AI news, 21 August: Broadcom in talks for $60bn-plus AI financing

Broadcom seeks massive AI chip financing. Chip designer Broadcom is in talks with lenders to raise more than $60 billion in debt tied to artificial-intelligence hardware, Bloomberg News reported on 20 August, citing people familiar with the matter. Reuters noted the package could ultimately approach $100 billion when junior and senior tranches are combined, and that private-equity firms Blackstone and Apollo are among those in discussions. The money would support capacity for firms including Anthropic and others that use Broadcom’s custom chips, building on a June partnership aimed at expanding AI compute. The talks are not a closed deal; Broadcom did not immediately comment to Reuters. The scale underlines how far companies are turning to complex debt structures to fund the AI build-out.

Micron commits $10 billion to AI memory research. Memory-chip maker Micron said on 20 August it will invest $10 billion over the next decade in a new research lab in Boise, Idaho, according to Reuters. The facility is meant to advance memory technologies and compute systems that feed AI accelerators, especially high-bandwidth memory in heavy demand as data centres expand. Micron plans to break ground in 2027 and host hundreds of researchers, linking the lab to its global R&D network. The move sits alongside larger U.S. manufacturing commitments and reflects the Trump administration’s push for more domestic chip capacity.

Brazil splits AI supercomputers between China and the U.S. Brazil’s government announced roughly 2.3 billion reais (about $444 million) for its AI ecosystem on 20 August, Reuters reported. Just over half will fund supercomputing infrastructure in Rio de Janeiro with China’s Huawei and iFlytek; a separate tender of about 1 billion reais is expected to favour a U.S. supplier, with officials pointing to Nvidia. President Lula’s administration framed the dual track as a bid for “strategic autonomy” and data sovereignty rather than dependence on one country or vendor. The top-tier machine is slated to start operating by the end of next year.

Anthropic eases enterprise data rules amid privacy race. Anthropic plans to give business customers more control over data used with its advanced Claude models, a source familiar with the matter told Reuters on 20 August. Firms would still face a 30-day retention requirement for safety monitoring on frontier systems, but could keep that data on their own cloud infrastructure. The change, developed with more than 100 customers including Salesforce, follows OpenAI’s recent promise of stronger enterprise privacy protections that avoid retaining customer data while still flagging misuse. Bloomberg first reported Anthropic’s shift.

Amazon aims for drone delivery in nearly 500 U.S. locales. Amazon said it will expand Prime Air autonomous drone delivery to nearly 500 U.S. cities and towns by the end of 2026—about six times its current footprint—according to company statements covered by AI News and the Wall Street Journal. The drones rely on onboard detect-and-avoid systems rather than a pilot for every flight, carry packages up to five pounds, and target deliveries in as little as 30 minutes. New metro areas on the near-term list include Chicago, Atlanta, Cleveland, Syracuse and Boise. The service already operates under FAA Part 135 rules for commercial air carriers.

Study: AI writing marks a third of newer web pages. A Pew Research analysis released 20 August found that more than one-third (about 35%) of English-language web pages published after ChatGPT’s late-2022 launch show significant signs of AI authorship or heavy AI editing, TechCrunch reported. In a broader July 2026 sample that included older pages, the share was around 10%. Commercial .com sites showed far higher rates than .edu or .gov domains. Detection tools are imperfect, Pew and analysts note, but the directional finding—that much of the recent open web is machine-assisted—adds fuel to debates about information quality, search and the training data of future models.