Industry
Amazon drops secret data centre pacts and pledges $1 billion to host towns
AWS drops NDAs with government agencies and promises $1 billion in community investment as more than 100 US data centre bans are considered.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Data centre politics dominated the day as Amazon pledged a billion dollars and warned of blocked projects, while AI moved deeper into prescribing, hiring and advertising.
Amazon opens up on data centres as Washington closes doors
Amazon made a two-pronged intervention in the fight over data centre construction. Its Data Center Commitment programme ends nondisclosure agreements with government agencies, promises community open houses and annual reporting on energy, water and carbon-free electricity, while the Built Together initiative pledges more than $1 billion over five years for host communities, including free community-college access for over 300,000 students and efficiency retrofits targeting bill cuts of 20% to 40%. The timing is pointed: more than 100 data centre moratoriums are under consideration across the United States.
AWS chief executive Matt Garman paired the pledge with a warning, claiming projects worth $68 billion were blocked in the second quarter of 2026 alone and that America risks writing its own losing ticket in the AI race. He dismissed water and energy-bill complaints as myths and hinted, unnamed, at foreign misinformation. The claims sit against contested evidence, including reports of a 76% electricity price spike tied to AI data centres. Meanwhile the Senate killed the Ratepayer Protection Act, 57 to 43, three votes short, which would have directed regulators to consider making heavy electricity users cover the grid costs they add. Opponents called it toothless because it was non-binding, but with it dead there is no federal protection, and an Oregon utility's 30% data centre bill rise alongside a 1.3% residential cut shows utilities are already sorting these costs themselves.
AI takes on prescribing, interviewing and advertising
AI is edging into roles once reserved for professionals, with varying supervision. Nolla Health's Utah pilot lets an app scan a face, assess acne severity and write a prescription autonomously, though oversight fades gradually: two physicians approve everything for the first 100 patients, review comes after the fact for up to 500, and beyond that doctors check a sample of at least 10%. HackerRank says its AI interviewer has run more than 500,000 technical interviews for testers including Snowflake and Capgemini, and OpenAI has introduced a visual ad format in ChatGPT alongside measurement tools and brand suitability controls.
The public remains sceptical of this momentum. A Quinnipiac poll of 1,202 US adults found 47% want AI development slowed and 30% stopped entirely until safety is verified, with only 5% wanting it faster. Notably, 69% still think America must keep pace with China, and 72% oppose a data centre in their own community, a gap between national ambition and local resistance that mirrors the moratorium fight. An MIT committee, meanwhile, warns AI is eroding office hours, study groups and student-faculty trust, and calls for a complete overhaul of higher education.
Security and benchmarks: from forged cookies to graded code reviews
A critical flaw in Rejetto HTTP File Server, tracked as CVE-2026-61500 and scored 9.3, is now under exploitation despite being patched in July. Discovered by Horizon3 with help from Anthropic's Mythos AI model, the weakness stems from the server exposing outputs of a reversible random number generator, letting an attacker reconstruct the session-cookie signing key from a handful of login responses and forge administrator access. VulnCheck observed reconnaissance from a China Telecom address hitting canary servers in Japan and the US, so any unpatched internet-facing instance is a practical target.
On measurement, GitHub released ReviewBench, an open benchmark of 219 public pull requests across 19 languages for judging AI code reviewers, with severity and category labels and a reported 96.6% agreement on human-labelled true positives. Aleph Alpha released Kolibri, a German-English model with 78 billion parameters but only about three billion active per token via a mixture-of-experts design, trained on 768 Nvidia B200s and offered under Apache 2.0 for public administration, aviation and industry.
Hardware shuffles: quieter Windows, louder AMD, cheaper Switch
Windows 11's 26H2 update has quietly cut idle RAM use, with Windows Latest measuring 7 GB idle on a 16 GB machine where it previously sat at 10 GB, and Reddit users reporting reductions of 2 to 3 GB. The same update widens the Low Latency Profile, a brief CPU boost at app launch, to nearly all applications, with Edge and Outlook launches roughly 40% faster. Microsoft is also phasing out classic v3 and v4 printer drivers in favour of a built-in standards-based stack, with third-party driver updates limited to security fixes from July 2027, which may trouble offices with older printers.
AMD published generative AI benchmarks for its Gorgon Halo chips days before Nvidia's expected RTX Spark launch, claiming 1.1x to 32.2x advantages over Intel, though the comparison was not like-for-like and the 32.2x outlier rests on an unidentifiable model. Tom's Hardware notes AMD's measured throughput still trails Nvidia's DGX Spark. In brighter consumer news, the Nintendo Switch 2 has fallen to £354.99 at UK retailers, below even its original launch price, a rarity for Nintendo at a time when component costs have pushed every console maker's prices up.