01MODEL RELEASES3 stories
DeepSeek open-sources its inference stack DeepInfer: official claims up to 1.95× agentic throughput; Latent Space measured 1.56—1.95× on H800.
DeepSeek官方 X（2026-10-20）+ Latent Space 实测Confidence MidVendor benchmark
It is officially described as a full-stack inference engine built from day one for DeepSeek models, open-sourcing mainly two variants, vLLM and SGLang, and introducing DeepSeek Sparse Attention (DSA), with accompanying paper AI Infra in the Large Era; the three demo entries are multi-node vLLM, SGLang, and SGLang-Omni, with access URLs dsk.deepinfer.cn／dsk.deepinfer.cn/sglang.
Take · The real watershed is not the throughput multiple but that model vendors are starting to bring their own inference stacks—once this layer is absorbed by the model side, vLLM／SGLang will degrade from an ecosystem niche to a reference implementation, and third-party inference providers’ bargaining basis will disappear.
Recursive releases Flint: 40B MoE, 2.5B active, 262K context, can run a coding agent on a 48GB Mac.
RecursiveTechCrunch + HuggingFaceConfidence MidProgress · self-reported benchmark scores
40B MoE (2.5B active), 262K context, Apache 2.0; self-reported SWE-bench Verified 74.2% (no independent reproduction); its Eureka system completed architecture search in under 4 months, and the official says most of the model was built by AI.
Take · The first time 'models building models' has landed with downloadable weights and laptop-class deployment requirements; 74.2% is self-reported with no third-party reproduction, so its value should be credited to the cost curve (runnable on 48GB memory), not the score.
DeepSeek and Huawei jointly design next-gen Ascend 950DT/960DT; TileLang Ascend edition open-sourced
DeepSeek × 华为The Information（10-20 首发，多源转载）Confidence MidSingle-source first report · reposted by multiple sources
Huawei engineers on-site modify chip design based on DeepSeek training needs; TileLang Ascend edition open-sourced; 960DT capacity to prioritize DeepSeek in 2027Q1; use 950DT for pre-training trials at year-end. Note: ledger has deleted the 'planned purchase of 160,000 chips' entry (unsourced).
Take · This is not a procurement relationship but a co-design relationship—binding the training framework (TileLang), chips (Ascend), and model into one vertical chain is the real path to bypass the CUDA ecosystem.
02PRODUCT4 stories
Claude Duet research preview: always-on cloud VM; tasks continue after shutdown; supports approved payments
Anthropic深度版 Top5⑤（官方 + 台账 10-22 节）Confidence High
Official line is 'work with Claude': the agent resides in a cloud VM and browser and continues executing after the local device is shut down; claims access to 6000+ connectors; supports approved payment operations; US waitlist, research preview stage.
Take · “Runs even when powered off” turns agents from in-session tools into always-on production factors—once billed hourly for always-on operation, the pricing unit shifts from token to “agent workstation,” making O&M and audit costs explicit.
OpenAI models launch on Google Cloud: GPT-5.6 Sol/Terra enter Model Garden, Codex-Max also enters Antigravity
OpenAI × 谷歌云深度版·产业维度（Spiceworks 报道）Confidence MidStatement conflict · pending verification
Deep-dive note: two major models will enter Google Cloud Model Garden, and Codex-Max will also enter Google's coding platform Antigravity; Google engineers tested Sol and found it “ran autonomously for 10 hours”; Anthropic remains one of Gemini Enterprise's three model providers. This dimension also flags a contradiction: the SVP's statement conflicts with an expired announcement on the homepage and must be handled as a “statement conflict.”
Take · A competitor's models on its own cloud shows cloud vendors now prioritize “selling compute” over “selling their own models”; the direct implication for buyers is that the same model across multiple clouds becomes a viable option, reducing lock-in risk.
Walmart launches its own agent gateway (MCP server): vetted agents can check inventory and place orders, with an “agent profile” and quota authorization
沃尔玛深度版·产业维度 + 官方台账（TechCrunch）Confidence Mid
Walmart uses an access regime: agents must pass review, have profiles created, and be authorized by quota before they can search in-site inventory and place orders for customers; Amazon, meanwhile, chose to block AI agents from placing orders for customers. The in-depth version separately notes Walmart agent traffic +380% YoY.
Take · The first layer of regulation for agent commercialization lies not with governments but with platforms; whoever sets access standards first collects "toll fees"; for those building agents, choosing sides means choosing channels.
NVIDIA GTC Berlin: Vera Rubin in full mass production, DSX POD deployment cut from 6 months to 14 days, open-source Agent security runtime OpenShell
英伟达官方·GTC 柏林 Keynote 实录（台账 10-22 节）Confidence HighVendor claims
Vera Rubin enters full mass production; DSX POD deployment cycle reduced from 6 months to 14 days (17 partners to deliver in early 2027); OpenShell, an open-source kernel-level sandboxed Agent security runtime; official figures put Agent token consumption at 132.6 trillion per week, 34x early February. In the same period, DeepSeek open-sourced DeepInfer, while Google and Meta open-sourced TorchTPU.
Take · “14-day deployment” is the most concrete business number today—it turns AI factories from engineering projects into replicable standard components; but 132.6 trillion tokens/week is a vendor figure and can only indicate a trend, not serve as a baseline.
03INDUSTRY6 stories
Meta Q3: free cash flow turned negative for the first time since going public, capex guidance raised instead to 1350—1450 hundred million, while issuing non-voting Class C shares
Meta财联社 + Yahoo 电话会（台账 10-22 节，多源）Confidence High
Revenue $64.6 billion (+26%) beat expectations; quarterly capex of $34.9 billion consumed operating cash flow, with free cash flow at -$2.12 billion (better than the market expectation of -$9.4 billion); 2026 capex guidance was instead raised to $135–145 billion; announced issuance of non-voting Class C shares (1:1 share allotment, effective 2027-01-19).
Take · “Not enough cash flow, equity takes over” is the standard move as AI capex enters its second stage: first use operating cash flow, then off-balance-sheet leases, and finally equity—using all three shows the ceilings of the first two have been reached.
DeepSeek’s roughly RMB 100 billion financing has closed; CITIC is advising, with a listing filing as soon as this year
DeepSeek彭博（10-22，系 10-20／10-21 传闻的交割确认）Confidence HighProgress · Closing confirmed
The roughly RMB 100 billion (about $15 billion) round has closed; the target was raised twice within half a month, from RMB 50→80→100 billion; CATL and Tencent each invested about RMB 10 billion as the largest backers; new entrants include Shixi, SMIC Capital, CPE, Legend Capital, and Geely; CITIC is advising, with a listing filing as soon as this year.
Take · The shift from “rumored financing” to “closing confirmed” is the dividing line for valuation approach: in the rumor stage, look at the story; in the closing stage, look at the terms. Meanwhile, its chip tie-ups with NVIDIA and Huawei are a combined play; half the meaning of the financing is to lock compute supply in its own hands.
Existing supply is locked up early: in 2027, AI infrastructure on sale is only about 2 million units, and companies have signed away the next 3 years of capacity
微软 / xAI / 谷歌 / Anthropic / Waymo深度版 Top5④（RAND 10-15 研究）Confidence Mid
NVIDIA's USD 500 billion-level AI-Infra tier has only about 2 million units on sale in 2027; Microsoft to add about 2 million cards in 2024—2026; xAI Colossus 2's single site at 1.5GW is the world's largest, with a year-end target of 6GW; Google Gemini 3 is trained entirely on its own TPUs; Anthropic on Nov 12 points to 2TW; Waymo in 2027 will deploy another 2,000 Jaguar I-Pace in 5 cities (inventory about 35,000 units, lifecycle about 4 years).
Take · This explains an anomalous phenomenon: money is not scarce, demand is not scarce; what is scarce is the '2027 delivery slot'—so the real risk is not price cuts but production scheduling, and the shift to long-term contracts is precisely the financialized expression of tight scheduling.
France's 10-year government bond yield broke 5% intraday (first time since 2002); Morgan Stanley: five major companies' signed but not yet started lease commitments total USD 1.118 trillion
法国 / 摩根士丹利新华社 + 财联社 + 工商时报（台账 10-22 节）Confidence High
France's 10-year yield intraday at 5.005%, France-Germany spread about 157bp, debt/GDP 119%, Moody's 10-23 reassessment; Morgan Stanley (10-20): AI capex in 2027 about USD 1.58 trillion, five major companies' signed but not yet started lease commitments USD 1.118 trillion (Microsoft USD 329 billion / Oracle USD 288 billion / Meta USD 279 billion).
Take · AI capex is already on the same sheet as sovereign bond yields: a 100bp rise in financing costs extends the payback period for AI infrastructure by a notch—this is the real reason 'off-balance-sheet leasing' keeps getting flagged.
Three regulatory tracks on the same day: Anthropic proposes its own federal bill; UK FCA mandates reporting of near-miss incidents; local 'sandbox regulation' proposed for legislation
监管三线Politico（10-21 首发）+ 英国 FCA + 瞭望（单源）Confidence Mid
Anthropic draft: 'imminent catastrophic risk' causing more than 50 deaths or more than $1 billion in losses must be reported to the Commerce Department within 72 hours, and only federal courts can halt it; UK FCA adds AI 'near-miss incidents' to existing mandatory reporting rules; draft regulations in Jiangsu, Wuxi, Hangzhou and Nanjing write 'sandbox regulation' into legal provisions (Liaowang single source, medium-high).
Take · What is most worth watching is 'who writes the rules': regulated entities submitting their own draft bills is tantamount to taking ownership of compliance standards in advance—once passed, later entrants will pay the compliance tax it sets.
Neolix raises nearly $400 million Series E (led by Boyu) + Zhipu wins Xiamen RMB 10.58 million AI medical foundation bid
新石器 / 智谱央广网 + 中国政府采购网（台账 10-22 节）Confidence High
Neolix's round is the largest single domestic autonomous-driving financing this year, with cumulative funding exceeding RMB 10 billion; more than 100 X6T self-loading/unloading vehicles have been deployed (labor -20% / loading-unloading efficiency +50%), cumulative unmanned operations reach 260 million km, and overseas orders exceed 10,000 units; on the same day, Zhipu won the RMB 10.58 million bid for the Xiamen Health and Medical Big Data Center's AI medical foundation (first prefecture-level-city direct-from-original-manufacturer order).
Take · For unmanned vehicles, the key is 'mileage × per-vehicle economics'; for government-enterprise AI, it is 'whether the first order can be replicated'—RMB 10.58 million is not large, but the combination of 'direct-from-original-manufacturer + prefecture-level-city' is a replicable template.
04RESEARCH4 stories
Microsoft Research announces resolution of the 31-year-old KLS conjecture
微软研究院官方博客（台账 10-22 节）Confidence MidOfficially claimed · pending review
The official blog frames it as the resolution of a “31-year major mathematical challenge” (the Kannan–Lovász–Simonovits conjecture); the blog title does not give the proof path or methods used, so it is recorded as “officially claimed solved” and not treated as “accepted by the mathematics community.”
Take · The old problem with AI in mathematics is not “whether it can guess the conclusion,” but “whether the conclusion can be reviewed by humans”—the lesson from the last similar announcement is that between the publication date and the acceptance date lie months of peer review.
Three papers on the same day point to the same thing: what is really being optimized is the harness, not the model
HF Daily Papers深度版·技术维度（三篇同题）Confidence MidSingle source · not reproduced
HarnessAdapt-RL (based on Qwen3-4B): +12.4 in the training distribution, +6.2 on a new harness; out of distribution, after switching harness, +6.2 and baseline -1.8; there are also Self-Credit and Configurable-Perception World Models, the latter concluding that “given the evaluation protocol, the optimizable items are limited, and changing the evaluation protocol yields greater gains.”
Take · This is more important than any model release: it proves that a large part of the score belongs to the “scaffolding” rather than the “weights”—so using others’ benchmark scores for model selection is equivalent to using others’ scaffolding to make decisions.
AA announces Intelligence Index v5 to launch in late October: not comparable to v4.3.2; adds Terminal-Bench Science and private anti-contamination set
Artificial Analysis官方公告 + OpenRouter（台账 10-22 节）Confidence HighMetric reset
v5 adds Terminal-Bench Science and private anti-contamination set; officials made clear v5 and v4.3.2 are not directly comparable; meanwhile DeepSeek V4.1 Flash 1017 ranked 2nd on the OpenRouter weekly chart with 16.5T tokens in one week, and newly entered the Vals Index at 8th (61.3).
Take · The evaluator's own announcement of a change of yardstick is the one thing to remember today: all cross-version leaderboard comparisons are void from the release date; citing old scores must include the version number—otherwise it is measuring the same thing with two different rulers.
Material health AMBER: 86 items, 24 selected after deduplication (72% excluded); 6 of the 24 pending verification; on the official side, 10/80 sources failed
素材自检深度版·数据维度（20 路元素情报 + 官方台账批次）Confidence HighMethodology self-check
86 items of bot intelligence from 20 channels; after deduplication, 24 were selected; 8 source plans were not executed (36 single-source / 6 empty runs / 2 format mismatches); of the 24, 6 are pending verification (25%); on the official video side, 10+15+21+9 items, with 1 batch 'substandard'; source failures 10/80.
Take · Including 'how dirty today's material is' in the daily report is meant to let readers know which items should be discounted—without this section, it would hide the 72% elimination rate and deliver only a seemingly clean final product.
05INSIGHTS6 stories
Long-context cache dividend: 96% of tokens before the hit are cacheable; V4 Flash cache-hit price 0.02 yuan/million
商机深度版·商机维度（官方定价页）Confidence Mid
Official figures: 96% of tokens before the hit are cacheable; V4 Flash cache hits cost 0.02 yuan/million tokens (already one-fifth of the miss price); within the 72-hour cache window, 'write today, edit tomorrow' is almost zero-cost.
Take · This is currently the only cost dividend available without a model upgrade, but the conditions are strict: the cache only recognizes exactly identical preceding text; once the prompt changes, the hit rate immediately collapses to 0—so it is an 'engineering discipline dividend,' not a 'technology dividend.'
80% is contribution gross profit, not net profit: GPU contribution gross profit 89%, net margin 9—14 percentage points lower
复算深度版·数据维度（第三方复算）Confidence MidRecalculation
GPU gross profit contribution 89%, with net margin 9—14 percentage points lower; third-party recalculation range 73.7%—86.8%; in the same period, '807 亿' was depreciation and amortization, operating cash flow 693 亿, free cash flow only 90 亿.
Take · The same set of numbers has three versions; mixing them once yields a completely opposite conclusion: 89% is contribution gross margin, 73.7%—86.8% is the recalculation boundary, and only 90 亿 is freely disposable money—when citing, you must choose one of the three and state the definition.
Move security and research to outcome-based payment: business units only recognize results, while AI still charges by token
动作深度版 Top3③（Tenex 案例）Confidence Mid
Case definition: 1.75 万英镑; within 2 months, compressed 3 law firms’ 3 months of research into 72 hours; contracts changed from 'per hour/per token' to 'per outcome'; law firms charging by outcome is the norm, yet AI services still bill by usage.
Take · Billing by token leaves all cost risk with the buyer, while what the buyer really wants to buy is 'usable conclusions'—whoever dares to quote by outcome first will get budget sign-off first.
Evaluation sets are assets; prompts are consumables.
前瞻深度版·技术维度研判Confidence Mid
+6.2 to +12.4 points (HarnessAdapt-RL basis), while prompt optimization usually has marginal returns fall to zero after a few iterations; the evaluation protocol itself becomes an optimizable, tradable asset.
Take · The investment focus should be reversed: put budget into “how to evaluate” rather than “how to write prompts”; the former is also what can be reused and sold externally.
Security-side budget gap: attackers’ token cost is near zero, while defenders have only 10—15% of IT budget
风险深度版 Top3① + 官方 Patch the PlanetConfidence Mid
Official line: agents can discover critical vulnerabilities at scale in minutes (two PHP CVEs on the same day estimated to affect about 70% of websites); security budgets typically make up only 10—15% of IT spending.
Take · Offense-defense costs are now completely asymmetric: finding vulnerabilities has near-zero marginal cost, while patching requires labor and downtime—so “managed patching + vulnerability disclosure coordination” will become a new procurement category, not another security scanner.
China-U.S. pricing gap of 4.6—21x: V4 Flash 0.28 USD vs 3.2/5.0 USD; the price gap itself is product space.
商机深度版·定价维度（官方定价页）Confidence MidRecalculation
For the same capability tier, V4 Flash is quoted at USD 0.28, a 4.6–21x gap from the USD 3.2/5.0 comparison tiers; the price gap mainly comes from cache structure and activated parameters, not merely subsidies.
Take · The 4.6–21x is not a discount gap but 'two cost structures for the same task'—the middle segment is the arbitrage space: move cacheable, batchable workflows to the low-price tier and leave heavy work to the high-price tier.
Ecosystem Pulseself-updating
News600Skills72OpenHub65Reddit252026-09-02 → 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. Start using long-context caching this weekRearrange live conversation and document-processing pipelines around 'exactly identical preceding text': pin system prompts and static background first and put variables last, so the hit rate can rise from 0 to the official 96% upper-bound range; V4 Flash cache hits at 0.02 yuan/million are already one-fifth of the miss price, and within the 72-hour window, 'edit the next day after writing' is nearly zero-cost. Path: measure current hit rate → change prompt structure → re-measure cost; deliver the accounting within two weeks.
window This week (72-hour cache window + official discount period; changes involve only prompt structure, no model switch required) 2. Build a counterparty list for '2027 supply locked up'Based on five points—NVIDIA's 500-billion-level tier with about 2 million units on sale in 2027, Microsoft adding 2 million cards in 2024–2026, xAI 1.5GW→6GW, Google all-TPU, and Anthropic 2TW—make a table of 'who has signed away the 2027 delivery slots'; then overlay Morgan Stanley-cited 1.118 trillion in signed-but-not-started lease commitments (Microsoft 329 billion / Oracle 288 billion / Meta 279 billion). Purpose: find alternative paths for our own 2027 compute budget (self-build, second-hand, heterogeneous), rather than waiting for production scheduling.
window 6–8 weeks (before the 11–12 earnings season, when long-term contract figures first enter 10-Q/10-K notes) 3. Security side shifts to outcome-based billing to seize the 90-day post-disclosure windowTwo PHP CVEs (CVE-2026-109618/109619, estimated to affect about 70% of websites) have been disclosed, with fixed versions 8.5.12/8.4.27/8.3.36/8.2.35; the next 90 days are a concentrated upgrade period for existing systems. Deliverables: ① inventory of existing assets and version verification checklist ② managed patch package (priced on a fix-one-count-one basis) ③ disclosure coordination service (aligned with OpenAI's Patch the Planet framing).
window 3–6 months (vulnerability attention window + same window as regulatory near-miss reporting requirements)