The day's tech, sifted: Aug 17, 2026

What matters today: OpenAI signed a 20-year, 10-gigawatt data center deal in Ohio with SoftBank's SB Energy, with Nvidia agreeing to spend up to $105B to support the campus and invest $1.5B directly in SB Energy, the deal the financing guarantee Nvidia cut earlier this week was heading toward, and a comedown from the $3B Nvidia was reportedly weighing putting into SB Energy just two days ago. Stripe separately closed its $7B+ acquisition of AI model marketplace OpenRouter, a deal Stratechery reads as a bet on aggregating access to models rather than building them. Both land the same week OpenAI disbanded its AI risk assessment team and Anthropic said it has no plans to release a stronger internal model despite seeing AI risks rise, the industry's two biggest labs resizing both their financing and their safety structures in the same stretch of days.

AI / LLMs

• OpenAI disbanded its preparedness team at the end of July, the unit responsible for evaluating whether frontier models pose serious risks like rogue behavior or bio and cyber misuse; the Financial Times reports the responsibilities were split across existing teams instead, the third such safety group OpenAI has dissolved since 2024.
• Anthropic says it has no plan to release "Model 2," an internal model it considers more powerful than Mythos, even as it sees its models' automated research abilities accelerating: Axios reports the company still judges the risk of serious harm as low but is finding its own systems harder to evaluate, while OpenAI is separately slow-walking a model called Astra because it cannot rule out critical cyber capabilities.
• Anthropic confirmed its invisible Claude text watermark is a version of Google DeepMind's open source SynthID-Text approach, which biases word choice in a detectable pattern; the official explanation follows John Gruber's critique yesterday that Anthropic had rolled the feature out to comply with the EU AI Act while explaining almost nothing about how it works or what it costs writing quality.
• Claude's Riemann zeta bound push, which raised the number from 41.6% to 67.2% over a second 54-hour run, traces back to an Anthropic employee with one semester of high school geometry (paywalled): Jarred Sumner mostly just told Claude to "keep going," and a Stanford number theorist called the result the most impressive math output AI has produced so far, still nowhere near a proof of the hypothesis itself.
• Qwen 3.8 27B produced the best local pelican-on-a-bicycle SVG Simon Willison has seen, and took 21 minutes and 22,276 reasoning tokens to do it: the 27B open-weights model runs on a laptop but defaults to "wildly overthinking," with one tester watching it cycle through "FINAL FINAL APPROACH" before finishing a routine coding task.
• Face2Gene, an AI tool used by 70 percent of the world's geneticists across 130 countries, is increasingly how patients and doctors identify rare, hard-to-diagnose genetic disorders, the Wall Street Journal reports, comparing patient photos against a database of syndrome-linked facial features to surface diagnoses clinicians had not considered.

Devtools & Infra

• GitHub went down worldwide for hours, knocking out the website, API, Actions and pull requests; Microsoft confirmed the outage at 13:40 UTC, and the Hacker News thread on it pulled 616 points and 456 comments, developers locked out of PRs mid-workday and asking aloud whether it's time to look at alternatives.
• Buf shipped the first fully-featured, production-grade LSP for Protobuf, bringing go-to-definition, completion and semantics-aware highlighting to editors that have had none of it for a format used everywhere gRPC is.

Security & Privacy

• Criminals are hacking into public Wi-Fi routers at hotels and conference centers and changing their DNS settings to redirect users to fake login pages and steal credentials, Bruce Schneier writes, a low-tech attack that works precisely because public Wi-Fi trains people not to question a login prompt.
• Germany's antitrust regulator ordered Apple to stop steering users away from third-party app tracking consent, finding Apple's own App Tracking Transparency prompts gave its own apps friendlier consent flows than rivals got; Apple's prompts reportedly cost social apps nearly $10B in tracking revenue when they launched in 2021, and Apple has agreed to change the design.
• A cyberattack on France's tax agency exposed names, tax rates and income data for 678,000 individuals and businesses, prompting Prime Minister Sebastien Lecornu to call an interministerial crisis meeting; affected taxpayers started getting notified today.
• Switching a site's nameservers to Cloudflare silently injects its analytics tracking script, a Hacker News poster found, with no opt-in step required once traffic is proxied through Cloudflare's DNS; the thread pulled 326 points and accusations that a company marketing itself on privacy should not default to tracking by DNS change alone.

Startups & Industry

• Nvidia agreed to spend up to $105B supporting SoftBank's SB Energy Ohio campus and invest $1.5B directly in the company, as OpenAI signed a 20-year, 10-gigawatt lease on the site: the numbers land inside the reduced guarantee Nvidia set earlier this week, and below the $3B stake Nvidia was said to be weighing in SB Energy just two days ago.
• Stripe finalized its acquisition of OpenRouter, the startup that lets companies switch between AI models, for more than $7B, a five-fold jump from the $1.3B valuation OpenRouter raised at in May with backing from Sequoia, a16z and Alphabet's CapitalG.
• A tracked shipment of rare books led investigators to an Amazon facility that scans books for AI training data and destroys the physical copies afterward, 404 Media reports, one tracked order from book marketplace Biblio ending up at Amazon's Las Vegas facility.
• Nvidia disclosed a nearly $21B stake in SpaceX, 123 million shares that trace back to Nvidia's January investment in xAI, shortly before Elon Musk folded xAI into SpaceX; the stake's value has since dropped to about $17B as SpaceX shares slid from their June IPO.
• Chip giants have backed startup funding rounds worth more than $250B this year, with Nvidia alone in a record 59 rounds, Crunchbase reports; against that backdrop, Groq raised $350M at a $3.5B valuation, roughly half what it was worth last September, before Nvidia struck a licensing deal with Groq and hired away much of its talent.

Elsewhere

• Amazon reinstated mandatory arbitration for US customers and barred class-action lawsuits, five years after dropping the requirement when 75,000+ Echo users filed arbitration demands at once; the new terms treat 25 or more similar claims filed within six months as a "Mass Arbitration," batched instead of handled separately.

Hacker News

The GitHub outage covered above was the day's biggest HN story by a wide margin, 616 points and 456 comments (discussion), spawning a 159-point Ask HN on GitHub alternatives. AI's economics and executives filled most of the rest: a 373-comment thread formed around Dario Amodei's public case for AI regulation (discussion), a Futurism poll on why young people distrust AI CEOs fed the same mood, and a piece on the AI credit resale economy mapped the gray market reselling unused inference credits. Also today, covered above: Stripe's OpenRouter close, Nvidia's trimmed OpenAI financing guarantee, Anthropic's watermark explainer, Qwen 3.8 27B's overthinking habit and Germany's Apple ATT ruling.

A 404-point rebuttal argued RISC-V's edge is economics, not elegance, and a 326-point Tell HN caught Cloudflare silently injecting analytics on a nameserver switch, no opt-in required. A 288-point essay argued small models trade world knowledge for reasoning on purpose, citing hallucination rates near 80 percent even as math and code scores climb.

Threads

• Nvidia's OpenAI financing kept shrinking on the way to becoming real: a $250B guarantee cut to under $120B two days ago, a rumored $3B SB Energy stake, and today's signed numbers land at $105B and $1.5B, walking back exposure at every step while still closing the deal.
• OpenAI disbanding its risk assessment team and Anthropic ruling out a stronger internal model despite rising risk read as opposite responses to the same pressure: one lab thinning its safety structure, the other holding a model back rather than explain why it's safe to ship.
• Nvidia's $21B SpaceX stake and its now-severed leverage over Groq (a licensing deal and talent raid preceded Groq's valuation halving) show how deeply Nvidia's own fortunes are wound into the companies around it, a windfall when one IPOs, a shrug when another's price collapses.
• Germany ordering Apple to stop favoring its own apps in tracking prompts, Cloudflare quietly turning on analytics by DNS switch, and criminals hijacking public Wi-Fi DNS to steal logins are three different scales of the same problem: privacy defaults set by whoever controls the infrastructure, not the user.
• Amazon scanning and destroying rare books for AI training data and OpenAI framing a $105B Ohio campus around local jobs are both the buildout's usual asymmetry: what gets consumed quietly and what gets promoted loudly.