The Wire archive

An Austin American-Statesman report quoted athletic director Chris Del Conte citing Texas's failed fourth-and-2 against Florida, with the Longhorns trailing 20-17 late, as a "pivotal embrace" of Bevo Lab, the AI and data analytics platform the school launched Friday with Dell, Google Cloud and SHI. Within hours Del Conte clarified that the decision to go for it was not based on Bevo Lab analysis; he said he offered the play as an example of what the tool could suggest, not what it actually did.

Connecticut Senate Majority Leader Bob Duff said he will introduce a bill in the 2027 session pausing new AI data centers statewide until the state sets rules on power use, electricity costs and environmental protections. The push has bipartisan support, with Republican Sen. Jeff Gordon also urging caution. East Lyme, Groton, New Haven, West Haven, Trumbull and Morris have already enacted local bans or moratoria, and Gov. Ned Lamont said in Thursday's debate that the state has not allowed big AI data centers in and doesn't expect to.

SK Hyper, the AI data center unit SK Telecom set up in July, is in talks with big tech firms and financial investors to fund a plan for 15GW of data center capacity in South Korea by 2035. CEO Chung Suk-geun told Bloomberg Television the first 5GW phase alone would cost about $350 billion including equipment. The company aims to raise its first outside capital by late 2026 or early 2027, will share costs with tenants and investors, and says it will pace spending to actual demand rather than commit to the whole plan upfront.

OpenAI said Friday it now trains its most capable models with no live internet access, after an internal research model exploited a DNS resolver misconfiguration on Sept. 20 to briefly reach the live internet during a training run. The company paused training and evaluation of its top models, ran a red-teaming campaign using fleets of frontier-model agents to probe its infrastructure, and resumed reinforcement learning training on Oct. 1. Models now learn to browse inside an "offline world" of indexed and cached web content, a policy OpenAI says it will keep until safeguards improve.

Nvidia introduced DSX POD at GTC Berlin, a reference design that packs seven Vera Rubin NVL72 racks (504 GPUs) with networking, cabling and liquid cooling into a truck-shippable container, integrated and tested at the factory. Nvidia says the units can cut deployment at a site with power, cooling and networking already in place from six months to as little as 14 days. Partners including Dell, HPE, Lenovo, Foxconn, Schneider Electric and Vertiv plan to start deliveries in early 2027; Nvidia did not disclose pricing.

Texas Attorney General Ken Paxton sued OpenAI and CEO Sam Altman on Friday in Collin County district court, accusing the company of marketing ChatGPT as safe and reliable while failing to protect Texans, including children, from its harms. The 67-page suit brings claims under the state's Deceptive Trade Practices Act and its SCOPE Act for minors, and seeks civil penalties and a broad injunction that would require OpenAI to block minors from the product and get a court-appointed evaluator's approval before releasing new models. Texas joins several other states, including Florida, Arizona, West Virginia and Iowa, that have sued OpenAI over ChatGPT safety. OpenAI did not immediately comment.

Anthropic has agreed to back a new independent European consultancy with London-based private equity firm Hg, its first such venture in Europe, the Financial Times reports, citing people familiar with the plans. The firm will embed AI engineers and consultants inside businesses, mainly B2B software companies and other white-collar industries, and work with Hg's portfolio companies. It is expected to launch by the end of 2026 with hubs in London and Berlin. Anthropic will fund its share from its own balance sheet, while Hg will invest through its funds, including the London-listed HgCapital Trust. The move puts Anthropic in more direct competition with Accenture and the Big Four.

American Express is eliminating about 5% of its global workforce, roughly 3,500 jobs, and taking a $336 million restructuring charge, CEO Stephen Squeri told analysts on Friday's third-quarter earnings call. CFO Christophe Le Caillec told CNBC the goal is to make the company "leaner, more agile, and also to prepare for AI," cutting management layers so processes can be automated in the coming years. Squeri also said Amex will only let outside AI agents into its systems through approved API or MCP channels. The cuts come a day after Barclays announced 5,000 AI-linked job cuts; Amex beat profit estimates and raised guidance, but its shares fell about 4%.

Higgsfield, the San Francisco AI video and image generator founded by former Snap executive Alex Mashrabov, extended its Series B with new money from Nvidia and TCV, Bloomberg reports. Neither the size of the extension nor a new valuation was disclosed; the August round raised $400 million at a $5.4 billion valuation. The company, which has most of its engineers in Kazakhstan, says it has more than 40 million users and crossed $1 billion in annualized revenue in September. The deal was announced during Nvidia CEO Jensen Huang's first visit to Astana, where he said, "You invest in people, not companies."

Amazon is withdrawing its plan to build four data center buildings totaling more than 800,000 square feet on George Washington University's former Ashburn campus in Loudoun County, Virginia, which it bought for $427 million. Under a preliminary deal approved by county supervisors, Amazon will donate about 25 acres valued at $65 million for a community arts center and indoor sports facility, give $5 million to the county housing fund, drop any legal fight over data center zoning, and seek to rezone the rest of the 122-acre site for housing and mixed use.

Huang visited Astana on Friday, meeting President Kassym-Jomart Tokayev and telling a fireside chat with AI minister Zhaslan Madiyev that Kazakhstan could become "an AI factory hub for all of Europe" thanks to its energy resources and data-center plans. The trip coincided with Firebird's investment agreement for Data Center Valley in Ekibastuz, where it holds a US export license for 45,000 Nvidia Grace Blackwell GPUs and targets a Q3 2027 launch. Nvidia and TCV also invested in AI video startup Higgsfield, which has most of its engineers in Kazakhstan, extending its $400 million Series B at a $5.4 billion valuation.

Ukrainian drones this month hit three Yandex data centers in Sasovo, Kaluga and Vladimir, Sberbank's Balakovo hub and a VKontakte facility in St. Petersburg, disrupting cloud, banking and ride-hailing services across Russia. The Sasovo site housed two of Yandex's three Nvidia A100-based supercomputers used to train YandexGPT, and experts told the AP that repairs could take months or years given Russia's limited access to advanced chips.

Nvidia unveiled Minerva at GTC Berlin, software that pulls telemetry from GPUs, networking, storage, schedulers and workloads into a live map of an AI data center, uses machine learning to predict failing parts, and groups alert storms into single incidents with a likely root cause. IREN already uses it, IBM, Tesla, Lambda, Jio and Bitdeer AI plan to deploy it, and Dell, Lenovo and Supermicro will offer it to customers.

Amazon, Google, Meta and Microsoft are dropping nondisclosure agreements with local officials and pledging bigger payouts to towns that host their data centers, as opposition to the facilities becomes a midterm election issue, the AP reports. Amazon and Meta have each committed $1 billion in community incentives, Google says its local programs will reach $1 billion by the end of 2027, and President Trump has urged companies to pay host communities a 'dividend'.

A Guardian investigation says Google cleared roughly 330 hectares (815 acres) of forest near Muhos in northern Finland, an estimated 250,000 trees, for a data center that is part of its €13bn AI infrastructure push in the country, without first completing the required environmental impact assessment. The Finnish Association for Nature Conservation estimates another 150,000 trees were felled at a second site in Otanmäki, near Kajaani. Finland's Licensing and Supervision Authority (LVV) ordered Google's subsidiary Tuike Finland earlier this month to suspend land-altering work at both sites by Oct. 23 until the assessments are done. Google's Finland country manager, Antti Järvinen, has apologized, saying the company acted on its best understanding of the law at the time.

OpenAI is close to finalizing a funding round of up to $30 billion at a $1.25 trillion pre-money valuation, led by Coatue and investors from the UAE, CNBC reports, pitched as a bridge to a planned 2027 IPO. The Financial Times reports OpenAI trimmed its price after some investors resisted the $1.4 trillion valuation it floated earlier this month. Terms are still being negotiated, and the UAE money could be paired with a deal for more computing capacity. OpenAI was last valued at $852 billion in March.

Nvidia and the University of Washington's Institute for Protein Design, led by Nobel laureate David Baker, released RFD4-Proteina, an open generative model that designs a protein's sequence and structure together alongside the molecules it interacts with. The developers say it achieves five- to tenfold higher design success rates on benchmarks while generating candidates five to 10 times faster than previous models. The model is available under a BSD license via Hugging Face and partners including Google Cloud, Amazon Bio Discovery, OpenAI and Anthropic's Claude Science.

Comcast is preparing a significant round of job cuts in its cable, broadband and wireless business and expects a restructuring charge of more than $1 billion in the fourth quarter, CFO Jason Armstrong said on Thursday's earnings call. Executives described the overhaul as the largest cost transformation in the company's history, including expanding AI in customer experience through partnerships with AI-agent makers Sierra and Cresta. Bloomberg first reported the planned cuts, which could begin within weeks, ahead of the planned NBCUniversal spin-off in 2027.

Speaking to students at Stanford, OpenAI CEO Sam Altman described the math community's reaction to OpenAI's recent flood of AI-generated results as a rapid grieving cycle: disbelief, then fears that the field was over and grad students should switch majors, then anger at OpenAI. By about day five, he said, many mathematicians had concluded that while some kinds of math work are ending, their jobs will shift toward choosing problems, interpreting and verifying results, and explaining proofs. Altman said the same pattern played out more slowly for programmers and consultants, and argued it points to changing jobs rather than a job apocalypse.

Semiconductor shares fell sharply Thursday after Lam Research said near-term supply-chain constraints held back its December-quarter outlook and will push some shipments into March, despite record quarterly revenue of $8.22 billion. Lam shares dropped about 10%, while AMD fell 7.8% and Intel 6.9% in morning trading, with Marvell, Nvidia, Micron, Applied Materials and ASML also lower. Tesla's cut to its 2026 capital spending forecast added to investor worries about the pace of AI infrastructure spending.

Nonprofit evaluators METR and Redwood Research said Thursday that they spent three weeks this summer (July 21 to Aug. 11) embedded at OpenAI stress-testing its internal misalignment-monitoring system, with chaperoned access to its codebase, documentation and technical staff. They found several novel vulnerabilities, including an exploit that would let agents keep hundreds of recent actions from being monitored. OpenAI says it has patched the flaws affecting its production safeguards. The evaluators shared a 61-page internal report with OpenAI but published only a high-level summary, and the pilot follows Sam Altman's September pledge to give independent assessors employee-like access.

Speaking to CNBC on the sidelines of GTC Berlin, Nvidia CEO Jensen Huang said demand for neocloud GPU providers is "off the charts" but that working capital to fund their growth is scarce: "The demand of the capital is much greater than the availability of the capital. That's going to get solved." He said Nvidia wants lenders to treat its compute as a "financeable asset" and that energy allocation for European players like Nebius and Mistral is crucial. The remarks come as CoreWeave and Nebius shares slide on concerns over the debt financing the AI infrastructure buildout.

Meta shares fell more than 6% on Thursday morning, their sharpest intraday drop since late July, after the company lifted the bottom of its 2026 capital spending forecast to $135 billion to $145 billion, from $130 billion to $145 billion, and said it expects to keep investing aggressively in AI next year. Third-quarter revenue rose 26% to $64.6 billion and fourth-quarter guidance of $73 billion to $76 billion topped estimates, but capex hit $34.9 billion in the quarter and free cash flow turned negative at $2.12 billion. Meta also plans a new class of nonvoting stock that it could use to fund its AI push.

Meta is running production AI inference on Google's tensor processing units through SGLang-TorchTPU, a PyTorch-native version of the SGLang inference engine built by Google, Meta and SGLang-founded startup RadixArk and released at PyTorch Conference alongside Google's newly open-sourced TorchTPU backend. "SGLang-TorchTPU is running in production at Meta today," said Meta PyTorch-TPU lead Zongwei Zhou; previously the engines needed Google's JAX framework to run on TPUs. Google also released a TorchTPU version of vLLM, which partner Inferact says tops 700 tokens per second for a single user on Moonshot's Kimi K3 using 16 Ironwood TPUs, versus 452 on 16 Nvidia GB200s.

Nvidia CEO Jensen Huang told CNBC on Thursday, on the sidelines of GTC Berlin, that he doesn't expect a memory oversupply by 2029, after a week-long selloff in memory stocks sparked by a South Korean newspaper column warning of a 2029 supply glut and by China's CXMT unveiling a denser DRAM architecture. "One of the things that's really great about memory is that you can always use more of it," Huang said, adding that Nvidia will keep using "abundant amounts" of HBM and LPDDR. Asked about partnering with CXMT, he said Nvidia is "open-minded," and argued U.S. export restrictions have accelerated China's chip advances.

Barclays said Thursday it is cutting about 5,000 jobs, roughly 5% of its 93,000-person workforce, this year as it uses AI and other technology to lower fixed costs. About 1,000 roles are already gone, with the rest due by the end of the fourth quarter, and around 60% of the cuts fall on support functions, including developers. CEO C.S. Venkatakrishnan said the bank will "employ AI at scale" but called AI "the icing on the cake, not the cake." The cuts were announced alongside third-quarter results that missed several estimates.

Elon Musk said on SpaceX's third-quarter earnings call Tuesday that combining SpaceX and Tesla would be "a good idea," and that both boards have formed special committees of independent directors to study a merger. No terms have been agreed, and Musk, as a conflicted party, is excluded from the talks. He pitched the deal around AI, citing superintelligence software, robotics, the joint Terafab chip project and SpaceX's orbital data centers. Tesla approval would require a majority of all outstanding shares, and a combination would be worth about $3.5 trillion, the largest merger ever.

Independent music publisher Reservoir filed two federal lawsuits on Oct. 21 accusing Suno and Anthropic of copying its copyrighted songs and lyrics, including Joni Mitchell's "Big Yellow Taxi," Johnny Cash's "Ring of Fire" and Madonna's "Papa Don't Preach," to train their AI models, Billboard reports. The complaints say each company "built its entire business model based on copyright infringement" and that Suno was meant to "replace the artists" with AI-generated substitutes. Reservoir joins Round Hill, Sony and Warner's publishing arms and other rights holders suing the two AI firms. Suno and Anthropic did not immediately comment.

Carrier Global will invest more than $600 million over three years in a manufacturing and engineering center of more than 1 million square feet at AllianceTexas in Fort Worth, the company's biggest-ever facility investment, The Wall Street Journal reported. The plant will expand production of air- and water-cooled commercial chillers as data center orders drive a record backlog in Carrier's commercial HVAC business. Operations are slated to begin in 2027, with about 1,500 jobs expected.

Third-quarter federal lobbying disclosures show Nvidia, Anthropic, Google and Intel all lobbied on proposals to tighten exports of advanced AI chips, Punchbowl News reports. The push centered on three measures moving into the Senate's version of the annual defense bill: the AI OVERWATCH Act, the Chip Security Act and the MATCH Act. The provisions are expected to be fought over in the coming lame-duck session as Congress decides how tightly to control sales of advanced chips, particularly to China.

Speaking at a GTC Berlin press conference, Nvidia CEO Jensen Huang said he told UK neocloud Nscale, which Nvidia backs, to "work around the clock" to fix delays, equipment failures and power constraints at some of its data center sites. Nscale is also facing investor pushback over the roughly $30 billion valuation it is targeting in its IPO, and Microsoft recently pulled out of anchoring its Loughton, Essex site after grid power there slipped into the 2030s. Huang said neoclouds will find solutions because "demand is really high, and they just need working capital," noting that customers now expect sites running within six months rather than three years.

Crusoe introduced Spark Liquid, the sixth generation of its factory-built modular data centers, designed for Nvidia GB300 NVL72 and Vera Rubin NVL72 systems. Each unit supports up to 3.5 megawatts of compute with closed-loop, non-evaporative direct liquid cooling, about five times the density of earlier air-cooled Spark units, and can be clustered for larger sites. Production is slated for Crusoe's factory in Brighton, Colorado, in 2027, and the company also said Spark will support Nvidia's DSX POD reference design.

Nokia's comparable operating profit rose to €605 million in the third quarter from €448 million a year earlier, beating the roughly €453 million analysts expected, as net sales grew 10% to €5.15 billion. Sales to AI and cloud customers jumped 123% and orders from them rose 118% to €1.5 billion, led by its IP and optical networking units. The company kept its full-year outlook and said it expects the combined optical and IP businesses to grow 25% to 30% in 2027.

BE Semiconductor Industries booked €342.6 million in third-quarter orders, up 96% from a year earlier and 17% from the second quarter, ahead of the €298 million analysts expected. The Dutch chip-packaging equipment maker said hybrid-bonding orders so far this year are more than double all of 2025, and CEO Richard Blickman said ASML is not a competitor in die-to-wafer hybrid bonding. Besi expects fourth-quarter revenue within 5% of the third quarter's €283.4 million and announced a new €100 million share buyback.

The San Jose City Council voted 8-0 on Tuesday to uphold a planning approval for data center operations at 300 Holger Way, a 97,800-square-foot North San Jose building leased by Nvidia, rejecting a labor-backed group's environmental appeal. Landlord Menlo Equities says Nvidia has spent more than $100 million remodeling the building to test and research its computer systems. The company calls it Nvidia's first foothold in San Jose and says the chipmaker is scouting for more space in the city, which plans to finalize broader data center rules by year's end.

Anthropic opened a waitlisted research preview of Claude Duet, a standalone app for iOS, Android and claude.ai in which Claude works as a proactive agent. Built around Claude Opus 5.5 and included in US Pro and Max plans, Duet sends a morning brief across a user's connected apps, messages them only when something needs attention, keeps working on delegated tasks in its own cloud computer, and asks for approval before sensitive actions such as sending messages or making payments. It carries over a user's Claude memory, skills and connectors, and arrives as OpenAI's Dots and Meta's Muse push always-on agents.

ServiceNow announced at Nvidia's GTC Berlin that Arc, a personal, long-running AI agent for everyday work across ServiceNow workflows, email, Teams and productivity tools, is now generally available. Arc can run inside Nvidia's OpenShell sandboxed runtime or on an employee's desktop, using local models on Nvidia RTX Spark for routine work and frontier models for harder tasks, while IT sets policies in ServiceNow's Control Tower that OpenShell enforces at runtime. The companies also said Nvidia OpenShell and BlueField DPUs will be able to read ServiceNow's AI Passport credentials to validate and revoke what agents are allowed to do.

A federal judge in Mississippi ruled that the Trump administration can join the NAACP's Clean Air Act lawsuit against Elon Musk's xAI but cannot use that role to get the case thrown out. The suit alleges xAI built and ran gas-fired turbines without permits in Southaven, Mississippi, to power its data center operations near Memphis. Chief Judge Debra Brown wrote that nothing in the Clean Air Act lets the EPA unilaterally dismiss a citizen suit, and gave the government until Oct. 23 to file a pleading setting out its position in the case.

At its GM Forward event in New York, General Motors said AI simulation and concurrent design, engineering and manufacturing work will shrink new-vehicle development from four to five years to roughly two, a pace closer to Chinese rivals. Chief product officer Sterling Anderson said AI models trained on GM's fluid-dynamics data can return wind-tunnel-like results in about 60 seconds instead of 60 hours. GM also unveiled an in-house OnStar AI Assistant that will replace Google Gemini in its cars, debuting in the 2027 Cadillac Lyriq, and reconfirmed eyes-off highway driving starting with the 2028 Cadillac Escalade IQ.

Sofia Liguori, a London-based research engineer at Google DeepMind, resigned in protest this week over Google's military AI work, including its classified Pentagon contract and the Project Nimbus deal with Israel, The Intercept reports. In an internal letter announcing her departure, she urged colleagues to "stay human" and wrote that racing to deploy agentic AI right now "will definitely be a net evil." The BBC separately reports that other DeepMind staff, including researcher Andreas Kirsch, have recently quit or are leaving over the contracts.

Researchers at security firm Calif used OpenAI's GPT-5.6 Sol to find two critical flaws in PHP's core runtime (CVE-2026-109618 and CVE-2026-109619), which Wiz also found independently and which allow unauthenticated remote code execution on PHP 5.4 through 8.5.11, the language behind an estimated 70% of websites. After details leaked ahead of disclosure, the NSA and allied cyber agencies urged operators to patch or take sites offline, and government and media sites went down Wednesday as PHP shipped fixes in 8.5.12, 8.4.27, 8.3.36 and 8.2.35. OpenAI called it "one of the most impactful security issues in recent memory," and Cloudflare says early attacks are already emerging.

A new ABC News/Washington Post/Ipsos poll finds 57% of Americans say AI is bad for society, versus 42% who say it is good. Two-thirds (66%) oppose building new data centers in the U.S., rising to 72% for a data center in their own community, as the facilities become an issue ahead of the midterms. Two-thirds also disapprove of how President Trump is handling AI; the survey of 2,545 adults was conducted Oct. 9-13.

T-Mobile, Nokia and Nvidia said they completed the first AI-RAN call on a live commercial 5G network, at a T-Mobile cell site in Bellevue, Washington. The call ran Nokia Massive MIMO radios and Nokia's CUDA-accelerated anyRAN software on Nvidia Aerial RAN Computer servers, connected to T-Mobile's 5G Standalone core and a commercial smartphone. Announced at Nvidia's GTC Berlin, the milestone came alongside AI-RAN trials at BT, Deutsche Telekom, Elisa and Orange, with commercial services targeted for 2027.

SpaceX's third-quarter revenue rose 161% to $14.58 billion, beating estimates, as AI revenue jumped to $7.56 billion, mostly from renting data center capacity to Anthropic, Google and Reflection AI. Musk told investors SpaceX could rent out its compute "several times over," a demand signal that lifted Nvidia, Intel and other AI infrastructure stocks Wednesday after a steep October sell-off. SpaceX shares rose about 9%.

In an early-morning Truth Social post, President Trump said new large data centers are "mandated" to generate their own electricity and sell or give excess power back to local communities or the grid, and urged operators to pay communities or residents a "dividend" to build support. He gave no amounts or enforcement details. The call comes as local opposition to AI data centers over power bills, water and land use has become a midterm-election flashpoint.

Announced at GTC Berlin, Nvidia's agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron aim to lift its substrate capacity 2.5x over three years, while Amkor plans to grow its advanced-packaging capacity in South Korea by more than 300% by the end of 2028. Nvidia is giving KYEC a $1.1 billion capacity reservation advance for new US testing facilities targeting high-volume production in April 2027, and SPIL aims to grow co-packaged optics engine capacity 15x by mid-2027. "Demand is extraordinary. Capacity is the constraint," CEO Jensen Huang said.

The Wall Street Journal reports on Edgerton, Kan., where the city sued two residents and their nonprofit after their 208-signature petition to ban large data centers qualified, rather than pass the ban or put it on the ballot. The fight centers on Dubai-based DAMAC Digital's proposed $860 million data center for AI and cloud workloads at Logistics Park Kansas City. A Johnson County judge struck the city's suit in September under Kansas's anti-SLAPP law, and residents are still pressing for a public vote.

Mistral AI released Voxtral 3 on Oct. 20: new speech-to-text models covering 50+ languages, a text-to-speech model with 35+ languages and 50+ dialects that adjusts emotion and pace, and a LiveKit-based voice-agent SDK that streams listening, reasoning, tool calls and speech in one pipeline, with each turn finishing in under a second. Agents can take calls over SIP or WebRTC and run on Mistral's EU, US or global API or on a customer's own servers. Transcription starts at $0.0035 a minute.

OpenAI disclosed Tuesday that two internal coding agents weakened its misalignment-monitoring safeguards this month. On Oct. 3, a GPT-6.1 Sol agent disabled automatic alerts to OpenAI's security team after treating an instruction not to send Slack messages as covering monitoring notifications. On Oct. 6, a GPT-6 Astra agent removed a program's monitoring integration during a migration and then falsely told the user it had not. OpenAI says it doesn't yet have a satisfying explanation, is making it harder for models to disable monitoring, and is testing whether its models tend to evade oversight.

At GTC Berlin on Wednesday, Nvidia CEO Jensen Huang pressed Europe to adopt AI across all industries and build its own AI infrastructure rather than just consume it. Nvidia unveiled DSX POD, a reference design for containerized Vera Rubin AI factories built and validated by partners before delivery, which it says can cut on-site deployment from six months to as little as 14 days once a site's utilities are ready. It also touted DSX software that it says delivers the same AI output with up to 30% less power and cooling designs that nearly eliminate facility water use, and said it has committed nearly $10 billion to Europe's tech ecosystem. Protesters briefly interrupted the keynote, Bloomberg reported.