Gates sounds alarm | Meta weighs settlement | EPA curbs data center notice

Bottom Line Up Front

Bill Gates published an essay saying AI has passed thresholds on biological and cyber risks, psychological harm and job destruction, and floated taxing robots and reserving some occupations for people. Meta and state attorneys general have discussed settling the teen addiction case midway through its second week of trial in Oakland. The Environmental Protection Agency proposed dropping mandatory public notice on air pollution permits for data centers and other small emissions sources, leaving the decision to state and local officials. Huawei offered Egypt more than 2,000 Ascend AI chips for government data centers, prompting U.S. officials to assemble a counteroffer from American chipmakers.

I.Top Stories

Gates says AI has crossed thresholds on biology, cyber and job destruction

Bill Gates published an essay Wednesday saying AI has passed the points at which several dangers should have been checked, and told MIT Technology Review that guardrails are not keeping pace with the technology. "We've crossed the threshold in terms of [AI's] bio-capabilities, cyber-capabilities, psychosocial capabilities, job-market-destruction capabilities, and even the lack of control," he said. Gates said he views bioterrorism as roughly 50 times more likely than a natural pandemic and that any model able to design novel molecules should be monitored. He floated reserving some categories of jobs for human beings, with the list varying by country. He also proposed taxing robots and the AI tokens (the billed units of model usage) consumed when software replaces human work. The essay is the first of several he plans on the subject.

Read at MIT Technology Review ↗ • Read at Gates Notes ↗

Meta and state attorneys general discuss mid-trial settlement in teen addiction case

Meta and the state attorneys general prosecuting it have discussed a mid-trial settlement of the case accusing the company of designing Facebook and Instagram to addict teenagers, people familiar with the talks told Bloomberg. The trial is in its second week in Oakland federal court, where the top legal officers of 29 states seek financial penalties and mandatory changes to how the platforms operate. Meta has denied the allegations and accused the attorneys general of seeking an "outlandish payout." Late Tuesday several attorneys general called press conferences, among them Nevada's Aaron Ford. His office said it would announce "a major consumer protection settlement involving a leading technology company." The states' claim under the federal Children's Online Privacy Protection Act covers collecting data from users Meta knew were children and using it to train machine learning and generative AI models.

Read at Bloomberg ↗ • Read at The Straits Times ↗ • Read at The Economic Times ↗

EPA proposes ending public notice on data center air permits

The Environmental Protection Agency proposed a rule dropping the requirement that states notify the public and take comments on air pollution permits for data centers and other minor pollution sources, The New York Times reported. The Clean Air Act makes that notice mandatory; the proposal would leave it to state and local officials' discretion. Developers routinely seek those permits for the diesel generators and gas turbines that back up their facilities, which emit nitrogen oxide, carbon monoxide and heavy metals. The agency said the change would "reduce administrative burden and responsibly speed up permitting" without altering emission standards. Joe Goffman, a former EPA assistant administrator for air, said the statute promised the public a voice: "The administration is basically saying: 'You all may have thought this was an ironclad promise, but it no longer is.'" The comment period has closed.

Read at The New York Times ↗ • Read at The Guardian ↗ • Read at Techstrong IT ↗

Huawei pitches Egypt on 2,000 Ascend AI chips as Washington readies counteroffer

Huawei has offered the Egyptian government more than 2,000 AI chips for data centers serving military, surveillance and other public sector uses, Bloomberg reported, citing people familiar with the matter and documents. The deal would be among the first known exports of Huawei's Ascend line. U.S. officials have approached Nvidia, AMD and Microsoft to put together an American counteroffer. Huawei's proposal also includes a partnership with the Chinese speech recognition company iFlytek covering vehicle and person recognition. Egypt has been subject to U.S. export restrictions on AI chips since 2023, so shipments of American technology there require Washington's approval.

Read at Bloomberg ↗ • Read at MarketScreener Sverige ↗

Mosseri tells Meta trial he called teen safety opt-in rates low without giving numbers

Instagram head Adam Mosseri became the first senior Meta executive to take the stand in the 29 state child safety trial in Oakland on Tuesday, disputing prosecutors' claim that his company hid unflattering findings from the public. Jason Slothouber, a prosecutor for Colorado, pressed him on Take a Break, a tool that prompts teenagers to close Instagram after long scrolling sessions. Slothouber noted that 1.8% of teens signed up and that Mosseri had promoted the tool without disclosing that figure. "I've definitely said publicly that the opt-in rates are low," Mosseri responded. "I don't know if I've said numbers." His testimony continues Wednesday.

Read at The Guardian ↗

II.China Watch

China's legislature takes up traffic law rewrite that pins self-driving violations on carmakers

China's top legislature began its first review of a rewritten Road Traffic Safety Law on Aug. 25, per Xinhua, the state news agency. The draft, before the Standing Committee of the 14th National People's Congress, adds a dedicated chapter of special provisions for autonomous vehicles and defines both autonomous driving and assisted driving functions. Where a vehicle breaks traffic rules with its autonomous function switched on, the penalty falls on the manufacturer or importer rather than the person in the seat. Cars running with that function off, and vehicles offering only assisted driving, stay under the existing rules for conventional cars. Individual clauses of the 2004 law were amended in 2007, 2011 and 2021, but this is its first systematic overhaul.

Read at 36Kr ↗

EHang pulls revenue forecast as regulators slow approvals for pilotless passenger flights

EHang withdrew its full year revenue target of 600 million yuan ($89 million) and declined to set a new one, Caixin reported. Chief Financial Officer Conor Yang told an Aug. 25 earnings call that the company would wait for clearer regulatory conditions. Since late June, Chinese regulators have taken a more cautious line across the low-altitude economy, Beijing's label for the drone and air taxi sector, which has slowed project approvals, delivery sign-offs and the start of commercial service. EHang booked 104 million yuan ($15 million) of revenue in the first half, 17.3% of the original target and 25.7% below a year earlier. The company had planned to begin ticketed flights on its two seat pilotless EH216-S in March and reaffirmed the 600 million yuan target as recently as June.

Read at Caixin ↗

XPeng robot unit raises more than $900 million with Tencent and Alibaba among backers

XPeng's robotics business closed a first funding round of more than $900 million on Aug. 24, valuing the unit above $6.3 billion, per Pandaily, which described it as the largest single private round yet in China's embodied intelligence sector, AI built into physical robots. IDG Capital led, and Tencent and Alibaba entered as strategic investors alongside Gaorong Venture Capital. Separately, XPeng's second quarter results showed a wider net loss and third quarter guidance below expectations. The unit still sources some of its high precision bearings, servo motors and sensor chips abroad. Chairman He Xiaopeng, who took over as chief executive of the robot business in June, is targeting mass production of the IRON humanoid by the end of 2026, with first deployments in XPeng's own stores and campuses.

Read at Pandaily ↗

III.Policy Tracker

GSA seeks device fingerprinting for Login.gov to screen out agentic AI and bots

The General Services Administration asked vendors on Tuesday for device fingerprinting technology to strengthen fraud controls on Login.gov, the government's shared sign-in service, saying fraud attempts have grown in sophistication and volume as agency adoption expanded. Device fingerprinting identifies visitors from browser, device and network signals rather than IP addresses or cookies alone. The request for information said the capability is needed against account takeover, synthetic identity fraud, residential proxy abuse and emerging threats from agentic AI, software that carries out tasks on a user's behalf. It should also classify automated traffic including large language model based agents built on ChatGPT, Claude and Gemini. Responses are due Sept. 11.

Read at FedScoop ↗

Pennsylvania sues Snap over compulsive use and Snapchat's 13-plus rating

Pennsylvania Attorney General Dave Sunday sued Snap in Philadelphia County Court on Tuesday, alleging the company misleads app stores and the public about how often sexual content, nudity, suicidal ideation and drug use appear on Snapchat while keeping an age 13-plus rating. The complaint, brought under state consumer protection law and seeking court orders rather than damages, singles out SnapStreaks and disappearing messages as features that make the app habit forming for teenagers. "Snapchat has designed its platform to lure children into constant, compulsive use that is detrimental to healthy adolescent development," Sunday said. About 50% of teenagers ages 13 to 17 use the app daily, according to the Pew Research Center.

Read at The Hill ↗ • Read at WPXI ↗ • Read at 90.5 WESA ↗

Paxton campaigns on data center limits in Texas Senate race

Texas Attorney General Ken Paxton, the Republican Senate candidate endorsed by President Trump, posted a four-point data center plan on X on Monday, The Guardian reported. It would bar Chinese technology from powering U.S. data centers, hold companies criminally liable if AI chatbots undermine children's safety and cut permitting rules so the facilities do not raise Texans' electricity bills. Paxton also backed Gov. Greg Abbott's state plan to keep new data centers out of rural neighborhoods, repeal their sales tax exemptions and require them to add generating capacity rather than draw it down. X labeled the graphic accompanying the plan as AI generated. In an Annenberg Public Policy Center survey last month, 61% of Americans opposed a new data center in their area, up from 49% in March, including 54% of Republicans.

Read at The Guardian ↗ • Read at Ken Paxton on X ↗

New York police logged immigration searches of Flock plate data as "investigation"

State and local police in New York searched the license plate reader database run by Flock on behalf of Immigration and Customs Enforcement, sometimes obscuring why, according to a report published Tuesday by two New York civil rights groups, the Surveillance Technology Oversight Project and LatinoJustice PRLDEF. The authors reviewed logs posted by Have I Been Flock'd, a site publishing Flock search records obtained through public records requests, and found officers in at least two towns citing ICE or "Wanted by ICE" as recently as July. Others logged civil or administrative immigration reasons for drivers suspected of no crime. A search logged at the New York State Fusion Center on March 24, 2025, cited a Department of Homeland Security agent and the southwest border; a state police officer ran the same plate an hour later, giving the reason as "Investigation."

Read at StateScoop ↗

IV.Capability & Research Watch

OpenAI publishes benchmark results for its Broadcom built Jalapeño inference chip

OpenAI said Tuesday that Jalapeño, the custom chip it developed with Broadcom and introduced in June, ran 1.5 to 1.9 times more AI work per watt and at 1.7 to 3.6 times lower latency start to finish than the best figures previously recorded on Nvidia's GB200 and GB300 systems. The comparison covered three open models: GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. The company measured the chip on InferenceX, a benchmarking platform for inference, the process of running an already trained model to answer a query or drive an agent. Richard Ho, OpenAI's vice president of hardware, told reporters the design avoids the usual trade-off between latency and throughput. Ho said deployment starts at small volume late this year and ramps through 2027.

Read at The Verge ↗ • Read at OpenAI ↗

Researchers argue AI agent design erodes the human oversight it depends on

Margaret Mitchell, Avijit Ghosh and Samir Passi posted a position paper to arXiv on Aug. 24 arguing that keeping a human in the loop, a commonly proposed answer to the risks of increasingly autonomous agents, does not work as currently built. Current practice works against oversight on two fronts, they write: agent systems are designed in ways that make close supervision harder, and extended use of automation wears down the cognitive capacities oversight requires. The authors call for treating overseers' needs as equal in priority to agent capability, and set out design features and organizational protocols meant to support critical judgment and counteract skill atrophy.

Read at arXiv ↗

V.Industry & Market Watch

Anthropic's IPO pitch would put its addressable market above $30 trillion

Anthropic is likely to tell investors that the total revenue opportunity for its products exceeds $30 trillion, people familiar with the matter told The Wall Street Journal. That is above the $28.5 trillion SpaceX presented before its own offering. Total addressable market is the annual revenue a company would collect at 100% market share. Anthropic is building its number from everything AI models could conceivably be put to work on. For scale, the 191 technology companies in the S&P 1500 generated $2.4 trillion in revenue last year, according to FactSet data. Anthropic more than doubled revenue to $11.6 billion in the second quarter and could seek to raise as much as $100 billion at a valuation near $2 trillion.

Read at The Wall Street Journal ↗ • Read at Quartz ↗

Emerald AI raises $150 million to make data centers dispatchable by utilities

Emerald AI said Tuesday that a round it described as oversubscribed brought in $150 million and valued the company at $1.05 billion, with the money going toward scaling commercial deployments. Its software schedules AI training and inference jobs against a site's batteries and onsite generation, cutting the facility's draw when the local grid is stressed while priority jobs keep running. That lets a utility treat the data center as a resource it can dispatch. The company ran five demonstrations on live commercial load over the past year, at sites in Arizona, Illinois, Virginia, Oregon and London, with partners including Nvidia and Oracle. Emerald estimates flexible operation could free more than 100 gigawatts on the existing U.S. system. The International Energy Agency puts data centers behind nearly half the growth in U.S. electricity demand through 2030.

Read at SiliconANGLE ↗ • Read at Emerald AI ↗

Gatik closes $200 million Series D, targeting more than 100 driverless trucks by year end

Gatik said Tuesday it closed a $200 million Series D that will take its driverless fleet from dozens of trucks to more than 100 by the end of this year. The company runs short regional freight routes between distribution centers and retail stores, and says it has completed 85,000 driverless orders with 99% arriving on time. PepsiCo signed a multiyear agreement in June and now has 41 driverless Gatik trucks hauling Frito-Lay freight around Dallas, Phoenix and northwest Arkansas, joining Walmart, Kroger, Tyson Foods and Canada's Loblaw on the customer list. Contracted revenue across that base runs past $600 million, and the trucks operate in Texas, Arizona, Arkansas and Canada.

Read at SiliconANGLE ↗ • Read at Gatik ↗

Gartner projects AI data centers to take half of chip revenue within four years

AI data center spending takes 36.5% of semiconductor revenue this year and will pass half within four years, Gartner said in a forecast published Monday. The research firm expects global semiconductor revenue to pass $1.6 trillion this year, nearly double last year's total. Memory remains the largest contributor for now, at an estimated 54% of 2026 semiconductor revenue. Construction Dive also reported that chief information officers are looking to cheaper models to rein in spending as compute costs rise.

Read at Construction Dive ↗

Hexaware chief sees AI cutting prices for routine IT work by up to 25%

Customers could demand fees 20% to 25% lower for some of the work Hexaware does today as AI takes over routine tasks, Chief Executive Srikrishna Ramakarthikeyan told Bloomberg. He said the Indian outsourcing firm will need fewer people for the same jobs. More than half of its revenue is already produced at least partly by AI, a share he expects to reach about 80% within three years. The company is moving away from billable hour contracts toward fixed price and platform deals, and expects the cost of running AI models to become a line item in customer technology budgets. Hexaware shares have fallen more than 20% since its early 2025 listing, while Infosys, Wipro and Tata Consultancy Services are each down about 40% over the same span.

Read at Bloomberg ↗ • Read at Yahoo Finance ↗

LinkedIn users clicked its AI slop button a million times in two weeks

More than 1 million people used LinkedIn's "seems like AI slop" reporting button in the first two weeks after launch, chief product officer Hari Srinivasan wrote last week. The feature, announced last month, lets members flag posts that appear to be machine generated. Srinivasan said members are seeing 40% fewer views of content the platform classifies as AI slop than they were a few weeks earlier.

Read at Fortune ↗

VI.Global & Geopolitics

Albanese wins state buy-in on AI data center rules after Queensland carve-out

Australian Prime Minister Anthony Albanese said he secured agreement from state and territory leaders to press ahead with Australia's first national AI laws, with data center standards to be legislated early next year, Bloomberg reported. Leaders stopped short of settling the standards at Wednesday's national cabinet meeting in Sydney and agreed to further work. Queensland and the Northern Territory, which had pushed for coal and gas to back the facilities instead of the renewable supply the federal plan would require of developers, won a carve-out allowing them to do so. Northern Territory Chief Minister Lia Finocchiaro told ABC Radio before the meeting that Canberra should not "impose rules on states and territories when they, under their own steam, are able to make these decisions for themselves for their people." She cited the Beetaloo Basin gas reserve as ample supply. Industry Minister Tim Ayres said the standards were meant to prevent a race to the bottom and to keep Australia's electricity system from going backward as it modernizes.

Read at Bloomberg ↗ • Read at InDaily ↗ • Read at ABC News ↗

Vietnam writes AI rules into its publishing law

Vietnam's National Assembly approved amendments to the Publishing Law on Aug. 24 that prohibit using AI to commit illegal acts in publishing, require notices on AI created material and widen the grounds for revoking editors' certificates, JournalismPakistan reported. The vote came at the first extraordinary session of the 16th National Assembly, where 472 of the 475 deputies present backed it. The law takes a two track approach, encouraging publishers to adopt AI, electronic publishing and digital infrastructure while placing new technology obligations on publishers, editors and digital distribution services. Most provisions take effect March 1, 2027.

Read at JournalismPakistan ↗ • Read at Quốc hội ↗

IMF's Georgieva says AI investment is broadening beyond the U.S. and propping up global growth

Kristalina Georgieva, managing director of the International Monetary Fund, said the AI investment boom is broadening past the United States, with other countries putting money into data centers and related infrastructure and economies tied to the AI hardware supply chain gaining from the demand. She said that expansion is cushioning some of the economic damage from the Iran war and the closed Strait of Hormuz, and that the world economy has absorbed the energy shock better than the fund had feared. Georgieva warned the shock is not over, with oil and gas reserves falling ahead of the northern hemisphere winter. The fund estimated in July that the world economy would grow 3% this year, down from 3.5% in 2025, and will publish updated forecasts in October.

Read at Financial Times ↗ • Read at CryptoBriefing ↗