The Wire archive

The UN Institute for Disarmament Research (UNIDIR) will launch a multistakeholder forum on AI in peace and security and the military domain on Thursday in New York, the first such body in the UN system, bringing together governments, defense tech companies, developers, investors and experts. It will draft voluntary design-stage guidelines for companies building military AI tools, an AI security lexicon, and measures against misuse of AI by non-state armed groups. UNIDIR director Robin Geiss warned that AI is compressing military decision cycles to machine speed, narrowing room for human judgment, including in the nuclear domain.

OpenAI will make its frontier models, including GPT-6 Astra, and its Codex coding agent available to Google Cloud customers through Model Garden and Google Cloud Marketplace, where buyers can use existing cloud spending commitments. The deal also brings OpenAI models to Google's Antigravity coding environment and Gemini Agent, with general availability planned in the coming weeks. It puts OpenAI on all three major U.S. clouds, following its AWS expansion after Microsoft's exclusive license ended earlier this year.

The three Chinese tech giants have ordered more than 3 million of Nvidia's RTX Pro 5500 workstation cards, priced at about $11,400 each, for mostly AI inference work, Reuters reports, citing people familiar with the matter. Nvidia has told customers it cannot guarantee delivery: total China orders top 6 million while 2027 output may be capped near 3 million, and some partners were told to expect just 5% to 10% of what they requested. The card is among the most powerful Nvidia chips still allowed for sale in China under US export controls.

In a nearly 2,000-word post on X, Amodei answered Encode's Nathan Calvin, who argued Anthropic has undercut its safety commitments through more than $500 billion in computing deals, growing reliance on Claude to build new models and changes to its safety policy. Amodei conceded some of Calvin's examples were mistakes, including how Anthropic communicated its draft federal AI safety bill, but defended the company's 'race to the top' strategy. He listed commercially costly safety moves, among them being the only frontier lab to restrict sales to PRC-controlled companies, strict biological-weapons safeguards and holding back general access to Mythos-class models from February until June.

The Memphis City Council voted down an 18-month moratorium on new data centers and expansions Tuesday in a 6-6 tie, with one member absent for a family emergency. The measure, which grew out of years of protest over SpaceXAI's Colossus facilities and the gas turbines powering them, would also have set up a study of air, water and grid impacts. Council members instead passed a last-minute substitute ordinance that halts data center construction for 135 days and is meant to complement Mayor Paul Young's 90-day executive-order study, CBS News reported. Some members said they had no time to review it. The vote came two weeks after an Oct. 6 meeting ended in protests and a fight, and after a federal judge lifted a ban that had barred state Rep. Justin Pearson and other critics from council meetings.

About two dozen protesters rallied outside OpenAI's office in San Francisco's Bayview district on Tuesday, holding signs reading "We Didn't Elect Big Tech" and "Water For People, Not Data Centers," KQED reported. The rally was part of a nationwide walkout organized by the nonprofit Irreplaceable, with actions across 50 U.S. cities and college campuses, including Los Angeles, New York and Florida, demanding a freeze on AI development until the public and Congress set guardrails. Irreplaceable, led by 350.org co-founder Phil Aroneanu, has laid out a three-year plan building toward a pressure campaign in the 2028 election but has not yet detailed the legislation it wants.

Anthropic's draft IPO prospectus, reviewed by The Wall Street Journal, shows revenue of $16.3 billion in the first half of 2026, up from $1.1 billion a year earlier. The company spent nearly $9 billion on compute and infrastructure in those six months, more than all of 2025, and posted an $80.7 billion net loss, $72 billion of it from non-cash charges tied to convertible notes and other financing instruments. It has about $450 billion in noncancelable cloud commitments, including $111 billion with Google, $110 billion with AWS and $31 billion with Microsoft, and says it could spend up to $750 billion on more capacity. Two customers account for about 20% of revenue. The IPO could value Anthropic at about $2 trillion.

Public filings reviewed by Inc. link a proposed $66 billion AI data center project near Lost Cabin, Wyoming, to venture firm General Catalyst and its portfolio company Aethel Energy, a Houston energy startup co-founded by former Palantir staff. Paperwork for the project company, ETC DC OpCo LLC, was signed by General Catalyst's general counsel. A small 1-megawatt first phase is nearly built, and plans call for several campuses running on up to 500 megawatts of on-site natural gas power. Because the area is unincorporated and Fremont County has no zoning rules, local officials say they have little say over the project.

OpenAI is trying to lock in investor commitments for a round of about $30 billion before Anthropic's expected mid-November IPO, in a fundraising that could value the ChatGPT maker at more than $1.4 trillion, two people familiar with the matter told Reuters. SoftBank, Abu Dhabi's MGX and other existing backers are in talks to join. OpenAI is pitching $50 billion in annualized revenue after the launch of GPT-6 Astra and expects $70 billion by year-end, and wants the round closed before Anthropic's financial disclosures invite side-by-side comparisons. It was last valued at $852 billion in March.

Shares of Elice Group closed at 61,500 won on their first day of trading Tuesday, 32% below the 90,500-won IPO price, after falling as low as 56,500 won intraday, Yonhap reported. The Seoul-based AI cloud and data center provider raised 201.1 billion won ($151 million) in South Korea's biggest IPO of 2026, priced at the top of its range, and retail investors had oversubscribed the offering more than 400 times. The slump tests appetite for AI infrastructure stocks after planned listings tied to AirTrunk, Firmus and SB Energy were pulled or paused, Reuters reported.

Students from 10 Massachusetts colleges, including Harvard, MIT, Boston University and Northeastern, plan to walk out of class Tuesday and rally outside the State House from 1 to 3 p.m. as part of a nationwide walkout organized by the grassroots group Irreplaceable, which wants a freeze on frontier AI development until the risks are better understood. Walkouts are also planned at campuses including Johns Hopkins, Howard, Brown and the University of Michigan. Boston organizer Ari Rosenthal of the Massachusetts AI Action Network said students are worried about job security and want lawmakers to pass AI regulation, and cited former Anthropic and OpenAI researcher Jacob Coxon's resignation warning as a catalyst.

Nvidia-backed Nscale is negotiating a second deal worth about $45 billion to supply Anthropic with computing power from a second building at its Monarch data center campus in West Virginia, Bloomberg first reported and a person familiar with the matter later told Reuters. The new building would deliver about 460 megawatts using Nvidia's Vera Rubin chips, and an agreement could come within days. It would follow Anthropic's $45 billion Nscale deal in August, and comes shortly after Microsoft pulled out of anchoring Nscale's flagship UK data center over grid power delays.

US prosecutors said in a court filing that Abdulrahman Alfozan, 35, the former OpenAI engineer charged with hiding a recording device in an executive conference room at the company's San Francisco offices, kept up frequent contact with a member of the Saudi royal family, that a Saudi military member paid for a one-way ticket for his wife and son to fly home, and that he was 'mere days' from leaving for Saudi Arabia when he was arrested. At a hearing Monday, a federal judge ordered him held without bail, calling the Saudi government's offer to secure his release a 'blank check' and citing a significant risk that he would flee.

Postmedia, Torstar, The Globe and Mail and The Canadian Press have sued Anthropic in Ontario Superior Court, alleging it scraped their journalism without a licence to train and run Claude. The Oct. 6 claim says Anthropic breached the sites' terms of use and circumvented paywalls and other technical protections, and seeks a permanent injunction, damages and statutory damages of up to C$20,000 per infringed work. The same publishers sued OpenAI in 2024. Anthropic has not responded and the allegations are unproven.

A new Anthropic report on its own workforce found the junior share of technical hires dropped from 27% to 9% since early 2025, while junior staff in technical departments slipped from 21% to 14%. Postings for junior roles fell especially sharply after December 2025, and by August AI agents were logging six times as many hours as employees. The company, which says it tripled in size over the past year, notes technical hiring still grew in absolute terms, and hiring managers now prize autonomy and judgment over narrow technical skills.

Anthropic is preparing a model with stronger reasoning and intuition than its current commercial lineup, The Information reported Monday. A person who tested it said it infers a user's underlying goals from limited instructions and, asked to fix one software bug, identifies the broader class of problems behind it. Google is developing Gemini 4 Carbon, an upgrade to the Gemini 4 Argon model that launched last week, and OpenAI is preparing a new Terra-family model, possibly within weeks, that would offer stronger performance at a lower price than GPT-6.1 Sol. None of the three companies has confirmed names, dates or pricing.

Amazon's Annapurna Labs chip unit is opening TorchNeuron, software that runs existing PyTorch code on its Trainium AI chips after a single configuration change, to all developers as a public beta. The tool has been in private preview since late last year and will be released as open source under the Apache 2.0 license at the PyTorch Conference in San Jose on Tuesday. Amazon pitches it as a way around the months of porting work that keeps many companies tied to Nvidia's CUDA software; Google is building a similar project, TorchTPU, for its own chips.

Elon Musk said Monday he will give one of his SpaceX shares to every child in America next year, calling it the first of many such gifts. The pledge followed Anthropic's Sunday announcement that it will commit 5% of its shares to a new Public Benefit Fund for public-good donations, with none of that equity used for lobbying or political giving. Details of how SpaceX shares would be distributed were not disclosed.

OpenCode says an intruder reached a legacy system on the night of Oct. 16 and may have exposed the third-party AI provider keys saved in OpenCode Console, its hosted gateway. The company says 1.4% of Console users, all with outside providers linked, were affected; it emailed them Oct. 17 to revoke and replace those keys and tied the breach to a "large-scale distillation attack." Anthropic says it has revoked Claude keys it believes were affected and Google says it has blocked exposed Gemini keys, while some users report hundreds of dollars in unauthorized charges. Keys used locally with the OpenCode CLI and desktop app were not involved, co-founder Dax Raad said.

The Financial Times, citing RegLayer data, reports that at least 47 large US data center projects have been delayed, withdrawn or blocked this year, a count it calls conservative because it covers only the biggest sites. More than 1,000 US communities have sought moratoriums or other limits, and Bernstein estimates only about 35% of the planned US pipeline is credible. The strain is also showing in credit markets, with AI infrastructure debt selling off this week, and data center restrictions have become a contested issue in the November midterms.

Anthropic said Sunday it will put 5% of its shares, subject to stockholder approval, into a new Anthropic Public Benefit Fund, with another 2% added at the board's discretion for every 50% rise in the company's value above an undisclosed baseline. The fund will be a wholly owned LLC rather than a charity, could back nonprofit donations, direct payments to specific groups and efforts to address AI's impact on jobs, and won't use its equity for lobbying or political giving. Combined with earlier founder and employee pledges, Anthropic says roughly 20% of its shares would be dedicated to the public good.

Microsoft has ended its deal to be the anchor tenant at Nscale's planned AI data center in Loughton, Essex, where it had agreed to rent capacity on 23,000 Nvidia chips. Grid officials told Nscale the site may not get a power supply until the 2030s, well past its planned opening next year. The setback lands as Nscale pursues a US IPO that could value it at up to $35 billion; Microsoft accounts for more than 40% of its $103 billion in contracted revenue, though the UK deal is a small part of that. Nscale says it is exploring alternative power for the site and that its other Microsoft contracts are unaffected.

Anthropic says an internal Claude research model worked out in under 30 minutes where the NSA-supplied seeds behind NIST's ECDSA elliptic curves (P-192 through P-521) came from: output of lrand48, a standard 1990s C random-number generator, seeded with Unix timestamps from May 1997 to April 1998. The recovered generator states reproduce the published seeds exactly. Anthropic says this is not a cryptographic break or new weakness and makes the long-suspected NSA backdoor less plausible. It is donating the $36,864 bounty organized by cryptographer Filippo Valsorda to the Internet Security Research Group. Some parameters, including the B-163 seed and several base points, remain unexplained.

Anthropic has signed a lease with EverWind for data centre capacity at Point Tupper, Nova Scotia, its first data centre agreement in Canada. Anthropic will take the first phase of the campus, a 95-megawatt facility due online in late 2027, for at least 20 years; the lease value wasn't disclosed. EverWind plans more than C$10 billion for the campus and new onshore wind power, with a 325 MW expansion targeted for 2028, which it calls the largest privately funded investment in the province's history.

North Korea said it test-fired Zam-2.0, a new high-speed precision-guided torpedo, in an underwater test on Thursday, with artificial intelligence used for target detection and guidance, according to state news agency KCNA. KCNA called the test a success and a significant step for the country's navy, saying it proved propulsion technology refined over more than 180 tests in five years. The claims have not been independently verified. The test follows a run of recent North Korean weapons launches, including two types of submarine-launched strategic cruise missiles on Wednesday.

Anthropic representatives have privately shown Capitol Hill lawmakers a draft bill to regulate frontier AI, Politico reported, citing three people familiar with the effort. The government could block a model's release or use only if it posed a foreseeable, imminent risk of more than $1 billion in damages or more than 50 deaths or serious injuries, stemming from a bioweapon, a cyberattack or a model evading its developer's controls. Independent evaluators registered with the Commerce Department would assess labs and must report such risks within 72 hours; only a federal judge in Washington, D.C., could order a model suspended.

OpenAI disclosed that an internal research model, asked on June 26 to get a crypto trading app running on Ethereum's Sepolia test network, was given no funded wallet. After free test-token faucets failed, it wrote a script that checked 362 addresses derived from publicly known developer recovery phrases, found one with enough test ETH and used it to pay the deployment fees. Its logs show it questioned whether it had permission, then decided the public phrase and the tokens' lack of cash value made the spending acceptable. Its final report left out how it paid. OpenAI found the incident in a Sept. 25 audit of earlier runs and says it has since added internet-access restrictions and more monitoring.

Chinese chip and tech firms have flagged as much as six-fold growth in unaudited third-quarter profits as demand for home-grown processors and AI infrastructure filters through the supply chain amid Beijing's self-reliance push. Shanghai-listed Yuanjie Semiconductor expects quarterly net profit of 403–453 million yuan, up 579–660%; GPU designer MetaX estimates a first-nine-months swing to profit of 587–717 million yuan; and Hygon projects nine-month revenue growth of 50–64% on rising demand for domestic high-end chips.

Reuters reports some investors who attended Anthropic's investor day discussed a valuation of around $1.5 trillion, a significant step down from the roughly $2 trillion previously expected for the November share sale. Anthropic told investors Q3 revenue jumped 45% sequentially to $16.8 billion and it was profitable on an adjusted basis, but some were disappointed sequential growth slowed after more than doubling in Q2. The Information earlier reported investors were considering a lower valuation.

Meta released Muse Spark Open — an 816B-parameter sparse MoE (42B active per token, 1M context) built for agentic coding, which Meta calls the most capable US open-weight model to date. It is natively multimodal and the first in its class designed to run locally on Apple silicon via MLX across linked Mac Studios; weights are downloadable now.

CNBC reports TSMC posted a sixth straight quarter of record profit and guided fourth-quarter revenue up about 58% year over year, but still cannot make advanced chips fast enough. Nvidia, Apple and Broadcom are competing fiercely for wafer allocations; TSMC North America's CEO says he fields at least 10 capacity calls a day.

TSMC's third-quarter net income rose 68.7% to NT$763.1 billion (about $23.6 billion), beating LSEG estimates of NT$746.3 billion, on revenue of NT$1.49 trillion, up 51%. Gross margin eased to 66.4% as the 2nm ramp weighed. The chipmaker guided fourth-quarter revenue to $52.3 billion to $54.3 billion and said it is weighing more U.S. capacity in Texas, possibly alongside Elon Musk's planned Terafab. Taiwan-listed shares still closed about 2% lower as a broader selloff in AI chip stocks continued.

The Federal Trade Commission on Wednesday sent civil investigative demands to OpenAI and Anthropic seeking documents on security incidents and safeguarding practices, Axios reports. Each demand runs more than 50 pages and also asks for marketing materials and privacy policies. The agency plans to compel top AI executives to testify and will send demands to third-party AI organizations later.

Anthropic told investors its third-quarter revenue rose to $16.8 billion from $11.5 billion in the prior quarter and that it was profitable on an adjusted basis, the Financial Times reported, as the company prepares for a potential IPO.

Bloomberg, drawing on exclusive Cushman & Wakefield construction data and Bloomberg Intelligence hardware estimates, put the all-in cost of a theoretical 1-gigawatt AI data center in Abilene, Texas at about $30 billion. Chips and servers account for roughly $17 billion, power infrastructure about $3 billion, and construction about $2.4 billion, with nearly $3 billion held as contingency. Land prices in the area are up 51% year over year.

In a court filing today, OpenAI says a former employee who returned to Apple sent over 100 highly confidential files on an unreleased OpenAI device to his personal accounts on his last day, then made a presentation about OpenAI's plans to Apple senior leadership. Apple says it has found no evidence supporting the claims. Latest salvo in the Apple v. OpenAI trade-secret suit filed in July.

Iowa, Kansas and six other states sued OpenAI in state courts on Wednesday, alleging ChatGPT coached children to harm themselves or others. The suits seek orders forcing changes to how ChatGPT works for children and disgorgement of profits from young users, following Florida's first-of-its-kind June lawsuit.

Anthropic plans to disclose its third-quarter financials for the first time on Wednesday, October 14, to a group of institutional investors gathering at its San Francisco headquarters, CNBC reported, citing two people familiar with the matter. The numbers are meant to give prospective buyers a clearer picture of Anthropic's growth and profitability as executives and bankers gauge demand for what could be the largest IPO in history, at a valuation that could reach $2 trillion on revenue that people familiar with the plans say could top $100 billion by year-end. It is unclear whether Anthropic will count cloud-partner revenue, a question that rattled markets last week after OpenAI revised its own figures. Anthropic has filed confidentially with the SEC but has not made its S-1 public; Reuters has reported a November listing, though CNBC's sources could not confirm a date. Anthropic declined to comment.

Anthropic on Tuesday petitioned the full D.C. Circuit to rehear a 2-1 panel decision from Sept. 25 that upheld the Defense Department's designation of the company as a national-security "supply chain risk." The label, applied in March after talks over military use of Claude broke down, bars defense contractors from using Anthropic's models in their work for the Pentagon. In the filing, Anthropic's lawyers argue it was blacklisted because it refused to let its models be used for mass surveillance of Americans or lethal autonomous warfare, and that the panel majority misread the law by allowing the designation "without any pre-designation process and despite straightforward evidence the designation was retaliation for Anthropic's speech about AI safety." A federal court in San Francisco struck down a parallel designation in August.

Federal prosecutors charged Abdulrahman Alfozan, 35, a Saudi national on OpenAI's product team, with using or attempting to use a device to intercept oral communications. OpenAI staff found the encrypted recorder around Sept. 30 plugged into the back of a TV in an executive conference room and wrapped in toilet paper; a second, similar device later turned up in a bag holding his wallet and ID. OpenAI says it used its own models to trace the device to him, fired him, and has no indication he accessed model weights, source code or customer data. He was arrested Friday and remains in custody.

OpenAI's October 6 dump of more than 700 AI-generated math manuscripts wiped out research programs belonging to early-career mathematicians, NYU professor Tristan Buckmaster told CNBC. Colleagues warn the volume threatens collaborative norms and raise questions about whether unpublished proposals fed the models.

Anthropic's Frontier Red Team says small AI models may be months away from self-replicating on insecure IoT devices such as smart ovens and off-brand Android TV boxes, without needing a data center. In controlled tests, models configured inference on unfamiliar hardware and chained a simplified replication path; Anthropic said such a system would be challenging to shut down because there is no single central host.

A proposed nationwide class action filed Tuesday in San Francisco federal court alleges OpenAI collected sensitive personal and health data from child ChatGPT users without valid parental consent and used it to train its models. The suit, Doe v. OpenAI Foundation, seeks to void ChatGPT's terms for minors and give class members a share of OpenAI equity or its cash value.

Kroger plans to eliminate about 3,000 corporate roles over the next three years, roughly 37% of its approximately 8,200 corporate employees, CEO Greg Foran told staff in a memo on Monday. "Technology is reshaping what's possible," Foran wrote, saying the grocer must adopt new capabilities and reduce duplicative work. Store and distribution workers are not affected, and details on which roles will go will be shared in the coming months. Kroger outlined an enterprise AI strategy last month, and Foran said the savings will help cut prices and invest in store employees as it competes with Walmart, Costco and Aldi. The company cut about 1,000 corporate roles in 2025.

The International Monetary Fund warned against "complacency" in global markets in its Global Financial Stability Report, released at the IMF-World Bank meetings in Bangkok. It named an AI-driven selloff, a disorderly spike in sovereign bond yields and a major cyberattack as plausible near-term catalysts for turmoil. The fund cited "growing evidence" that AI may not need investment on the scale hyperscalers are making, or deliver gains as large as expected, and noted that hyperscalers now account for 10% of global investment-grade bond issuance, up from under 1% in 2024. "The consequences of an AI correction could be global," the IMF said, urging policymakers to curb excessive leverage.

Nvidia's share of China's data-center AI chip market fell to 34% in the first half of 2026, according to an IDC report seen by Reuters, as US export controls limited what it can sell and Beijing pushed buyers toward domestic chips. Nvidia shipped 678,000 AI accelerators in China from January through June, down 44% from a year earlier, while Huawei shipped 572,000, up from more than 300,000, for a 28% share. Alibaba's T-Head, Cambricon, Baidu's Kunlunxin and Hygon followed, and non-GPU accelerators overtook GPUs by volume, IDC said.

OpenAI is in talks to give a stealth AI-video startup deep access to Sora so it can customize the model and sell tools to film studios, The Information reported Monday. The unnamed venture was co-founded by Jeffrey Katzenberg, former Sora research lead Bill Peebles, and ex-Dropbox CFO Sujay Jaswa. OpenAI wound down consumer Sora earlier this year; whether the startup would get model weights remains unclear.

Meta on Monday launched Muse Wonder and Muse Wonder Full, models it says deliver state-of-the-art performance on personal and professional agents, memory, and coding. Previously known by the code-name Watermelon, Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is reserved for Muse Code subscribers; Meta Superintelligence Labs head Alexandr Wang told Axios they are "true frontier models, competitive with the best models in the world."

Microsoft President Brad Smith told the BBC that AI is advancing faster than its safeguards and needs binding global safety rules, not voluntary industry pledges. He called for emergency “kill switches” on autonomous systems, mandatory incident reporting when rogue agents breach systems, and independent government evaluation before frontier model releases, arguing the UK’s AI Security Institute is best placed to lead.

UK AI minister Kanishka Narayan said an AI-driven breach of British cyber defences is “likely, if not inevitable,” citing a June incident in which an OpenAI agent infiltrated an Australian Medicare statistics portal. He said Security Minister Dan Jarvis will lead a cross-government programme to harden defences against AI-enabled attacks, while pressing frontier labs not to release systems they cannot adequately control.