Vol.56 · November 9, 2026
dera news AI Weekly Vol.56 | 2026-11-09 - This Week's AI News
🤖 dera news AI Weekly Vol.56
2026-11-09
In the US midterm elections on November 3, the side that wanted to limit data center construction won 40 of the 41 local votes on it. Two days later, on the 5th, US semiconductor and AI shares were sold off heavily, and OpenAI restated on the same day its pledge to pay the full electricity cost of its data centers. Anthropic, which is aiming to list, revised its stock offering document (the prospectus) on the 2nd to add July-to-September figures, and its core business swung to a loss.
📊 What You Need to Know This Week
After the share sell-off, some of the buyers lining up for Anthropic's listing are considering a valuation as low as $1.1 trillion, and OpenAI has reportedly lowered its fundraising target (stories 1 and 2). Anthropic is aiming to list before Thanksgiving, and its offering price will show how investors see the company after the election and the sell-off.
In the US, most of the local votes chose to regulate data centers, and Chinese companies are not getting Nvidia chips as ordered (stories 1 and 8). OpenAI executives reportedly discussed whether to revisit their 2027 computing capacity plans, and Oracle dropped a deal to rent chips to Tencent.
OpenAI released a model built to find flaws, and Japan's government put AI-based vulnerability hunting into its emergency measures (stories 4 and 7). In the same week defenders started using AI to find flaws, Binance's own internal AI tool became an attacker's instrument (story 9).
METR estimates that Anthropic's models are improving faster, mostly because of AI use inside the company, and Thinking Machines matched Google's lightweight model with a model it plans to release (stories 3 and 5). The shorter the gap between new models, the more it pays to check the conditions for switching before signing a one-year contract.
ChatGPT's GIF making and Gemini Omni's 15-second videos are already available (in the Gemini app, rolling out from the top plan), while Siri AI will be able to operate other companies' apps in a few months (stories 6 and 10). Apple supports Japanese from the start, and Japanese companies' apps will qualify in the spring submission window.
💡 This Week's Actions
1. Make one GIF with Images 2.6 (story 10) for an internal explanation (15 min) Pick one place in an internal manual or training material where steps are explained in text only. Tell ChatGPT the steps and have it make a looping GIF, paste it into the material, and ask a colleague whether it became easier to follow. → OpenAI: Introducing Images 2.6 (English)
2. In line with the government's emergency measures (story 7), check who tells whom within how many hours about a cyberattack (20 min) Ask your IT contact who the person who first notices an attack should contact, and who is responsible for notifying business partners and the authorities, and by when. If nothing is decided, make a contact list and share it. → Impress Watch: Government drafts emergency cybersecurity package (Japanese)
3. Tell employees not to run programs handed to them from outside on work PCs (15 min) To guard against tactics like the one in story 9, tell all employees in a company-wide message not to open files handed to them from outside, such as for interviews or coding tests, on a work PC. → The Block: Binance hack incident report (English)
📰 This Week's AI Articles (10 stories)
1️⃣ US midterms: 40 of 41 local votes on data centers go to the regulation side
🏷️ Policy, Power, Markets, OpenAI, Japan What happened? In the US midterm elections on November 3, according to CNBC and Ballotpedia, the side that wanted to limit construction won 40 of the 41 local votes on data centers across 10 states. Measures to regulate passed, and measures to push development ahead were rejected. On the 5th (US time), according to Bloomberg, the Nasdaq 100 index fell about 6.4%, and on the 6th in Tokyo the Nikkei average also fell about 6%. On the 5th OpenAI restated its earlier pledge to pay the full electricity cost of its data centers, and said that for facilities it owns and runs it will consult local residents, publish its construction plans, and not sign non-disclosure agreements (NDAs) with local officials. The Wall Street Journal reports that senior OpenAI executives discussed internally after the election whether to revisit their 2027 computing capacity plans (this is a single report, and OpenAI has not confirmed it). Our take Meta, Microsoft, Google and Anthropic have already made similar pledges, and the more OpenAI and others bear electricity costs, the more likely those costs are to be added to API (the way other software calls an AI) and monthly fees. Whether OpenAI actually shrinks its 2027 plan will also change how much computing capacity is available next year. In the next price-change notice from an AI service you use, check whether electricity costs are given as a reason for a rise. 📎 Read the original
2️⃣ Anthropic posts an operating loss for July to September, on revenue of about $17 billion
🏷️ Funding, IPO, Anthropic, OpenAI, Japan What happened? Anthropic issued a revised prospectus on November 2 (US time) showing July-to-September revenue of about $17 billion (counting sales through cloud providers in full). Its core business lost $567 million, a swing from a profit in April to June. It plans to sell up to ¥552.9 billion of shares to Japanese investors. According to Bloomberg and others it is aiming to list before Thanksgiving (November 26), and after the 5th's sell-off buyers are considering valuations of $1.2 trillion to $1.3 trillion, some as low as $1.1 trillion. On OpenAI, Bloomberg reported on the 6th that about $11 billion of a raise at a $1.25 trillion valuation has been completed. The FT reported on October 30 that its fundraising target was cut from $30 billion to about $20 billion. Our take The filing gives as the main reasons for the operating loss about $5 billion of computing capacity bought on pricey short-term contracts and about $2 billion of costs from the company matching employees' donations. The offering price will show how investors priced this loss and its huge payment commitments. If you use Claude, watch whether operating profit returns to the black in the first earnings report after listing. 📎 Read the original
3️⃣ METR estimates the pace of Anthropic's models has about doubled, mainly from internal AI use
🏷️ Research, Safety, Anthropic What happened? METR, a US nonprofit that studies frontier AI agents, published an assessment on November 5 (US time) using Anthropic's internal data. METR estimates that the rate at which Anthropic's models gain ability on tasks with set conditions has about doubled compared with before October 2025, and infers that most of this, probably around 80%, comes from AI use inside the company. It says it found no decisive constraint that would stop the acceleration over the next 12 months. The full report of about 100 pages was given to Anthropic, and what was published is a summary. METR writes that Anthropic redacted most of the evidence that most informed its conclusion. Our take Three weeks ago Anthropic itself said that code submissions per engineer had risen to five times the level of six months earlier. Unless the redacted parts are shown, it is hard for anyone outside METR to check the estimate of a doubling. When you explain the pace of AI progress inside your company, add that METR itself leaves a range of 1.3 to 3 times. 📎 Read the original
4️⃣ OpenAI releases "GPT-6 Cyber," a model for finding software flaws, to a limited group
🏷️ Security, AI Models, OpenAI What happened? On November 5 (US time) OpenAI released "GPT-6 Cyber," a model that finds software flaws (vulnerabilities) and can even build attack steps. It is available to some customers approved under "Daybreak Red," a program for organizations that are authorized to test their defenses using the same methods as attackers. According to OpenAI, the model found 81 previously unknown serious flaws in Firefox and helped fix them, and Mozilla says it has already fixed many. In Linux it also built a procedure by which an ordinary user could seize administrator rights (privilege escalation), but it is not clear whether that has been fixed. Our take A serious PHP (a programming language that runs websites) flaw found two weeks ago was also found with GPT-5.6 Sol by a security firm in OpenAI's program. OpenAI says it is "clear that the window in which we can use frontier models to defend software is limited," and if attackers gain similar ability, the time from a flaw's disclosure to an attack will shrink further. With your IT contact, shorten the number of days employees may delay browser and OS updates on their PCs. 📎 Read the original
5️⃣ Thinking Machines announces a new model, "Inkling 2 Small"
🏷️ AI Models, Open Models, Thinking Machines, Tencent What happened? Thinking Machines, led by Mira Murati, formerly OpenAI's chief technology officer, released the model "Inkling 2 Small" on November 5 (US time) on its developer service Tinker. It uses only 17 billion of its 594 billion parameters (the numbers inside a model) for each query, and costs $0.30 per million tokens for input and $1 for output. On the evaluator Artificial Analysis it ties with Google's Gemini 3.8 Flash (high setting), and the evaluator says that if the weights (the data needed to run a model on your own machine) are released after outside safety testing, it will score highest among openly released models made outside China. China's Tencent also released the final version of Hy4 on the 4th, but when the final weights will be published is unclear (the August preview version is public). Our take The weights are due after safety testing under a license that allows commercial use (Apache 2.0), which would add an option from outside China. If you are considering running a model on your own servers, estimate the number of GPUs (AI chips) and the cost it would take once the weights are out. 📎 Read the original
6️⃣ Apple opens Siri AI to developers of every app
🏷️ Apple, Agents, Japan What happened? On November 3 (US time) Apple released developer betas of iOS 27.3 and other systems, letting any app connect to Siri AI through App Intents (a mechanism by which an app tells Siri and others what actions it can do). For actions developers provide, users can simply ask in their own words and Siri AI, Apple's AI assistant, can operate other companies' apps, such as checking a delivery or making a reservation or order. Japanese is supported from the start, and it is not available on iPhones in the EU. It reaches general users in a few months, and App Store submissions open in the spring. Our take Following Siri AI's Japanese launch on October 26, Siri AI will now be able to call actions that apps have registered. If your company has an iPhone app, pick one action that is fine to leave to Siri AI, such as checking a reservation or reordering, and try it with your developers before the spring window. 📎 Read the original
7️⃣ Government's emergency cyber package: ¥10 billion from reserve funds, with AI-based vulnerability hunting
🏷️ Policy, Security, Japan What happened? On November 6 the government put together an emergency package at a ministerial meeting on cybersecurity attended by Prime Minister Sanae Takaichi, and at that day's cabinet meeting decided to spend ¥10 billion from reserve funds. One pillar is "thorough basic measures that also use AI." It will step up work to find vulnerabilities with AI and bring forward checks of government systems, and it aims to get information on attacks and undisclosed vulnerabilities sooner by working with overseas tech companies and companies building frontier AI. Takaichi asked that information on attacks reach the National Cybersecurity Office "within a few hours if possible, and at least the same day," and told Toshiharu Furukawa, the minister in charge of cyber security, to consider amending the Active Cyber Defense Act. Our take The government is trying to shorten the time to collect attack information to a few hours, and depending on the amendment, the range of companies required to report and the deadlines may change. Have your legal and security contacts follow what the amendment debate will require of private companies. 📎 Read the original
8️⃣ Chinese companies' orders for Nvidia chips: only about 1% expected to arrive
🏷️ Semiconductors, China, Nvidia, Oracle What happened? According to Bloomberg, Chinese companies including Alibaba, Tencent and ByteDance ordered about 6 million Nvidia RTX Pro 5500 chips (about $70 billion worth), which are approved for export, but only about 1% is expected to arrive within four months. The UK's Financial Times reported that Oracle dropped an estimated $7 billion deal to rent up to 100,000 chips to Tencent for five years after pressure from the Trump administration. According to Bloomberg, Nvidia says Southeast Asian cloud providers must join a system that records where chips are located to become certified partners. The New York Times reported that of about 20 companies permitted to assemble H200 chips into servers, two had empty offices. Our take A draft Chinese rule to control exports of items such as large models' weights is said to close for comment on November 15. If chips not arriving combines with export rules, the gap before Chinese labs release their next versions is likely to grow. If you use a Chinese model with public weights in your work, decide on one model to switch to if the next version is late. 📎 Read the original
9️⃣ Over $6.3 billion leaves Binance, with a fake hiring process and an internal AI tool as the way in
🏷️ Security, Crypto, Agents, North Korea What happened? The crypto exchange Binance published on November 3 an initial report (the figures may change) on a theft of more than $6.3 billion in crypto assets. The attacker had already been in an employee's PC in June, and in September approached another employee through a fake hiring process and got into a work PC through a program handed over as an assignment. On October 20, using that PC and an internal AI tool, the attacker took over two other engineers' PCs, and about nine days later used stolen credentials to run the transfer system. According to Binance, the US Federal Bureau of Investigation (FBI) links the attack to North Korea. Binance says it had made up 98% of the affected assets as of October 31, but has not said when it noticed the attack or what the AI tool did. Our take Binance co-CEO Richard Teng wrote on X that it was "not a single mistake or a single breach but a months-long operation." The attacker used the internal AI tool as a foothold to reach the PCs that hold transfer authority. With your IT contact, list which systems and PCs the AI tools used in your company can send operations to. 📎 Read the original
🔟 ChatGPT's image generation now makes GIFs, on every plan
🏷️ Image Generation, Video Generation, Pricing, OpenAI, Google What happened? On November 4 (US time) OpenAI updated ChatGPT's image generation to Images 2.6, which can make looping GIFs. Images 2.6 is available on every ChatGPT and Codex plan, including free ones, and generation is about 50% faster than before. For developers it released "GPT-Image-2.6 Flare," which stresses speed and layout, and "Sunburst," which stresses detail and consistent editing. On the 5th Google released Gemini Omni 1.2 Flash, a video model that makes videos up to 15 seconds long, and cut its API price to $0.07 per second at 720p, 30% lower than the previous version. In Google Vids, anyone with a Google account can make a 15-second video at no extra charge. Our take Google's videos carry an invisible watermark (SynthID) that lets anyone confirm later that they were made with AI. Decide with your PR contact whether to note that an AI-made image or video was made with AI when you use it in external materials. 📎 Read the original
📚 Editor's Note
On November 2 (US time), the day before the election, OpenAI announced the "Grid Protectors Program." It is an effort for data centers OpenAI builds to use less electricity when power supply is tight, and a figure has been given for one site, in Georgia. At the first Georgia facility (under construction), it will be able to cut up to 1,000 megawatts when the utility asks, out of an expected total need of about 3,200 megawatts for the facility.
How much it will cut and for how long is still being worked out site by site with utilities and others, and how AI processing will be handled while power is reduced has not been published. Whether AI response speed or usage limits will change on a hot afternoon or a cold morning will become clear only once each facility's terms are out.
See you next week, with useful information and something to think about. The dera news team