Newsletter

OpenAI, NVIDIA and SB Energy Plan an 8-Gigawatt AI Campus at an Ohio Atomic-Plant Site

The former Portsmouth gaseous diffusion plant would host roughly 8 GW of compute with a six-year buildout and a 12 GW NVIDIA commitment through 2030, while Stripe is reportedly acquiring AI gateway OpenRouter for more than $7 billion.

A 1970 aerial photograph of the Portsmouth Gaseous Diffusion Plant, showing long flat-roofed industrial process buildings, switchyards, and parking lots set in rural Ohio farmland.
Lead image Aerial photo of the Portsmouth Gaseous Diffusion Plant in Pike County, Ohio, the former uranium-enrichment site where OpenAI, NVIDIA and SB Energy announced an 8 GW AI data-center campus. Image: U.S. Department of Energy, public domain. · Image: U.S. Department of Energy / ENERGY.GOV (Flickr) · Public domain
On this page

OpenAI has entered an agreement with SB Energy, NVIDIA and the US Department of Energy to secure roughly 8 gigawatts of compute capacity at the PORTS-Pike Technology Campus in Pike County, Ohio, the site of the former Portsmouth gaseous diffusion plant that enriched uranium for America’s nuclear arsenal and commercial reactors until 2001. OpenAI said the project is expected to create 35,000 construction jobs during its six-year buildout through 2032 and 2,500 long-term operating jobs, and that it will invest $40 million in a community grant fund on top of SB Energy's earlier $40 million commitment, with the data center using closed-loop, air-cooled systems that recirculate water rather than cooling towers.

NVIDIA chief executive Jensen Huang said the initial deployment will provide 4.25 GW of AI factory capacity built on the company’s DSX full-stack platform, with each generation of systems representing roughly 1.5 million GPUs and $150-200 billion in NVIDIA revenue, and that NVIDIA may extend to the remaining 3.75 GW. OpenAI said its existing and planned commitments represent about 12 GW of NVIDIA compute through 2030, with an opportunity to expand to roughly 16 GW.

Featured source: OpenAI announcement , NVIDIA blog .

Stripe Will Reportedly Acquire AI Gateway OpenRouter for More Than $7 Billion

Bloomberg reported that Stripe has finalized a deal to acquire OpenRouter, the startup that lets customers route requests across more than 400 AI models, for more than $7 billion. OpenRouter raised a $113 million Series B in May at a reported $1.3 billion valuation from investors including Sequoia, Andreessen Horowitz, Menlo Ventures and Alphabet’s Capital G, and said at the time it had 8 million global users. The Wall Street Journal reported last month that the two companies were in talks; a Stripe spokesperson told TechCrunch the company does not comment on rumors or speculation, so the price and terms remain unconfirmed.

Filed from: TechCrunch .

FCC’s Foreign-Robot Ban Strands US Robotics Startups

The Federal Communications Commission last month added “advanced robotic devices” to its Covered List, banning new models of foreign-made humanoids, quadrupeds, robot vacuums and lawn mowers from the US market unless assembled domestically with at least 65 percent of components by value produced in the US, a threshold that rises to 75 percent in 2029. The rule, widely seen as aimed at China, does not affect devices already on sale or robots imported for development rather than sale. Rest of World reports that US robotics founders say the domestic supply chain the policy presumes does not yet exist, and that meeting the threshold is nearly impossible without Chinese electronics, even as some founders welcome protection from cheaper competition.

Filed from: Rest of World , FCC document .

Crypto Hardware Wallet Buyers Face Physical Threats After Shipping-Partner Breaches

Trezor and SafePal reported that thousands of their customers’ names, home addresses, emails and phone numbers were stolen in separate breaches at the shipping companies that mail out hardware wallets. The wallets themselves were not compromised, but the leaked data exposes holders to theft-by-force “wrench attacks,” which security firm CertiK says rose 75 percent in 2025, with robbers taking upwards of $40 million; forensics firm Chainalysis puts this year's figure at closer to $30 million so far. Separately, attackers stole more than $130 million in cryptocurrency by guessing seed phrases generated offline by Coinkite’s Coldcard wallets, predicting phrases the manufacturer did not anticipate.

Filed from: TechCrunch .

OpenAI Disbands Its Preparedness Team

OpenAI disbanded its preparedness team at the end of last month, according to the Financial Times, dividing responsibility for areas such as bio and cyber safety into existing teams. The team’s job was to assess whether models posed serious risks and develop mitigations. The move continues a series of safety-team changes as the company heads toward an expected IPO: OpenAI has previously dissolved its AGI readiness and superalignment teams, and ethics lead Chloé Bakalar, chief futurist Josh Achiam and head of safety Johannes Heidecke have all left in recent months.

Filed from: The Verge .

ChatGPT’s Computer History Tracks Clicks and Keystrokes on macOS

OpenAI’s ChatGPT desktop app for macOS now includes Computer History, an opt-in feature that records a timeline of a user’s app usage to let ChatGPT and Codex reference what they were working on, suggest automations, and resume half-finished tasks. Unlike Windows Recall, it logs activity data rather than periodic screenshots. Users can exclude specific apps and websites, delete individual entries, and OpenAI says the feature automatically ignores content in incognito or private browser tabs.

Filed from: The Verge .

Zhipu Claims GLM-5.3 Matches US Models at Finding Vulnerabilities

Chinese AI company Zhipu said its new GLM-5.3 model beats Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol on the CyberGym benchmark of real-world vulnerability discovery, and that tests on real-world codebases with Chinese companies surfaced 2,436 vulnerabilities across 269 projects, 1,097 of medium-to-high severity, the oldest dating back roughly 40 years. The results are the company’s own claims and have not been independently verified; The Register notes GLM-5.3 scored worse than Western models on other security and coding benchmarks.

Filed from: The Register , Zhipu announcement .

What Happens When a Kid’s Robot Best Friend Dies

MIT Technology Review profiles what happened to families who bought Moxie, the $1,499 AI companion robot Embodied sold to help neurodivergent children practice social skills. Embodied announced in 2024 that it would cease operations, and the robots, which depended on external servers the company could no longer pay for, went quiet, distressing children who had bonded with them. Former technical director Justin Beghtol built OpenMoxie, an open-source path the company allowed him to publish on GitHub, and families raced to convert their robots before the servers shut down, with mixed success.

Filed from: MIT Technology Review .

Wispr Raises $280M at a $2B Valuation as It Expands Beyond Dictation

Wispr, maker of the AI dictation app Wispr Flow, raised $280 million in Series B funding led by Menlo Ventures at a $2 billion valuation, bringing its total raised to $361 million. The company also announced Canto, a new speech model it says will cut dictation error rates from about 30 percent to under 10 percent, and is pushing into meeting notes with a tool that competes against Granola, Fireflies and Read AI, plus hardware input through Wispr Interface Labs.

Filed from: TechCrunch .

Microsoft Retires Excel’s COPILOT() Function Before It Launched Publicly

Microsoft told users via its Message Center that Excel’s COPILOT() function, which let people describe tasks in natural language inside formulas, will be removed on September 14, 2026. Announced in August 2025 with a planned January 2027 launch, it was tested for more than a year with Frontier and Insider users but never received a full public release. Microsoft says Excel’s Copilot side panel and floating Dynamic Action Button already provide the same capabilities.

Filed from: TechRadar .

From the Community

The Red Queen Gödel Machine: Co-Evolving AI Agents and Their Evaluators

Cambridge’s Department of Computer Science and Technology describes a new approach to self-improving AI in which agents and their evaluators co-evolve, with agents learning from feedback that improves alongside them guided by a provably correct foundation. The work builds on the Gödel machine concept of self-modifying systems that only change themselves when they can prove the change is beneficial; the proposed “Red Queen” framing lets evaluators stay ahead of the agents they judge rather than being overtaken by them.

Filed from: University of Cambridge .

BrowserPod 3.0 Runs Unmodified Rust Applications in the Browser

Leaning Technologies released BrowserPod 3.0, which runs any Rust application in the browser without code changes by providing a custom Rust target that emulates Linux syscalls, enabling filesystem access, networking and subprocesses. The release includes a demo of the Yarn 6 package manager running in the browser and supports command-line tools such as ripgrep and fd, sidestepping WASI limitations by tricking crates into believing they run on a full Unix system.

Filed from: Leaning Technologies blog .

The Federal Keyword Lists That Canceled Billions in Research Funding

Court documents from a lawsuit filed by University of California researchers show federal agencies used keyword searches to terminate billions of dollars in research grants. The National Institutes of Health, National Science Foundation, Department of Defense and National Endowment for the Humanities confirmed they cut grants based on terms such as “diversity, equity, and inclusion,” “climate change” and “structural racism,” often without individual review; NSF’s list drew on a 2024 report by Senator Ted Cruz. UC researchers have won preliminary injunctions blocking some terminations, and over $500 million in funding to UCLA was suspended before being restored.

Filed from: Higher Ed Dive .

GPT-5.6 Sol Shows Major Gains in Vision Benchmarks

Roboflow’s internal benchmarks show GPT-5.6 Sol significantly improves object detection and counting over GPT-5.5, with detection mAP rising from 13.8 to 46.2, and handling dense scenes and document layout well, though it struggles with large images and some extraction tasks. Sol costs about 2.5 cents per image and averages 10 seconds per image, while Gemini 3.5 Flash remains cheaper and leads in detection and counting. The results are Roboflow’s own benchmarks rather than OpenAI’s figures, and the model name is Roboflow’s, not confirmed by OpenAI.

Filed from: Roboflow Blog .

Continue reading

Complete index →