Show only each day's banner story

Jensen Huang leaves Nvidia after 33 years, Colette Kress becomes interim CEO

Nvidia will keep producing chips under interim leadership, but for the first time in 33 years it must find a new permanent leader.

Nvidia's board announced Jensen Huang's immediate resignation as president and CEO and his departure from the board. Chief financial officer Colette Kress has become interim CEO while retaining her previous role; the board is considering internal and external candidates for a permanent successor. Harvey Jones has been appointed the first chair of the board in the company's history. Nvidia said customer commitments and product plans would continue without interruption. According to NBC News, the decision followed reports about Huang's extramarital relationships and an investigation into his relationship with businesswoman Alice Huang, whom sources cited by news outlets linked to Chinese intelligence. NBC stressed that it had not independently verified the report of a second relationship. China rejected the espionage allegations as unsubstantiated. According to Congressman Brian Mast, Huang agreed to brief lawmakers on his contacts with Alice Huang and Nvidia's security protocols. Semafor, citing two officials, reported a US Justice Department antitrust investigation into Nvidia and two other investigations involving Megaspeed and Supermicro. According to Semafor, there were no signs at the time of possible criminal charges against Nvidia or Huang. Nvidia said it had no reason to believe national security or trade secrets had been compromised, but would continue its independent review.

The fast Gemini 4 model is now available to developers and all Search users, while complex agents can distribute work among models with different costs and reasoning depths.

Google releases Gemini 4 Hydrogen for fast agents and upgrades Search

The fast Gemini 4 model is now available to developers and all Search users, while complex agents can distribute work among models with different costs and reasoning depths.

Gemini 4 Hydrogen is designed for everyday programming and multi-step agent workflows where speed matters more than the extended reasoning of Gemini 4 Carbon and Argon. The model is available on all AI Studio and Gemini API tiers, in Gemini for business, the Gemini app, Google Search, and for paid Google AI Pro and Ultra subscribers in Antigravity. It has a one-million-token context window, with pricing independent of context length: $1.50 per million input tokens and $7.50 per million output tokens. According to Google, Hydrogen scored 51 on the Artificial Analysis Intelligence Index at its default high reasoning level, versus 50 for Argon, and 54 at the xhigh level. Compared with Gemini 3.8 Flash, it completes tasks 2.5 times faster and scores 57% on Terminal-Bench 4.1 versus 18.2%. These results and the claims about fewer errors come from the developer. The new Orchestrator in Antigravity switches between Carbon and Hydrogen within a session. Google's measurements show that this reduced quota consumption by 24% and execution time by 19% while achieving nearly the same success rate as Carbon. In Search, Hydrogen has replaced 3.8 Flash for all users worldwide: the system better understands multi-part questions, explores subtopics in parallel, and produces more visual answers with tables, images and additional links.

According to the company, the new capital will allow it to increase production of interoperating unmanned systems in several countries.

UFORCE raises $700 million at a $6 billion valuation to expand autonomous systems

According to the company, the new capital will allow it to increase production of interoperating unmanned systems in several countries.

UFORCE announced a $700 million Series A round at a $6 billion valuation, co-led by Valor Equity Partners and the UK's National Wealth Fund, according to CNBC. According to the company, it develops aerial, maritime and ground unmanned systems and SI to coordinate them. UFORCE CEO Oleg Rogynskyy said the funds would expand manufacturing in Ukraine and allied countries. In March, the company raised $50 million at a valuation of about $1 billion, according to CNBC. Rogynskyy said Ukrainian military orders would remain the priority as international supplies grow.

Sources 1

OpenAI has secured additional capital ahead of a possible IPO, but its spending and operating loss continue to grow alongside revenue.

OpenAI raises $15 billion at a $1.25 trillion valuation

OpenAI has secured additional capital ahead of a possible IPO, but its spending and operating loss continue to grow alongside revenue.

OpenAI has raised $15 billion at a $1.25 trillion valuation, CNBC reported, citing sources. A round of up to $30 billion at a $1.4 trillion valuation had previously been discussed; it is viewed as an interim step before a possible IPO in 2027. According to figures the company provided to investors, ChatGPT has about 100 million paying users, or 8.5% of its weekly active audience. Third-quarter revenue rose about 50% from the second quarter to $10 billion. Adjusted operating loss increased from $10.1 billion to $10.9 billion, excluding $2.6 billion in stock-based employee compensation.

Sources 1

SpaceX is expanding its computing infrastructure in two states at once, but the Arizona project's timeline depends on approvals from authorities and the local community.

SpaceX to invest more than $74 billion in two data centers in Texas and Arizona

SpaceX is expanding its computing infrastructure in two states at once, but the Arizona project's timeline depends on approvals from authorities and the local community.

SpaceX has announced two projects worth more than $74 billion in total. About $50 billion is earmarked for a data center in Cochise County, Arizona, and more than $24 billion for expanding the project in the Rio Grande Valley and the Starbase headquarters in Texas. According to Reuters, the Arizona project is expected to create more than 5,500 construction jobs and 1,000 jobs after launch, with a dry cooling system planned instead of groundwater extraction. The Texas facility will generate its own electricity; SpaceX has also pledged $220 million to expand the water supply in nearby Brownsville. The Arizona project still requires government approval, and a petition by local residents opposing it has gathered more than 2,000 signatures, Reuters reported.

Lenders are financing new computing capacity before construction begins, accelerating data center delivery for major SI customers.

Rowan secures more than $1.5 billion for a 5.5 GW data center pipeline

Lenders are financing new computing capacity before construction begins, accelerating data center delivery for major SI customers.

Rowan Digital Infrastructure upsized its term loan by $600 million in a deal led by Apterra Infrastructure Capital and opened a $590 million revolving credit facility and a $355 million letter of credit facility led by Citibank. The funds are intended for early project stages: acquiring land, securing power connections and ordering equipment with long lead times. The company's pipeline covers 5.5 GW of projects across several US regions; Rowan says more than 1.2 GW has already been contracted by technology customers. Over the past 18 months, the company has raised more than $8 billion from over 40 lenders.

Naver is turning computing infrastructure into a separate business with long-term revenue rather than merely support for its own services.

Naver wins its first SI factory contract worth $1.5 billion

Naver is turning computing infrastructure into a separate business with long-term revenue rather than merely support for its own services.

Naver signed a 2 trillion won ($1.5 billion) contract with a US company specializing in SI inference. The 20 MW agreement provides computing services on next-generation GPUs, with initial deliveries in 2027; revenue will begin after commercial launch in the first half of the year. Naver is developing its SI factory with Brookfield and Nvidia and plans to expand capacity from 55 MW in 2027 to 200 MW in 2028. The company has already secured contracted capacity above its 2028 target and is in talks with more than ten potential customers. At the same time, Naver's quarterly revenue rose 10.7% to 3.47 trillion won, while operating profit fell 20.9% amid investment.

Demand for SI computing is rapidly increasing Nebius revenue, but further growth requires even larger investment and debt financing.

Nebius increases quarterly revenue 6.7-fold but raises spending plan to $27 billion

Demand for SI computing is rapidly increasing Nebius revenue, but further growth requires even larger investment and debt financing.

Nebius reported third-quarter revenue of $980.8 million, up 571% year over year and above the LSEG consensus of $907.5 million. Net loss narrowed from $119.6 million to $37.5 million. The company expects full-year revenue at the low end of its $3–3.4 billion range. Its 2026 capital expenditure forecast was raised from $20–25 billion to $27 billion, while the 2027 range was set at $40–45 billion. Shares fell 16% after the report as investors worried about the scale of planned investment and the cautious capacity rollout forecast.

Sources 1

Baidu is simultaneously reducing the cost of complex reasoning and accelerating expressive video generation for enterprise developers.

Baidu releases a low-cost reasoning model and a micro-expression video model

Baidu is simultaneously reducing the cost of complex reasoning and accelerating expressive video generation for enterprise developers.

At Baidu World 2026, the company opened two application APIs through its Qianfan platform. ERNIE Lightning Deep Thinking uses Baidu Sparse Attention and a coordinated fast-and-slow reasoning scheme, compressing the KV cache during decoding. According to Baidu, optimization for Kunlun reduces inference engine costs by more than 30% and improves return on investment by 50% compared with other chips. ERNIE Yijing Micro-Expression generates facial expressions and movements from text, images and speech, supports scripts of up to 4,000 Chinese characters and uses an emotion description system with six dimensions and 200 subcategories. According to the company, its fast version generates one minute of video in about one minute and costs 50% less than the industry average.

Sources 1

Gemini is becoming a wearable assistant that sees the surroundings and handles everyday tasks without requiring users to take out a phone.

Samsung and Google open preorders for Gemini glasses starting at $395

Gemini is becoming a wearable assistant that sees the surroundings and handles everyday tasks without requiring users to take out a phone.

Samsung's Gemini glasses will go on sale in the United States and South Korea on December 10. Warby Parker models start at $395, while Gentle Monster models range from $449 to $489; 30 design variants are available. The devices weigh about 41 grams and include a 12 MP camera, open-ear speakers, four microphones and Snapdragon AR1 Gen 1, with up to nine hours of battery life. Gemini answers questions about what the wearer sees, translates, assists with navigation, saves whiteboard content to Samsung Notes and processes notifications. The camera is disabled if the external recording LED is covered, bystanders' faces are automatically redacted before Gemini responds, and Samsung says visual data is not used to train models or target ads. The full set of SI features requires a paired phone.

Scientists will be able to update SI climate forecasts faster and test more policy options without repeating months of expensive computations.

MIT cuts climate SI model training from eight full simulations to one

Scientists will be able to update SI climate forecasts faster and test more policy options without repeating months of expensive computations.

MIT researchers developed an algorithm that selects synthetic emissions scenarios for training climate emulators. These models simulate the behavior of full Earth system models in seconds and allow scientists to test thousands of climate variants. The new method requires one expensive full-model run instead of six to eight. The algorithm iteratively adjusts emissions trajectories until the emulator's error falls to a specified level; individual training scenarios do not need to be realistic. In tests published in PNAS, the model generalized better to unfamiliar situations, including local restrictions on one pollutant while others increased, whereas conventional emulators failed. Next, the authors plan to test the method on a scenario generated by a full climate model and on other model types.

Sources 1

Off-the-shelf SI agents drastically shortened the time needed to attack industrial equipment in a test, making basic controller protection even more essential.

Claude and Kimi shut down a test wind farm in under two hours

Off-the-shelf SI agents drastically shortened the time needed to attack industrial equipment in a test, making basic controller protection even more essential.

Shinobi Security replicated a wind farm with six turbines generating nearly 10 MW using the same types of controllers and network devices as a real site, Computer Weekly reported. Given VPN access, the models independently exploited default passwords, misconfigurations and an open repository of engineering documents without searching for unknown vulnerabilities. Claude disconnected the farm from the grid in 66 minutes using $20 worth of tokens, while Kimi achieved the same result in under two hours for $8. According to Computer Weekly, one agent spoofed weather station readings so that safety automation stopped the turbines. The research will be presented on November 16; the experiment highlights the importance of changing default passwords, updating controllers and restricting access to operational documentation.

US-EU cooperation on computing supply chains continues, but the scope of common rules and the independence of European regulation remain to be negotiated.

EU objects to expansion of the Pax Silica alliance

US-EU cooperation on computing supply chains continues, but the scope of common rules and the independence of European regulation remain to be negotiated.

The United States is asking Pax Silica members to sign a declaration in San Francisco that would expand the alliance for SI, semiconductor and critical-mineral supply chains into a broader economic security pact, according to Euronews. The draft provides for supply-chain diversification, coordinated responses to economic coercion and a regulatory working group. Some EU countries consider the definitions of sensitive technologies and critical infrastructure too broad and fear additional US restrictions. A separate dispute concerns the term "superintelligence": the European Commission proposes treating it as equivalent to SI as already defined in the AI Act. EU governments have given the Commission a mandate to negotiate, but according to Euronews, authority to sign may not be agreed in time for the ceremony.

Sources 1

SI developers could gain a clear legal way to use creative works, while authors would receive a payment mechanism and the right to opt out.

Russian presidential council proposes author compensation for SI training and labels for generated content

SI developers could gain a clear legal way to use creative works, while authors would receive a payment mechanism and the right to opt out.

The Council for Codification and Improvement of Civil Legislation drafted amendments that explicitly classify digitization, storage and processing of works for neural network training as use of copyrighted and related-rights material. Three mechanisms are proposed: a contract with the author, an open license or use without consent with guaranteed remuneration through accredited collective management organizations. Authors could prohibit use of their works through such an organization's website. The draft also requires labels on materials produced by generative systems and recognizes rights to generated results when a person has made a creative contribution. The amendments synchronize the Civil Code with Federal Law No. 243-FZ on sovereign and national models and are planned to take effect on March 1, 2027.

SpaceX is attempting to build its own SI computing platform for Earth and orbit instead of buying a ready-made server rack.

SpaceX designs Kessel rack with an 84-layer board instead of NVLink cables

SpaceX is attempting to build its own SI computing platform for Earth and orbit instead of buying a ready-made server rack.

SpaceX and xAI are developing the Kessel computing system with 36 NVIDIA Vera CPUs and 72 Rubin GPUs while retaining the NVL72 topology, according to SemiAnalysis. Instead of NVLink cable assemblies, compute and switch modules are arranged around a passive 84-layer backplane. The platform is designed for terrestrial and orbital data centers. SpaceX intends to buy individual packaged chips and network controllers from NVIDIA while ordering boards and other components itself. This will allow the design to be adapted for launch and orbital heat dissipation but will require in-house validation and supply-chain management. According to the analysis, economic efficiency will depend on the balance between integration savings and the cost of complex multilayer boards.

Sources 1

Security teams can use agents to continuously find vulnerabilities and prepare tested fixes while engineers retain the final decision.

OpenAI introduces Daybreak for automated vulnerability discovery and remediation

Security teams can use agents to continuously find vulnerabilities and prepare tested fixes while engineers retain the final decision.

OpenAI has launched two products on its Daybreak cybersecurity platform, according to The Register. In early access, Red Team uses a fleet of agents to continuously test web applications: they analyze source code and application context, investigate attack paths, validate vulnerabilities from scanners and researcher reports, and test fixes. Defense Factory automates the defensive loop of finding, validating and fixing vulnerabilities. Agents based on GPT-6 Cyber scan codebases and prepare patches; before deployment, engineers review and approve a ready merge request together with a functionality test. The Register presents OpenAI's claim of "superhuman" testing scale as the company's position.

The court will decide whether a lawsuit can restrict the next generation of models from five major developers based on alleged risks.

Nonprofits sue five SI developers to stop new models

The court will decide whether a lawsuit can restrict the next generation of models from five major developers based on alleged risks.

The Alliance for Secure AI, Irreplaceable Education Fund, Public Citizen and the Heat Initiative have filed a lawsuit in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI and SpaceXSI. The plaintiffs seek to block the development and deployment of systems exceeding the companies' most advanced models at the time of filing, including recursive self-improvement systems in which SI conducts further research and development itself. The lawsuit alleges that the companies know their control methods are unreliable yet continue to create technologies with catastrophic risks, including cyberattacks and biological weapons. This is the plaintiffs' legal position, not a fact established by the court. According to CBS News, the defendants had not responded to the outlet's request for comment by publication time.

The chip-location tracking requirement now has evidence of technical feasibility from the manufacturer itself, although Nvidia has yet to provide an official written position.

Nvidia privately confirms to congressman that SI chip tracking is technically feasible

The chip-location tracking requirement now has evidence of technical feasibility from the manufacturer itself, although Nvidia has yet to provide an official written position.

Congressman Bill Foster told MIT Technology Review that Nvidia representatives confirmed during technical discussions that the anti-smuggling measures in the Chip Security Act could be implemented. The bill provides for verifying the location of advanced SI chips through inspections, company attestations or a signal reporting where a chip is switched on. It is among the candidates for inclusion in the US defense budget. In March, the Semiconductor Industry Association, of which Nvidia is a member, called such requirements new, untested and potentially infeasible. Nvidia already offers Fleet Intelligence to monitor chip performance across data centers and confirmed to Reuters that the system can identify the country where chips are located. The company did not respond to MIT Technology Review's request for comment on Foster's remarks; the congressman clarified that he had not received written confirmation from Nvidia.

Investors are willing to fund infrastructure that helps companies move from experimental chatbots to their own working agents.

LangChain raises $300 million at a $3 billion valuation

Investors are willing to fund infrastructure that helps companies move from experimental chatbots to their own working agents.

CapitalG, Alphabet's investment arm, led the round, according to The Wall Street Journal. LangChain's valuation rose from $1.25 billion in 2025 to $3 billion. The company develops a platform that large organizations use to build and deploy their own SI agents.

Capital for new SI companies continues to grow not only at model laboratories but also at applied teams across different industries.

Norwest closes a $4 billion fund for investments including SI startups

Capital for new SI companies continues to grow not only at model laboratories but also at applied teams across different industries.

Norwest's eighteenth fund is its largest in the firm's 65-year history. Backed exclusively by Wells Fargo, Norwest will invest in companies at different stages in the United States, India, Israel and other countries. Its priorities include SI applications and infrastructure, physical and voice SI, agentic cybersecurity, domain-specific models and aerospace projects. The firm's previous fund, closed in 2024, totaled $3 billion.

SI infrastructure now accounts for more than half of revenue at the world's largest contract electronics manufacturer.

Foxconn profit rises 34% on demand for SI servers

SI infrastructure now accounts for more than half of revenue at the world's largest contract electronics manufacturer.

Hon Hai Precision Industry, known as Foxconn, reported record quarterly net profit of NT$77.52 billion ($2.35 billion), up 34% year over year. Revenue rose 46% to NT$3.02 trillion. The cloud and networking division generated 54% of revenue, up from 51% in the previous quarter. The company linked its better-than-expected results to demand for SI servers and racks, although gross margin fell 0.33 percentage points year over year to 6.02%.

New Chinese teams are designing processors for models and general-purpose agents rather than for individual tasks.

China's Tongxin raises hundreds of millions of yuan for an AGI chip

New Chinese teams are designing processors for models and general-purpose agents rather than for individual tasks.

Tongxin Intelligent completed seed and angel rounds totaling several hundred million yuan. Investors include Tsinghua Alumni Seed Fund, Beijing Institute for General Artificial Intelligence, Songlin Technology, Pingzhi Information, Dark Matter AI and Yuzhang Capital. The funds will support core architecture design and research before prototype chip manufacturing. Spun out of Beijing Institute for General Artificial Intelligence, the company is developing a computing architecture for models, world models and general-purpose agents.

Demand for SI is already visible in macroeconomic statistics, not only in the revenues of individual technology companies.

UK economy grows 0.7% as the contribution of SI and cloud becomes visible

Demand for SI is already visible in macroeconomic statistics, not only in the revenues of individual technology companies.

UK GDP grew 0.7% in the third quarter after a revised 0.5% in the second, with annual growth of 1.9%, according to the Office for National Statistics. Services output increased 0.9%, while computer programming and consultancy grew 4.4%, lifting the information and communication sector by 2.1%. The ONS called programming and consultancy the largest driver of GDP growth in 2026 and noted that many companies reporting rapid turnover growth work in SI and cloud computing. Statisticians said the precise contribution of these technologies remains difficult to measure.

Sources 1

Indian companies can use OpenAI models for working data without sending supported computations outside the country.

OpenAI launches local API request processing in India

Indian companies can use OpenAI models for working data without sending supported computations outside the country.

OpenAI expanded its Indian data residency program: supported API requests and model computations now take place within the country rather than merely being stored there. Local inference initially supports GPT-6 Luna through the Responses API and Chat Completions API. Customers must create a separate project, select India for storage and processing and use the regional endpoint. The offering is intended for organizations with sensitive-data localization requirements, particularly financial companies. Data storage in India has been available since May 2025.

Sources 1

Alice AI is moving from advice to completing multistep assignments while leaving decisions about payments, bookings and account access to the user.

Alice AI begins handling entire assignments in "Act!" mode

Alice AI is moving from advice to completing multistep assignments while leaving decisions about payments, bookings and account access to the user.

Yandex gave several thousand active Alice Plus subscribers early access to the personal agent "Act!". It breaks tasks into steps, selects tools, makes phone calls, works with websites through a separate cloud-based Yandex Browser and collects files on Disk. Users can see its progress and call transcripts. Plugins connect the agent to Yandex and third-party services, while custom sources can be added through MCP. Alice AI requests confirmation before purchases, bookings or orders and hands control of the page to the user when account login is required. The mode is expected to become available to all subscribers later.

Sources 1

Verifiable data on actual use of the SI system has prompted a review of public claims and the terms of a major government contract.

NHS England removes claim that 137 hospital trusts benefit from Palantir

Verifiable data on actual use of the SI system has prompted a review of public claims and the terms of a major government contract.

NHS England removed a page describing 137 hospital trusts as live with Palantir's Federated Data Platform and already benefiting from it. In response to freedom of information requests from Medact, 29 organizations on the list said the system was not live or that they were not using all or most of its applications. The Office for Statistics Regulation is examining NHS claims about the system's adoption; the regulator had previously noted a lack of evidence for the claimed 117,000 additional operations. A possible extension of Palantir's £330 million contract through 2029 remains under discussion. Palantir says ending the contract would put politics above patients, while NHS England did not respond to the Guardian's questions.

Japan is considering a shift from recommendations to mandatory action following incidents involving advanced models.

Japan proposes penalties for refusing to cooperate after advanced SI incidents

Japan is considering a shift from recommendations to mandatory action following incidents involving advanced models.

A government working group discussed mandatory safety measures for companies working with advanced SI, according to Kyodo News. After incidents, authorities could issue legally binding orders to companies that fail to cooperate with investigations, with penalties proposed for noncompliance. The discussion concerns incidents involving advanced SI, while the companies to be covered have yet to be determined. Japan's law in force since May 2025 allows investigations into abuses and guidance to businesses but contains no penalties or binding measures. The working group aims to present an overall outline of possible changes by the end of the year.

The EU intends to use SI in public administration only with verifiable human accountability and dedicated parliamentary oversight.

European Parliament demands human oversight of SI in democratic processes

The EU intends to use SI in public administration only with verifiable human accountability and dedicated parliamentary oversight.

The European Parliament adopted a resolution on the institutional aspects of SI by 405 votes to 133, with 60 abstentions. The document calls for parliamentary scrutiny of SI's impact on democratic processes, protection of fundamental rights and human oversight in drafting and evaluating legislation. MEPs also backed European technological autonomy, an EU-led international agreement on SI risks and a budget-neutral European Parliament SI observatory. The resolution proposes updating electoral law and mechanisms protecting citizens' rights when public authorities use SI.

If the amendments are adopted, small companies will be able to offset part of the cost of implementing Russian SI services through tax deductions.

Russia's Ministry of Digital Development discusses a tax deduction of up to ₽500,000 for domestic SI

If the amendments are adopted, small companies will be able to offset part of the cost of implementing Russian SI services through tax deductions.

Russia's Ministry of Digital Development and Ministry of Finance are working out a profit tax and simplified tax system deduction for small and medium-sized enterprises that buy, use or create Russian SI solutions in 2027–2029. The proposed limit is ₽200,000 of expenses per year, rising to ₽500,000 for implementations with proven effectiveness, according to an Interfax source. A government working group has already discussed the initiative, but its funding parameters are still being coordinated. Expected lost budget revenue is estimated at ₽14.7 billion per year and ₽44.1 billion over three years. Maksut Shadaev said the measure should stimulate small-business demand for services based on national and sovereign models.

Sources 1

Autonomous platforms are moving beyond demonstrations to large vessels capable of independently carrying out extended surveillance missions.

HD Hyundai and Anduril launch a 600-ton autonomous vessel

Autonomous platforms are moving beyond demonstrations to large vessels capable of independently carrying out extended surveillance missions.

HD Hyundai launched an autonomous surface vessel developed with Anduril in Ulsan. The VAone system provides autonomous navigation, while the Lattice platform supports reconnaissance, surveillance and intelligence-gathering missions without a crew. Following sea trials off Ulsan, the company plans further tests in the United States early next year and operational data collection for entry into the US Navy's unmanned vessel market. In September, the Navy announced plans to procure 30 medium-sized autonomous vessels.

Sources 1

Demand for networks serving large SI centers has begun generating significant revenue for Cisco after several years of lagging accelerator manufacturers.

Cisco expects annual SI infrastructure revenue to double to $9 billion

Demand for networks serving large SI centers has begun generating significant revenue for Cisco after several years of lagging accelerator manufacturers.

Cisco's revenue for the first quarter of fiscal 2027 rose 24% to $18.5 billion from $14.88 billion a year earlier, beating the LSEG consensus of $18.1 billion. Net income increased from $2.9 billion to $4.6 billion, and adjusted earnings per share were $1.40 versus expectations of $1.33. Its forecast for hyperscaler SI infrastructure revenue in the current fiscal year has been raised to $9 billion, more than double the fiscal 2026 result. For the next quarter, Cisco expects revenue of $18–18.2 billion and adjusted earnings per share of $1.34–1.36, both above analyst estimates. Shares rose after the report.

Sources 1

Cerebras has gained a major cloud customer and is accelerating sales as it seeks a share of the SI accelerator market alongside Nvidia.

Cerebras grows revenue 70% and announces Microsoft deal

Cerebras has gained a major cloud customer and is accelerating sales as it seeks a share of the SI accelerator market alongside Nvidia.

Cerebras' quarterly core revenue grew 70% to $230.9 million, beating the LSEG expectation of $214.7 million. Loss per share was 67 cents. For the fourth quarter, the company forecasts revenue of $310–330 million with a core gross margin of 47–49%, and it expects annual revenue to triple in 2027. Cerebras CEO Andrew Feldman announced an agreement with Microsoft signed in October; the company expects the first revenue from it in the first quarter of 2027. Details were due to be disclosed on the analyst call. Cerebras shares rose 17% in extended trading after the report, although they remained down about 47% since the May Nasdaq listing.

Sources 1

Nscale is going public with rapidly growing revenue, but investors will have to assess billion-dollar losses and dependence on a few major customers.

Nscale discloses loss widening to $1.69 billion in updated IPO prospectus

Nscale is going public with rapidly growing revenue, but investors will have to assess billion-dollar losses and dependence on a few major customers.

Nscale has updated its paperwork for a listing on the New York Stock Exchange. Third-quarter revenue grew 1,344% year over year to $307.5 million, but net loss increased from $1.02 billion in the second quarter to $1.69 billion. Microsoft previously withdrew from an Nscale project in England over access to electricity and decided not to renew a contract in Texas, according to the Financial Times. Nscale nevertheless expects Microsoft to remain a significant customer in the future. The company also cites additional demand from Anthropic's September decision to expand its computing capacity order at Nscale's West Virginia site.

Sources 1

The simultaneous failure of email, workplace collaboration and Copilot again demonstrated organizations' dependence on Microsoft's unified cloud infrastructure.

Microsoft outage disrupts Outlook, Teams and Copilot

The simultaneous failure of email, workplace collaboration and Copilot again demonstrated organizations' dependence on Microsoft's unified cloud infrastructure.

Users in the Americas reported problems signing in to Exchange Online, sending and receiving email, and disruptions to Teams, Copilot and other Microsoft 365 tools. At its peak, Downdetector received more than 130,000 reports about Outlook, more than 46,000 about Microsoft 365 and about 3,000 about Teams. These figures reflect the number of reports, not the exact number of people affected. Microsoft identified unusually high load on part of its infrastructure as a possible cause and began redistributing traffic to restore connections. The company had earlier declared a separate Teams outage in North America resolved, although some users continued to report problems.

France is creating a dedicated venue for joint military SI development before its status within NATO has been formalized.

France establishes defense SI resilience center in Rennes

France is creating a dedicated venue for joint military SI development before its status within NATO has been formalized.

France's Ministry of the Armed Forces has announced a defense SI resilience center together with partner nations. The center in Rennes is expected to become operational in 2027 and will develop military use of SI and the exchange of practices among allies. France intends to seek NATO recognition for the center. The project was initially developed along the lines of NATO Centres of Excellence. When the alliance paused the creation of new centers to review its policy, France decided to maintain the schedule and launch it outside NATO's structure. According to Bertrand Rondepierre, director of the ministry's SI Agency, the goal is to accelerate the mastery and adoption of SI to support military operations alongside allies.

Sources 1

Sberbank is shifting the next stage of investment from individual SI services to models and its own computing infrastructure.

Sberbank places SI at the center of its 2027–2029 strategy

Sberbank is shifting the next stage of investment from individual SI services to models and its own computing infrastructure.

Sberbank's new strategy for 2027–2029 retains a target return on equity of at least 21% and places the main emphasis on SI and related tools. The previous three-year strategy called for ₽600 billion of SI investment; in the new one, German Gref announced major investment in the bank's own data centers and allowed for acquisitions in that market. According to Gref, Sberbank's income from using SI over the previous three years was ₽1.5 trillion. He forecasts that this figure will decline over the next three years because model development will require more investment. The bank also intends to continue selling loss-making non-core assets, retaining subsidies only for strategic areas.

Before buying through a link provided by a chatbot, shoppers should verify the store's address: even well-known models sometimes display fake or unsafe sites.

Podqi test finds risky shopping links in answers from ChatGPT, Gemini and Perplexity

Before buying through a link provided by a chatbot, shoppers should verify the store's address: even well-known models sometimes display fake or unsafe sites.

Podqi asked ChatGPT, Gemini and Perplexity shopping questions about 151 brands and analyzed 6,795 answers across two test periods in September. Risky sites included impostor websites, sellers of counterfeit goods, re-registered old domains and inactive addresses that could be taken over. Such sites appeared in 15% of Perplexity answers, 8% of ChatGPT answers and 4% of Gemini answers; these shares include recommendations, neutral mentions and warnings. Risky links appeared about half as often for large brands, while ChatGPT issued a warning in 24% of cases where such sites appeared. Google and OpenAI told The Verge that they were continuing to improve link quality and protection against deceptive sites; Perplexity said it did not promise 100% accuracy but considered accuracy a priority. The study was conducted by intellectual-property protection company Podqi, and one of the tested brands was its client.

Sources 1

Spacecraft have taken the first step from deploying ready-made structures to autonomously assembling new objects in orbit.

Robot assembles a structure in space without astronaut assistance for the first time

Spacecraft have taken the first step from deploying ready-made structures to autonomously assembling new objects in orbit.

A team led by Sergio Pellegrino at the California Institute of Technology built a robotic arm that joined parts of a radio antenna in orbit without astronaut assistance. The work was carried out under a DARPA initiative exploring new methods of construction in space. Previously, large orbital structures were either assembled by astronauts or launched folded for automatic deployment. According to the researchers, autonomous assembly opens the way to more complex space stations and other structures that cannot simply be unfolded after launch.

Explicitly separating commands from data gives small open models a way to ignore hidden instructions in documents without a separate external filter.

IBM releases Granite 4.1 with LLMON markup against hidden injections

Explicitly separating commands from data gives small open models a way to ignore hidden instructions in documents without a separate external filter.

IBM has introduced the experimental granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models. LLMON marks instructions, data and constraints as separate blocks using reserved special tokens, while an execution block explicitly links a command to the relevant data. This lets the model distinguish commands from document content and tool outputs. Lossless JSON conversion tools and examples for adapting the markup to new tasks have also been released. In IBM's tests of static indirect injections, LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open model, while retaining the general capabilities of Granite 4.1 Instruct. The company emphasizes that robustness is learned behavior, not a guarantee, and that adaptive multi-turn attacks require further work. The models are intended for experimentation, not production use.

Russian companies have another way to embed controlled SI agents in internal systems without moving workflows to an external cloud.

VK AI Space 26.3 combines agents, quality checks and offline operation

Russian companies have another way to embed controlled SI agents in internal systems without moving workflows to an external cloud.

VK Tech released VK AI Space 26.3, an enterprise platform comprising seven modules. It connects language and multimodal models, knowledge bases, single-agent and multi-agent assistants, quality evaluation and security mechanisms. It can be deployed on a customer's own hardware without external network access. Agents access APIs, databases and files, execute multistep processes and pause for human confirmation when necessary. Administrators can test answers on evaluation datasets, restrict actions, filter requests and tool calls, assign roles and quotas and review audit logs.

Russian developers will be able to use generative tools without a separate registry prohibition while retaining responsibility for the safety of the final product.

Russia's Ministry of Digital Development will not restrict SI-created software in the national registry

Russian developers will be able to use generative tools without a separate registry prohibition while retaining responsibility for the safety of the final product.

Maksut Shadaev said at CNews Forum that generative development alone would not prevent a product from entering the Russian Software Registry. The Ministry of Digital Development estimates that about 90% of software companies already use such practices, but mostly for prototypes and individual tasks rather than major new products. The minister separately stressed the need to control the final code, token costs and a specific company's responsibility for developing the product. Criteria for sovereign and national models are still being formulated in secondary legislation; Shadaev estimated that Russian large models lag the leaders by 6–12 months.

Sources 1

Retailers must admit useful SI bots while distinguishing them from malicious traffic to prevent sales events from ending in downtime.

StormWall: one third of DDoS attacks on retailers disguise themselves as SI traffic

Retailers must admit useful SI bots while distinguishing them from malicious traffic to prevent sales events from ending in downtime.

The number of DDoS attacks on Russian retailers rose 2.5-fold year over year and fourfold compared with the same period in October between November 1 and 11, according to StormWall. Bots disguised as SI service traffic participated in 34% of attacks; retailers allow such traffic for indexing and visibility in model answers. Average attack volume was 230 Gbps, with a maximum of 1.2 Tbps. StormWall linked the increase to November sales and said most campaigns involved ransom demands; some small stores were unavailable for several hours.

Sources 1

Growing SI infrastructure is creating commercial demand for new plants producing lower-carbon materials and for electric freight.

Meta funds lower-carbon materials and transport for data centers

Growing SI infrastructure is creating commercial demand for new plants producing lower-carbon materials and for electric freight.

Meta signed long-term purchasing agreements for data center construction and supply chains. A contract with Sublime Systems will support production of at least 180,000 tons of lower-carbon cement and the company's first commercial-scale plant; an agreement with Stegra covers primary steel produced without coal in the iron reduction process. The company is also expanding electric freight routes and open-access charging stations in Virginia, Texas, California and Arizona. A contract with MacroCycle Technologies will support a plant in the southeastern United States that processes up to 5,000 tons of hard-to-recycle plastic into PET annually. Some agreements use environmental attribute certificates: the supplier produces verified lower-carbon materials and Meta pays for the environmental benefit.

Sources 1

Muse can now handle routine phone errands independently, but only for paid subscribers and only in the US.

Muse starts calling US businesses for paid subscribers

Muse can now handle routine phone errands independently, but only for paid subscribers and only in the US.

Meta has made phone calls to businesses on a user's behalf available to all paid Muse subscribers in the US. The agent can call, for example, an internet provider, the DMV or a restaurant without online reservations. The feature was previously tested with a limited group of users. Subscriptions start at $20 per month for 500 million Muse tokens per week; the $100 tier provides 3 billion tokens. Meta said last month that Muse had more than 10 million users.

A small company can delegate purchasing and customer acquisition to an agent while retaining approval of every payment and advertising campaign.

Replit Agent learns to buy products and run ads for small businesses

A small company can delegate purchasing and customer acquisition to an agent while retaining approval of every payment and advertising campaign.

Replit has added two features for owners of companies built on its platform. Through Stripe's Link and Visa, the agent finds products, fills a cart on a merchant's website and stops before payment so the owner can approve the final amount. The browser session is isolated, and permission is granted for a specific purchase rather than arbitrary spending. The feature is available to Core and Pro subscribers in the US for one-time purchases. The second feature creates and launches Google Ads campaigns across Search, YouTube, Display and Discover. The agent analyzes the website, prepares copy and images, suggests a budget and adds conversion tracking; campaigns launch only after user review. The advertising tools are available to Pro subscribers worldwide.

Sources 1

VTB already uses SI for most routine inquiries and plans to gradually allow it to perform personalized financial actions with customer consent.

VTB's SI assistant handles 87% of customer inquiries in branches

VTB already uses SI for most routine inquiries and plans to gradually allow it to perform personalized financial actions with customer consent.

According to VTB deputy president and management board deputy chairman Georgy Gorshkov, the internal assistant helps branch staff resolve 87% of customer questions, with customer satisfaction at 97%. In the contact center, the system analyzes interaction histories and suggests answers to operators, which the bank estimates saves more than 15,000 working hours per year. The assistant combines source retrieval and large language models to prepare answers from a knowledge base. In 2027, VTB plans to add scenarios for savings, investments, purchases and loyalty programs, as well as individual actions with customer permission. Gorshkov stressed the need for continuous model testing and monitoring.

SGLang developers now have another way to move model serving and fine-tuning to AWS Trainium infrastructure without changing their familiar API.

AWS releases SGLang-Neuron for running models on Trainium chips

SGLang developers now have another way to move model serving and fine-tuning to AWS Trainium infrastructure without changing their familiar API.

SGLang-Neuron is now publicly available in the AWS Deep Learning Container. The plugin from RadixArk and Annapurna Labs makes it possible to serve models and generate reinforcement-learning rollouts on AWS Trainium through the familiar SGLang server interface. Trainium support can evolve separately from the SGLang core thanks to the external plugin design. The plugin uses native PyTorch and Neuron Kernel Interface kernels, with prefill and decode separated through NIXL over Elastic Fabric Adapter. This allows context processing and generation to be distributed among Trainium servers while retaining SGLang's scheduler and request lifecycle.

Bing users can choose between SI summaries and ordinary links instead of a mandatory search format.

Bing gets an extension to disable SI answers in search

Bing users can choose between SI summaries and ordinary links instead of a mandatory search format.

Microsoft released an extension for Edge and Chrome that adds an SI feature toggle to Bing and restores traditional results without summaries and chat. The extension is already available, but installation makes Bing the default search engine. Users can also add the "-ai" operator to an individual query to disable SI.

Sources 1