Boss lens: labs are locking generators themselves, which means power and lead times—not just model prices—now gate capacity. If your workload sits on one cloud or region, ask about power and expansion queues (inference).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 9, 2026
11-08 07:50 – 11-09 07:50 (UTC+8)
Top 3
Unconfirmed Boss lens: GLM showing up inside China Mobile Cloud’s MaaS catalog could shorten procurement for domestic compute—but versions, price and go-live are unpublished, so treat it as intent rather than a live SKU (inference).
Boss lens: judgment-heavy services are still hiring, while template-and-document work is already thinner—separate “can draft with AI and check it” from pure data-entry roles (inference).
All
FT: Anthropic took over 1.7GW of gas-generator commitments for about $3.4B, called the largest known single power-kit purchase by an AI labUnconfirmed
Boss lens: labs are locking generators themselves, which means power and lead times—not just model prices—now gate capacity. If your workload sits on one cloud or region, ask about power and expansion queues (inference).
CollapseFT: Anthropic took over 1.7GW of gas-generator commitments for about $3.4B, called the largest known single power-kit purchase by an AI lab
The Financial Times reported on Nov 8, as relayed by CNBC and others, that Anthropic in late September took over commitments for about 1.7 gigawatts of heavy-duty gas-fired generators from San Antonio supplier OnePWR Solutions in a deal valued at about $3.4 billion, including engineering and construction capacity to install the machines. OnePWR had ordered the units from Caterpillar about two years earlier but needed financing to complete the purchase; Anthropic first planned to buy power from OnePWR, then took over the orders instead. It did not acquire OnePWR or its staff, and the eventual siting of the generators was not disclosed. Reuters said it could not immediately verify the report; Anthropic and OnePWR did not comment. This remains a people-familiar-with-the-matter account, not an official announcement.
China Mobile Cloud and Zhipu sign a deal: a joint Token center, and GLM models will join Mobile Cloud’s MoMA MaaS platform
Boss lens: GLM showing up inside China Mobile Cloud’s MaaS catalog could shorten procurement for domestic compute—but versions, price and go-live are unpublished, so treat it as intent rather than a live SKU (inference).
CollapseChina Mobile Cloud and Zhipu sign a deal: a joint Token center, and GLM models will join Mobile Cloud’s MoMA MaaS platform
IT Home reported on Nov 8 that China Mobile Cloud and Zhipu AI signed a cooperation agreement on Nov 6, witnessed by Mobile Cloud GM Sun Shaoling and Zhipu CEO Zhang Peng. The two sides said they would focus on adapting models to domestic chips, jointly build a Token development center, and work to raise token-production efficiency. Zhipu’s GLM models will be listed on Mobile Cloud’s MoMA model-serving platform (its MaaS offering). They also plan a joint product lineup and an AI-coding collaboration, plus integrated compute-plus-model-plus-application packages for finance, government and other sectors. The report did not name which GLM versions, a launch date, or pricing; this is still a signed intent, not a live catalog.
WSJ: U.S. professional and business services added about 70,000 jobs in October, the most since July 2022; document-preparation employment was down about 17% year over year through September
Boss lens: judgment-heavy services are still hiring, while template-and-document work is already thinner—separate “can draft with AI and check it” from pure data-entry roles (inference).
CollapseWSJ: U.S. professional and business services added about 70,000 jobs in October, the most since July 2022; document-preparation employment was down about 17% year over year through September
The Wall Street Journal’s Ian Hodgson reported on Nov 7 (with follow-on coverage from the Washington Post and Techmeme inside this window) that U.S. professional and business services added about 70,000 jobs in October, the largest monthly gain since July 2022. Consulting firms, law offices and accounting businesses all hired, even though those lines rank high on AI-exposure lists. The same reporting cycle, relayed by multiple outlets, said employment in document preparation was down about 17% year over year through September. The BLS public series shows that industry’s employment falling from about 33,500 in September 2025 to about 27,800 in September 2026, which matches the percentage. White-collar hiring as a whole did not collapse, but routine document work is already shrinking.
SemiAnalysis: Envision’s Ulanqab Galaxy campus delivered its first 120MW in about a year at roughly one-fifth of U.S. cost; up to 800MW of IT capacity is planned
Boss lens: 120MW in a year at about 5.5 cents/kWh shows how much location still moves cost and lead time if you ever site inference or backup nodes (inference). Parts of the original may sit behind a subscriber wall.
CollapseSemiAnalysis: Envision’s Ulanqab Galaxy campus delivered its first 120MW in about a year at roughly one-fifth of U.S. cost; up to 800MW of IT capacity is planned
SemiAnalysis published a site visit on Nov 7 (tour in September; circulated in-window via Techmeme and others): Envision Group’s Galaxy campus in Ulanqab, Inner Mongolia, delivered its first 120MW phase in August 2026, about a year after breaking ground, at roughly one-fifth of U.S. construction costs. The campus is planned for up to 800MW of IT capacity across four built-to-suit phases for two unnamed Chinese internet giants, with completion expected by the end of 2027. Envision’s Inner Mongolia data-center plans total about 3GW; Galaxy is the first. All-in regional power is about $0.055/kWh. Envision has about 420MW of its own wind against a 240MW data-center load and aims to cover more than 60% of the first two phases from those turbines. A 5MW hall pilots solid-state transformers that take 10kV AC to 800V DC at the rack. A local report cited by the firm put nearly 30,000 workers across about 22 data-center projects in the surrounding Chahar High-tech Zone.
Bloomberg’s Gurman: Apple CEO Ternus told software chief Federighi to speed the AI roadmap; Siri could act more like a goal-chasing agent as early as next yearUnconfirmed
Boss lens: a more agentic Siri next year could cut hand-offs in iPhone/Mac workflows—but this is still sourced reporting, not a shipping feature you should budget against (inference).
CollapseBloomberg’s Gurman: Apple CEO Ternus told software chief Federighi to speed the AI roadmap; Siri could act more like a goal-chasing agent as early as next year
Mark Gurman’s Bloomberg Power On newsletter, as relayed by 9to5Mac and MacRumors on Nov 8, said new Apple CEO John Ternus held a closed-door meeting with software and AI chief Craig Federighi in recent weeks and pressed him to accelerate the software roadmap beyond merely making necessary Siri improvements. Gurman wrote he was told Apple wants Siri to keep track of goals, decide what comes next and keep going—more like OpenAI’s Dots, Meta’s Muse and Grok—and that Siri could become “more like an agent” as early as next year. The same report said Apple is experimenting with a pipeline that gives engineers about 50% more time to code and explore, and less time on integration, as coding agents improve. Vision Pro and Siri AI lead Mike Rockwell is still seriously considering stepping down next year. These are source-based accounts; Apple has not published an official timetable.
The Information: excluding Anthropic’s IPO, AI companies sold or priced $300B+ of stock in 2026, with about $100B more on tapUnconfirmed
Boss lens: a huge amount of AI equity already hit the market this year, so cheap follow-on capital may get harder—if someone cites “the industry is still raising easily,” ask which deal and whether it has actually priced (inference).
CollapseThe Information: excluding Anthropic’s IPO, AI companies sold or priced $300B+ of stock in 2026, with about $100B more on tap
Techmeme on Nov 8 summarized The Information’s Alex Eichenstein: excluding Anthropic’s IPO, AI companies had already sold or priced more than $300 billion of stock in 2026, with about $100 billion more still teed up, raising questions about how much supply the market can absorb. The Information article is paywalled; this item uses only the Techmeme headline-level summary and does not break out issuers or which deals were follow-ons versus new listings. The tally is separate from Anthropic’s own listing process.
DeepSeek Harness’ English community post: an official plugin marketplace and author-support program are coming; it hopes to be the operating system of the agent era
Boss lens: a marketplace plus author grants is a bid to become a platform others build on, not just a chat app—and the 1.54 million DAU counts official-API users, not self-hosted copies (inference).
CollapseDeepSeek Harness’ English community post: an official plugin marketplace and author-support program are coming; it hopes to be the operating system of the agent era
DeepSeek Harness posted an English article, “Growing Together with Our Plugin Community,” on Nov 8. It said the team would harden plugin-API stability and security and build an official marketplace so developers can publish and users can find plugins. It is also preparing a support program for plugin authors—token grants, technical discussions and promotion—with details to come at launch; a form now collects interest, and companies are invited to ship their products as plugins. The article restated that, as of Nov 4 and counting users on the official API, daily active users reached about 1.54 million and daily token use about 92 trillion; third-party plugins have grown to about 35,000, and about 60% of active users have built or installed at least one. Those usage figures had already appeared in an earlier Chinese disclosure; what is new in the English post is the marketplace and author-support plan, plus the line that Harness hopes to serve as “the operating system of the agent era,” with built-in features replaceable under an everything-is-a-plugin design. Team member Tianyi Cui repeated the OS analogy in Chinese. Bloomberg and others reported the DAU and token figures during this window.
Vista8: a weekend that used up 3 Codex reset cards and one Claude account added about 64,300 AI-written lines to screen recorder CapWise
Boss lens: one person plus coding assistants can get a tool to “basically usable” in a weekend—but the bill was three reset cards and a dead account. If you ask a team to sprint this way, set quota, backups and a reviewer first (inference).
CollapseVista8: a weekend that used up 3 Codex reset cards and one Claude account added about 64,300 AI-written lines to screen recorder CapWise
Product manager Vista8 (@vista8) posted on Nov 8 that over the weekend he used up three Codex reset cards and lost one Claude account, bringing his screen recorder CapWise (Linglu) to a basically usable state, with about 64,300 lines of code added by AI. He described the product idea as ready out of the box, with one-click export after recording, and said he planned to use the app to record a demo of itself. This is a first-person build log, not a third-party audit of hours or quality; he did not explain why the Claude account died. The post drew 94 bookmarks and 116 likes.
Guoyu: 24 hours consumed about 50% of one Claude Code account’s weekly quota, and about 90% of that work was fully automated
Boss lens: a fully automated day can burn half a weekly cap, so “let the agent run” hits ceilings fast—budget a human review pass and do not assume one seat covers a full-week sprint (inference).
CollapseGuoyu: 24 hours consumed about 50% of one Claude Code account’s weekly quota, and about 90% of that work was fully automated
Guoyu (@turingou) posted on Nov 8 with screenshots of what he said he shipped in the previous 24 hours: one day used about 50% of a single Claude Code account’s weekly quota, and about 90% of that work was fully automated. The post does not list each task in text or point to externally auditable deliverables, so the specifics rest on the screenshots and his own account. This is a user’s usage note, not an official Anthropic statistic. The post drew 94 bookmarks and 129 likes.
TestingCatalog: Google is internally testing a “next generation Gemini agent” for Gemini Enterprise, labeled Gemini for GoogleUnconfirmed
Boss lens: a next-gen agent as the Gemini Enterprise front door could raise what office suites can do on their own—but this is still an internal UI, not a reason to switch vendors today (inference).
CollapseTestingCatalog: Google is internally testing a “next generation Gemini agent” for Gemini Enterprise, labeled Gemini for Google
AI-news account TestingCatalog (@testingcatalog) posted screenshots on Nov 8 saying Google is working on “the next generation Gemini agent” for Gemini Enterprise, already in internal use, with UI copy reading “Gemini for Google - The next generation Gemini agent for enterprise.” The author expects a new generation of work-focused Google AI products, but did not give a ship date, pricing, feature list, or say whether outside customers can test it. This is an unreleased-product sighting, not a Google announcement.
The Information: amid a server shortage, Anthropic has been asking where to find 5–10MW of power in older buildings, reviving interest in aging telecom sitesUnconfirmed
Boss lens: even a frontier lab is hunting a few megawatts inside old buildings, which means empty space is not the same thing as available power if you ever lease a private or backup room (inference).
CollapseThe Information: amid a server shortage, Anthropic has been asking where to find 5–10MW of power in older buildings, reviving interest in aging telecom sites
The Information’s account posted on Nov 8 that, exclusively, Anthropic has been asking where to find 5 to 10 megawatts of power in older buildings amid a server shortage, and that the hunt is reviving interest in decades-old telecom facilities. The post did not name cities, say whether any site has been signed, or give a timetable. The article is paywalled; this item uses only the public post and does not add details from behind the wall. It is a different thread from the same day’s FT report on a 1.7GW generator purchase: one is taking over generating kit, the other is hunting mid-size power inside existing buildings.
U.S. judge lets Claude Max “5x/20x more usage than Pro” suit proceed; plaintiffs say weekly caps are about 3.5x and 6x
Boss lens: “5x/20x” is per five-hour session, not per week or month—check the weekly cap before you multiply the sticker price into a budget (inference).
CollapseU.S. judge lets Claude Max “5x/20x more usage than Pro” suit proceed; plaintiffs say weekly caps are about 3.5x and 6x
Courthouse News Service reported on Nov 6 (still circulating in this window) that U.S. Magistrate Judge Alex Tse in the Northern District of California found Anthropic’s promise of “5x or 20x more usage than Pro” for Claude Max plausibly misleading to customers who pay monthly, and let key claims proceed. Max 5x is about $100 a month and Max 20x about $200, versus Pro at about $17–$20. For one Claude model, the plaintiffs calculated that Max 20x delivers about six times Pro’s weekly usage and Max 5x about 3.5 times—because the advertised multiples apply per five-hour session, while weekly caps do not scale by the same factors. Tse noted the per-session explanation sits two hyperlinks beyond the main sales pages and is not required to subscribe. Claims for negligent misrepresentation and breach of contract may continue; some California unfair-competition and false-advertising claims were dismissed with leave to amend. The ruling lets the case go forward; it is not a finding that Anthropic must pay damages or change its ads.
Huawei Ascend: LoRA fine-tuning of Moonshot’s Kimi K3.1 Pro validated on a 128-card Atlas 950 SuperPoD; Swap Core caching said to add about 15%
Boss lens: LoRA on a 2.8T open model atop an Ascend supernode means some fine-tunes may no longer have to wait for foreign chips—but the speedup is the vendor’s number, so rerun it on your own data (inference).
CollapseHuawei Ascend: LoRA fine-tuning of Moonshot’s Kimi K3.1 Pro validated on a 128-card Atlas 950 SuperPoD; Swap Core caching said to add about 15%
Pandaily and follow-on reports around Nov 8 said Huawei Ascend has validated LoRA fine-tuning of Moonshot AI’s Kimi K3.1 Pro on a 128-card Atlas 950 SuperPoD. Kimi K3.1 Pro has about 2.8 trillion total parameters. Ascend cited roughly a 15% training-performance gain from Swap Core caching. Public repos also show related example setups (such as 16 nodes times 8 cards). The 15% figure is Ascend’s own claim, carried by a small number of outlets; it is not a Moonshot evaluation write-up, and the coverage does not include a full controlled-experiment appendix.
Tencent released a DSpine draft model for Qwen3.8-27B, saying agent-throughput tests ran about 2.71x–3.84x faster
Boss lens: a draft model that speeds an agent two- to four-fold without changing the main weights is worth an eval if you run your own inference—re-measure on your own dialogue lengths (inference).
CollapseTencent released a DSpine draft model for Qwen3.8-27B, saying agent-throughput tests ran about 2.71x–3.84x faster
Tencent posted a five-layer DSpine speculative-decoding draft checkpoint, tencent/Qwen3.8-27B-DSpine-b16, on Hugging Face to speed inference for Alibaba’s Qwen3.8-27B without swapping the target model. Tencent’s own agent-throughput tests put the speed at about 2.71x to 3.84x a standard autoregressive baseline. Related GitHub notes say support for Qwen3.8-27B inference was added in November 2026. The speedup is Tencent’s benchmark; output length and tool-call density change how much speculative decoding actually saves, so it should not be read as a guaranteed cut to every company’s bill.
Co-Design Bench: leading GPT-6 Astra meets about 70% of requirements on average, but fully completes only 26.4% of tasks
Boss lens: even the best coding agent fully finishes only about one in four real change requests—write down who reviews and what “done” means before you swap model names (inference).
CollapseCo-Design Bench: leading GPT-6 Astra meets about 70% of requirements on average, but fully completes only 26.4% of tasks
The Decoder reported on Nov 8 on a paper from Alibaba and several Asian universities: Co-Design Bench tests coding agents on real user change requests, asking them to rework existing web apps, sometimes from images, audio, video or documents. GPT-6 Astra led the ten models tested, meeting about 70% of a task’s requirements on average but fully completing only 26.4%; in most cases the agent did the bulk of the work and left at least one detail for a human. Astra fully solved about 41% of multimodal tasks and just under 18% of repository-scale overhauls. Claude Opus 5 and Qwen3.8-Max came next. The same Qwen3.8-27B scored 75% or 56% on larger overhauls depending on the harness. The authors argue older benchmarks have saturated near-perfect scores, while this test leaves a gap of more than 30 points. Each item is a single fixed request, with no mid-task course corrections from a user.
Nous Research open-sourced Arete-27B, a ~53M-token fine-tune of Qwen3.8-27B that it says cut propaganda content from about 22% to under 1%
Boss lens: vendors are now fine-tuning away a base model’s political slant—if your answers have compliance needs, write your own eval checklist rather than trusting the publisher’s percentage (inference).
CollapseNous Research open-sourced Arete-27B, a ~53M-token fine-tune of Qwen3.8-27B that it says cut propaganda content from about 22% to under 1%
Techmeme on Nov 8 relayed Nous Research’s release of Arete-27B, now on Hugging Face under Apache 2.0: a supervised fine-tune of Alibaba’s Qwen3.8-27B on about 53 million tokens, with weights, evaluation data and training data published. Nous says official Chinese propaganda content in answers fell from about 22% in the base model to under 1% after the tune, and argues the bias layer is “shallow.” Both the propaganda rate and the claim that general capability was not hurt come from Nous’s own evals; coverage does not include an independent replication. Other reports say Nous is taking its Hermes line into enterprise sales. This is a value-alignment fine-tune, not a new base-model pretraining run.
UK growth minister Tomlinson: Britain will not slow AI because of U.S. backlash; the official roadmap still calls for at least 6GW of AI data centers by 2030Unconfirmed
Boss lens: the UK says it still wants the build-out, but grid access is a separate question—if you ever site a European backup, ask about the connection queue and local-job conditions, not just the welcome speech (inference).
CollapseUK growth minister Tomlinson: Britain will not slow AI because of U.S. backlash; the official roadmap still calls for at least 6GW of AI data centers by 2030
The Financial Times reported on Nov 7 (circulated in-window; the full article is paywalled) that UK growth minister Dan Tomlinson said Britain would not slow its AI investment because of political backlash in the United States, and that the current pace is “too low for what we need to stay at the forefront.” The government’s Compute Roadmap forecasts a need for at least 6GW of AI-capable data-center capacity by 2030. Public summaries also say ministers are considering support via grid access and energy costs, while asking developers who want help to deliver local jobs and community benefits. Whether any new support has actually been decided is not established in the accounts we could read. The Guardian reported on Nov 5 that about 315 UK data centers totaling about 73GW are queueing for grid connections; that counts connection requests, not built megawatts, and the article predates this window, so it is background only.
The Information: Palo Alto Networks is weighing a roughly $500 million deal for Clutch Security, which manages passwords and codes used by AI agentsUnconfirmed
Boss lens: once agents can log in, passwords are no longer only a people problem—you do not have to buy a new vendor today, but someone should list which bots hold which keys (inference).
CollapseThe Information: Palo Alto Networks is weighing a roughly $500 million deal for Clutch Security, which manages passwords and codes used by AI agents
The Information’s account posted on Nov 8 that Palo Alto Networks is weighing its fifth acquisition of the year: a roughly $500 million deal for Clutch Security, a startup that helps companies manage passwords and authentication codes used by AI agents, described as an emerging business security risk. The post did not say how far talks have gone, whether an offer is in, or when anything would be announced. The article is paywalled; this item uses only the public post. It remains a “weighing” exclusive, not a closed transaction.
Bloomberg: Anthropic-backed super PAC Public First spent $63M+ in the 2026 midterms; 10 of the 17 congressional candidates it backed won
Boss lens: a lab that spent real money helping candidates win is likelier to keep pushing safety and disclosure rules—companies shipping to U.S. users should watch how the new Congress talks about data, transparency and liability (inference).
CollapseBloomberg: Anthropic-backed super PAC Public First spent $63M+ in the 2026 midterms; 10 of the 17 congressional candidates it backed won
Techmeme on Nov 8 summarized Bloomberg’s Emily Birnbaum: the Anthropic-backed AI-safety super PAC Public First spent more than $63 million in the 2026 U.S. midterms, and 10 of the 17 congressional candidates it backed won their races. The item sits against broader coverage of an AI-safety network competing with pro-industry super PACs, but this write-up uses only the spending and win counts listed by Techmeme and does not open the full Bloomberg story for opponent-spending comparisons. These are election results, not a newly passed regulatory statute.
Lxfater on open-source billion-context: a proxy between coding agents and model APIs that lets the model decide what context to compress or restore
Boss lens: compressing long sessions can cut token spend, but repeated folds can warp earlier details—65 bookmarks show demand, so trial it on a real project before you standardize (inference).
CollapseLxfater on open-source billion-context: a proxy between coding agents and model APIs that lets the model decide what context to compress or restore
Lxfater (@lxfater) on Nov 8 wrote up the GitHub project ranxianglei/billion-context, a context-compression proxy that sits between coding agents such as Claude Code, Codex, OpenCode and Kimi and the model API. Unlike a client’s one-shot summary, it injects four tools—compress, decompress, search_context and acp_status—so the model itself decides when and what to fold; details can be restored later, which the author says helps preserve the prefix cache. He also relayed a friend’s experience: the idea is sound, but after many compression rounds in a long session the context can drift into invented shorthand, more so on heavily quantized models, with occasional small bugs. Treat results as something to test from the repo. The post drew 65 bookmarks, more than its 50 likes.
AP: officials at the Commerce Department’s AI safety center CAISSI were asked to report informally to a White House office, raising independence concernsUnconfirmed
Boss lens: if safety reviews sit closer to the White House, the tone may become more political or more uniform—worth a watch by whoever handles U.S. compliance, not a reason to guess at new clauses yourself (inference).
CollapseAP: officials at the Commerce Department’s AI safety center CAISSI were asked to report informally to a White House office, raising independence concerns
Techmeme on Nov 8 summarized an Associated Press report by Garance Burke: officials at CAISSI, the Commerce Department’s AI safety center, were asked to informally report to a White House office, raising concerns about the center’s independence. The public Techmeme blurb does not name the office, say when the request began, or say whether it was put in writing. This item uses only that headline-level summary and does not open the full AP story. CAISSI is an internal government evaluation body; a change in reporting lines is not the same thing as a new rule on private companies.