task” selection and cost accountingEntry point: use the customer’s 3 most expensive real tasks to build a per-task cost benchmark (Kimi K3.1 Pro open weights can be deployed locally or on dual clouds; the 3.7% activation rate is the cost lever). Path: per-task cost baseline report → model × deployment combination recommendations → quarterly retest subscription, tied to “rework rate” rather than benchmark scores.
window From now through Q1: Microsoft has changed its stance; the premium for 1M context will be quickly squeezed out; those who build baselines early will gain pricing power.