Show only each day's banner story

Anthropic reports $17 billion in quarterly revenue and $895.4 billion in commitments ahead of IPO

Anthropic is growing revenue rapidly, but its future is now tied to hundreds of billions of dollars in compute commitments that it will have to service after going public.

In its amended Form S-1, Anthropic disclosed third-quarter revenue of roughly $17 billion, up about 1,127% year over year and 47% quarter over quarter, following 144% growth in the second quarter and 132% in the first. Its quarterly net loss reached $23.29 billion. Revenue for the first nine months was $33.24 billion, with a net loss of $104 billion, most of it tied to changes in the fair value of convertible notes and derivative liabilities. Purchase and lease commitments reached $895.4 billion, almost entirely for compute and infrastructure. Related expenses rose by $14.4 billion over nine months, up 329% year over year. Released ahead of the IPO roadshow, the filing adds not only rapid growth to the company's profile but also the scale of its long-term obligations to compute providers.

Sources 1

cnbc.com

Optical links between chips are becoming as strategic to SI infrastructure as the accelerators themselves, and Marvell is making the biggest bet in its history on that shift.

Marvell agrees to acquire optical component maker Lumentum for $115 billion in stock

Optical links between chips are becoming as strategic to SI infrastructure as the accelerators themselves, and Marvell is making the biggest bet in its history on that shift.

Lumentum shareholders will receive 4.5 Marvell shares for each share they own. The companies put Lumentum's enterprise value at $115 billion including debt, or about $1,132 per share, a 23.9% premium to the October 28 close before talks were reported. Marvell shareholders will own roughly 67% of the combined company and Lumentum shareholders 33%. Both boards unanimously approved the deal, which still requires shareholder and regulatory approval and is expected to close in the first half of 2028. Marvell is combining its custom SI chips and networking products with Lumentum's lasers, optical transceivers and switches. The companies are targeting a $550 billion market and see potential annual savings of more than $1 billion. Because the consideration is entirely in stock, the final value will depend on Marvell's share price.

Sources 2

cnbc.comthenextweb.com

OpenAI is moving its models beyond standalone chats into energy, investment and government checks, with local rules and officials' decisions remaining part of the process.

OpenAI signs four agreements with Abu Dhabi's government and companies

OpenAI is moving its models beyond standalone chats into energy, investment and government checks, with local rules and officials' decisions remaining part of the process.

ADNOC will become the first energy company to deploy the OpenAI Frontier platform through G42's Inception42. Mubadala, Inception42 and OpenAI will apply the models to investment analysis, opportunity discovery and risk assessment. A separate five-year agreement covers Frontier deployments across UAE government, business and education. The Abu Dhabi Government is launching the Machine-Readable Government initiative, converting policies and regulations into a form SI systems can process. The first pilot, with the agriculture and food safety authority, checks permit and certificate applications for food imports. Every assessment can be traced to the underlying rule, while officials retain final decision-making. Computation runs on GPUs in UAE data centers through OpenAI's local inference service.

Sources 2

cnbc.comthenextweb.com

South Korea is channeling semiconductor windfalls into a long-term national program for sovereign models, compute, robotics and workforce training.

South Korea proposes record 2027 budget with 21.3 trillion won for SI projects

South Korea is channeling semiconductor windfalls into a long-term national program for sovereign models, compute, robotics and workforce training.

President Lee Jae Myung presented parliament with an 820.9 trillion won budget, up 12.8% from the previous one. It allocates 21.3 trillion won to three areas: semiconductors, physical SI and national SI data centers, almost double the 2026 level. Yonhap puts a separate sovereign model program at 4.6 trillion won, while CNBC reports 4.7 trillion won. The plan also calls for training 2,000 chip-design specialists, expanding SI education and company-linked degree programs leading to employment. Defense spending would rise 8.2% to 73.3 trillion won, including hybrid systems combining people, robots and drones. The government aims to pass the budget by December 2.

Sources 2

cnbc.comen.yna.co.kr

Oracle is increasing both debt and share count to build cloud capacity, leaving investors to judge whether SI revenue will outpace the growing financial burden.

Oracle raises $20 billion in shares and bonds to expand SI infrastructure

Oracle is increasing both debt and share count to build cloud capacity, leaving investors to judge whether SI revenue will outpace the growing financial burden.

Oracle's filings split the offering equally between common stock and bonds. The bond issue comprises six tranches maturing from 2029 to 2046, one with a floating rate and five with fixed rates. Proceeds are intended for general corporate purposes, including capital expenditure, debt repayment and acquisitions. As of August 31, Oracle had $125.34 billion in notes payable and other borrowings, compared with revenue of $67.36 billion in fiscal 2026. It expects revenue to exceed $225 billion in fiscal 2030 and also said this week that capital expenditure would peak in the current and next fiscal years.

Sources 1

theglobeandmail.com

South Korean researchers gain a next-generation national computing resource designed for both classical science and training large SI models.

South Korea launches 614.7-petaflop Hangang supercomputer

South Korean researchers gain a next-generation national computing resource designed for both classical science and training large SI models.

The country's sixth national supercomputer cost 382.5 billion won ($276 million) and replaces the 2018 Nurion system. It contains 8,496 GPUs, including Nvidia GH200 Grace Hopper chips, and 9,936 CPUs. Theoretical peak performance is 614.7 petaflops, 24 times its predecessor, with 10 times the storage capacity and four times the network throughput. Hangang will run scientific simulations and complex SI workloads, including the K-Moonshot program applying advanced models to major research challenges. The Ministry of Science and ICT expects it to rank in or near the global top 10 in the Top500. Full service for researchers begins December 16 after pilot operations.

Sources 1

koreatimes.co.kr

Keeping the model and operating system in non-volatile on-chip memory lets a device begin recognition almost immediately and use less energy moving data.

Tohoku University and Aisin build chip that runs SI without external memory

Keeping the model and operating system in non-volatile on-chip memory lets a device begin recognition almost immediately and use less energy moving data.

The prototype 16nm system-on-chip integrates a dual-core Cortex-A53, Mali-G31, an SI accelerator and 32MB of non-volatile STT-MRAM. Boot data and model weights remain on the die, eliminating transfers from an SSD to DRAM. With Tiny YOLOv3, kernel startup took 65.7 ms instead of 2,535.3 ms, frame processing took 58.5 ms instead of 594.6 ms, and the full path from power-on to the first detection took 124.2 ms instead of 3,129.9 ms. Claimed energy consumption was 50.7 mJ per frame, about 1/50 of the comparison setup. Hybrid 8/16-bit arithmetic and weight reuse reduced weight data from 17.7 to 9.36MB, while mAP accuracy in the demonstration fell from 0.325 to 0.316. This is an evaluation-board result with a lightweight model, presented at IEEE A-SSCC 2026, not a production automotive chip. The researchers see applications in vehicles, robots and industrial devices requiring instant startup.

Sources 2

aisin.comxenospectrum.com

A large SI campus could temporarily return up to a gigawatt of power to the grid instead of only consuming electricity, provided flexibility is built into its design and contracts.

OpenAI pledges up to 1 GW of flexible demand for Georgia's power grid

A large SI campus could temporarily return up to a gigawatt of power to the grid instead of only consuming electricity, provided flexibility is built into its design and contracts.

According to OpenAI, the Grid Protectors Program will establish agreements with utilities and regulators to reduce future compute campuses' electricity use during grid shortages. The first site is the Effingham County project: OpenAI commits to paying all required infrastructure and electric-service costs, while its agreement with Georgia Power provides up to 1,000 MW of flexible demand. Depending on the site, batteries, on-site generation and demand controls may be used. Georgia Power estimates that its portfolio of large-load agreements, including this project, could save customers about $950 million annually from 2029 and at least $15 per month for a typical household using 1,000 kWh. OpenAI stresses that flexibility complements rather than replaces new generation and transmission.

Sources 1

openai.com

Even the strongest models still complete only a small share of long specialist tasks, while the new benchmark measures their usefulness and cost on work closer to real practice.

GDPx benchmark gives top model just 12% on professional tasks

Even the strongest models still complete only a small share of long specialist tasks, while the new benchmark measures their usefulness and cost on work closer to real practice.

SemiAnalysis released 206 tasks across 11 domains: chip and hardware design, cybersecurity, biology, finance, law, quantitative trading, materials science, alignment, software engineering and recursive self-improvement. Fourteen specialist partners built the tasks; the score averages pass@1 across three attempts with equal weight for each domain. Claude Opus 5.5 leads at 12.0% with an average cost of $51.41 per task, Gemini 4 Carbon scored 10.9% at $16.29, and GPT-6 Astra scored 9.6% at $53.99. Every model received zero in finance and law under strict binary grading, so partial scores are shown for those domains. SemiAnalysis cautions that safety classifiers lowered Opus and Astra's results in cybersecurity, biology and alignment. Benchmark updates are planned every three to six months.

Sources 2

gdpx.semianalysis.comunite.ai

Russian compute capacity gains a state emergency-reserve mechanism, while domestic services may be supported by foreign data centers when necessary.

Putin decree requires non-state data centers to reserve 10% of space and capacity

Russian compute capacity gains a state emergency-reserve mechanism, while domestic services may be supported by foreign data centers when necessary.

The reserve must be created within 10 days of the decree's signing. Operators can continue using the capacity until a government subcommission decides otherwise; after such a decision, they must make it available to support Russian information resources within 24 hours. The subcommission will also set a price ceiling, while Russia's Ministry of Digital Development inventories available space and capacity. The decree permits temporary use of foreign compute capacity to operate resources belonging to Russian citizens and legal entities. The decision will be made by the subcommission under the Government Commission for Digital Development in coordination with the FSB.

Sources 2

rg.ruvedomosti.ru

App-store review cannot guarantee safety when a server can replace an innocuous interface after approval with a generator of nonconsensual intimate images.

Wired uncovers network of over 400 apps hiding sexual deepfake features

App-store review cannot guarantee safety when a server can replace an innocuous interface after approval with a generator of nonconsensual intimate images.

Wired's investigation connected hundreds of domains and apps posing as pet-photo editors, language courses or fitness coaches. A server decided whether to show the advertised interface or a hidden web page generating explicit images and videos from uploaded photos; one trigger was arriving through a related Meta ad. The network was promoted through thousands of Facebook and Instagram ads and accepted payments through Apple and Google Pay. Researchers at AI Forensics estimate that the operator released apps for about a year, at times dozens per week. Based on domain records and service behavior, Wired links them to Chinese-language developers and Hong Kong-registered KeLing Technology Limited. Researchers had previously found SI-generated child sexual abuse material among the ads.

Sources 1

wired.com

Korean compute-infrastructure builders are expanding abroad, while Southeast Asia gains another site for large SI workloads.

LG CNS builds 30 MW SI data center in Jakarta with room to grow to 400 MW

Korean compute-infrastructure builders are expanding abroad, while Southeast Asia gains another site for large SI workloads.

The hyperscale Jakarta facility was built for KMG, a joint venture between Indonesia's Sinar Mas Group and Korea Investment Real Asset Management. LG CNS calls it the first SI data center built by a South Korean company for an overseas client. The site supports high-performance computing on next-generation GPUs and can expand from 30 to 400 MW. LG CNS plans to pursue similar projects in Malaysia, Thailand, Vietnam and Japan, as well as subscription GPU services and liquid-cooling infrastructure.

Sources 1

koreaherald.com

Grok's infrastructure continues to expand across multiple sites, and the announced dates mark Colossus 2's shift from construction to a phased launch.

Musk gives launch dates for two new Colossus 2 sites

Grok's infrastructure continues to expand across multiple sites, and the announced dates mark Colossus 2's shift from construction to a phased launch.

According to Elon Musk, the Gun Range site of the Colossus 2 computing complex in Memphis should come online by the end of November and Spam in December. The first is on a former local police shooting range, while the second occupies a building formerly used by Hormel. Colossus 2 expands computing capacity for Grok. Musk previously said SpaceX had raised $40 billion in debt to buy SI chips and that its SI business overtook Starlink in revenue last quarter.

Sources 1

thenextweb.com

Robotics investors are backing not only machines but also recordings of human actions, which models need to learn ordinary real-world work.

GenRobot AI discloses $100 million round to collect robot training data

Robotics investors are backing not only machines but also recordings of human actions, which models need to learn ordinary real-world work.

According to TechNode, the Series A+ round closed in the third quarter and was disclosed on November 2. Strategic investors included SAIC Motor, Baidu, Eastworlds, Shoucheng Holdings and Galaxy General Robotics, with GL Ventures, CITIC Goldstone, CCB International and Huatai Zijin providing the financial tranche. The Beijing company makes equipment that records video and people's hand and body movements during household, manufacturing and logistics tasks. Its Matrix platform synchronizes, cleans and labels the recordings for robot training and evaluation. GenRobot says it holds more than one million hours of data and works with over 30 SI companies; its open 13,000-hour dataset receives more than one million downloads per month.

Sources 2

aiweekly.coaichatdaily.com

A small agent can learn to recover from errors directly inside real apps and transfer that skill to another mobile operating system without a huge cloud model.

Apple trains compact SI agent to operate apps on live iPhones

A small agent can learn to recover from errors directly inside real apps and transfer that skill to another mobile operating system without a huge cloud model.

Apple researchers split the agent into a planning module and a module grounding actions in interface elements. Reinforcement learning took place in live iOS apps across hundreds of concurrent virtual devices rather than on recorded trajectories. A judge model assessed both intermediate steps and final outcomes, while GiGPPO used visually similar states from parallel attempts to stabilize training. The 8-billion-parameter agent outperformed variants trained on prepared examples and those rewarded only for the final result. When transferred to Android, its generalization was almost seven times the SFT baseline; on real iOS apps it tied Gemini-3.1-Pro and outperformed Gemini-3.5-Flash. This is a research result on UINavBench, iOSWorld and AndroidWorld, not a consumer feature announcement.

Sources 1

machinelearning.apple.com

The value of a large SI contract cannot be measured by the number of formally connected organizations: data submission, genuine daily use and proven patient benefit must be distinguished.

29 NHS hospital trusts dispute active use of Palantir platform

The value of a large SI contract cannot be measured by the number of formally connected organizations: data submission, genuine daily use and proven patient benefit must be distinguished.

The Guardian and Medact sent more than 100 freedom-of-information requests. All 29 trusts that NHS England publicly listed as benefiting from the £330 million Federated Data Platform replied that the system was not live or that they were not using its applications. The Office for Statistics Regulation has taken the NHS claims into its casework process. NHS England maintains that nearly 140 trusts reported benefits from the tools and explains that the list includes back-office analytics and data submission, not only patient care. Infrastructure for data submission can count as live without operational use of the applications. Palantir cites 117,000 additional operations, a 14.3% reduction in discharge delays for long-stay patients and a 5.6% improvement in 28-day cancer diagnosis. NHS previously acknowledged that it could not prove a causal link between the platform and additional operations.

Sources 1

theguardian.com

One trial demonstrated both the practical value of recognition to police and the dispute over mass checks of passersby, where even a rare error becomes a real person's stop.

Western Australia police scan 1.38 million faces and arrest 122 people in recognition trial

One trial demonstrated both the practical value of recognition to police and the dispute over mass checks of passersby, where even a rare error becomes a real person's stop.

Over five months, the real-time system was used in 100 deployments across 43 locations. Police reported 122 arrests, 146 interactions with registered sex offenders and nine false alerts; according to police, no one was arrested because of a false match. Images of people not on the watchlist were pixelated and deleted immediately. Commissioner Col Blanch called the false-alert rate very low while acknowledging the technology is imperfect. Civil Liberties Australia objects to calculating the error rate against all 1.38 million scans and proposes assessing court outcomes for those arrested. The Human Rights Law Centre points to the consequences of every wrongful stop and questions whether mass biometric surveillance is proportionate. Camera vans have been withdrawn while results and safeguards undergo an internal review.

Sources 1

theguardian.com

Open weights let one base model be adapted to different rules and cultures, but such tuning changes not only safety but also the norms users receive in answers.

Western startups adapt open Chinese models for local companies

Open weights let one base model be adapted to different rules and cultures, but such tuning changes not only safety but also the norms users receive in answers.

Hirundo edits open-model weights to change answers on politically sensitive topics and reduce cybersecurity and dangerous-knowledge risks. Lazarus AI removed mandatory censorship from Qwen variants: it identified topics using leaked guidelines and generated retraining answers with Gemini Flash. Former chief scientist Eric Hartford said ReAligned had been deployed at roughly a dozen US companies. Hirundo describes the task as adaptation to Western customers' requirements, not creating an anti-China model. Researcher Oriane Peter argues that universally safe or "Western" alignment is impossible: alignment depends on culture and application, while a single center can produce cognitive homogeneity. Graham Webster of DigiChina notes that political alignment often has little practical significance for ordinary tasks such as building a coffee shop website.

Sources 1

restofworld.org

Filmmakers gain a workflow in which actors perform on set while the visual world around them can be rebuilt later without additional shooting days.

iQIYI opens Shanghai studio combining live filming with SI post-production

Filmmakers gain a workflow in which actors perform on set while the visual world around them can be rebuilt later without additional shooting days.

The roughly 110-square-meter studio supports multi-camera filming, synchronized high-resolution video and audio, digital continuity logging and real-time footage review. Recorded scenes automatically flow into the NadouPro platform, which offers nearly 30 models, including Seedance 2.5 and Kling 4.0, plus hundreds of production agents. The first project is the internet feature Cop in Trouble, whose filming began on October 28. Live performances provide the source material and reference for generating appearance, composition, lighting and space; sets, costumes and details can change after filming while preserving the acting. iQIYI calls the site China's first SI film studio and aims it at small teams as well.

Sources 2

sg.finance.yahoo.comaichatdaily.com

Gemini will appear in an everyday Samsung wearable that translates, provides directions and collects notes without requiring the user to take out a phone.

Samsung to launch Gemini smart glasses in mid-December, initially in Korea and US

Gemini will appear in an everyday Samsung wearable that translates, provides directions and collects notes without requiring the user to take out a phone.

The glasses run Android XR and use Gemini for real-time translation, navigation and environmental recognition through a camera and microphone. Documents and meetings can be saved and organized in the notes app on a Galaxy smartphone. They use the Snapdragon AR1 Gen 1 chip and have a claimed nine-hour battery life. The limited launch in two countries will let Samsung assess demand and feedback before expanding geographically. The designs were created with Gentle Monster and Warby Parker; Samsung showed four variants at Google I/O and Galaxy Unpacked.

Sources 1

3dnews.ru

ChatGPT advertising is joining the conventional automated market, giving major buyers an easier route through existing systems and OpenAI another way to fill ad inventory.

OpenAI connects Magnite and PubMatic to remnant ad sales in ChatGPT

ChatGPT advertising is joining the conventional automated market, giving major buyers an easier route through existing systems and OpenAI another way to fill ad inventory.

According to The Information, the agreements let advertisers set budgets, place bids and upload creative for ChatGPT through the familiar Magnite and PubMatic platforms, while OpenAI retains control over ad placement and delivery. The arrangements cover inventory not sold directly; current volume through the new partners is small, and participants expect it to grow as integration deepens. OpenAI is expanding its advertising infrastructure after agreements with Criteo, LiveRamp, StackAdapt and Amazon Ads. A company spokesperson told The Information that the new connections make it easier to buy and manage ads through tools advertisers already use.

Sources 2

sg.finance.yahoo.comtheinformation.com

Protecting compute infrastructure is shifting from a single backup server to distributed and underground sites, although such physical resilience substantially raises construction costs.

Rostelecom discusses underground data centers protected from drones

Protecting compute infrastructure is shifting from a single backup server to distributed and underground sites, although such physical resilience substantially raises construction costs.

According to Vedomosti, the operator is considering underground sites and holding talks with Nornickel. The proposed protection has three layers: a backup center storing data copies, a distributed architecture across geographically separate locations, and underground facilities for the most critical workloads. Denis Patrikeev, lead consultant at iKS-Consulting, estimates construction will cost 30-40% more than an above-ground center of comparable capacity and reliability because of excavation, waterproofing and ventilation. He estimates a service life of 15-20 years. This is a plan under discussion, not an operating facility.

Sources 1

3dnews.ru

The two countries are expanding cooperation from cyber defense to a shared military SI cycle, from testing through deployment in their armed forces.

UK and Germany to jointly test and deploy military SI

The two countries are expanding cooperation from cyber defense to a shared military SI cycle, from testing through deployment in their armed forces.

British Defence Secretary Wes Streeting announced an expansion of the joint Cyber Lighthouse project to superintelligence. The two countries plan joint testing, development and eventual deployment of the technology in their armed forces. Government and industry representatives will meet in spring 2027 to determine how the work will proceed. Streeting also noted that German companies had invested £800 million in the UK defense industry since the 2024 defense pact, including Helsing, which opened a Plymouth factory producing SI-enabled underwater vehicles for tracking submarines.

Sources 1

euronews.com

Source verification is becoming a mandatory part of legal work, as generated citations regularly reach court and consume participants' time.

About 80% of South Korean judges encounter SI-fabricated legal precedents

Source verification is becoming a mandatory part of legal work, as generated citations regularly reach court and consume participants' time.

The Supreme Court surveyed 360 judges from October 6 to 16. Among those handling civil, family and administrative cases, 196 judges, or 80.7%, had seen citations of nonexistent precedents, while 159, or 65.4%, had seen distorted precedents. Another 16.9% had encountered citations of nonexistent laws. About 69.5% of judges reported false precedents cited by litigants themselves and 57.6% by lawyers. Nearly 92% supported institutional measures: source verification, grounds for fines and disciplinary action for repeat violations. The National Court Administration links such citations to delays in proceedings.

Sources 1

en.yna.co.kr

Users will no longer need to move a project manually between conversation and execution, although the simpler interface will share one usage allowance for chat and coding.

OpenAI to merge Chat and Work tabs in desktop ChatGPT

Users will no longer need to move a project manually between conversation and execution, although the simpler interface will share one usage allowance for chat and coding.

In an email to users, OpenAI said it would first enable the combined interface for a small group and gather feedback before a wider rollout. A single window will let users move from discussing an idea to execution: attach files and plugins, request a website or presentation, and guide ChatGPT's work on the web. Codex will remain a separate part of the desktop app. ChatGPT and Codex will draw from a shared token pool; OpenAI expects most users to see little change, but limits may run out more frequently during intensive work.

Sources 1

techcrunch.com

The weapons maker is moving SI from auxiliary systems and data analysis toward possible functions in the firearm itself, but this remains an announced direction rather than a finished product.

Kalashnikov considers adding SI to assault rifles

The weapons maker is moving SI from auxiliary systems and data analysis toward possible functions in the firearm itself, but this remains an announced direction rather than a finished product.

CEO Alan Lushnikov told journalists that SI is already used in the company's developments to control automated systems and process data, and that its scope will expand. The technology is not yet used directly in assault rifles; the company is considering that possibility for the future.

Sources 1

ixbt.com

Defenders are facing not a single script but parallel agent actions that correct mistakes and create tools during a breach.

Positive Technologies links three-hour breach of Russian company to SI agents

Defenders are facing not a single script but parallel agent actions that correct mistakes and create tools during a breach.

In Vzglyad's account, Positive Technologies said attackers accessed more than 10 Linux and Windows nodes, while the path from an external service to the leak of confidential information took about three hours. PT ESC IR head Denis Goydenko called it the department's first investigation of a successful attack believed to have been carried out entirely by SI agents; the company bases that conclusion on the speed and pattern of activity. According to the company, initial access came through a configuration error in an external service. The agents tried multiple lateral-movement paths in parallel and corrected their own mistakes, while malicious tools were created during the attack. This is the investigating company's assessment of the likely mechanism, not independently confirmed attribution of the perpetrator.

Sources 2

vz.rukommersant.ru
Previous day