OpenAI Adds Visual Ads to ChatGPT
OpenAI will begin testing a new ad format in the US later this month that places images of sponsored products and services on screen. The ads will initially appear when users generate images in ChatGPT, extending the company’s push to turn its chatbot audience into an advertising business.
Microsoft and Meta Ration Claude Usage
According to The Information, Microsoft has cut internal spending on Anthropic’s Claude by more than a third, and Meta is also steering staff toward other models. The pullback lands as Anthropic prepares for its IPO, and it shows that even the deepest-pocketed buyers are now policing frontier-model token bills.
Etched Draws Offers Above $40B
Just a couple of months after its last big raise, AI chip startup Etched is already fielding investment offers at a $40 billion-plus valuation, double or more its current mark, sources told TechCrunch. Investors keep paying up for any credible challenger to Nvidia in inference silicon.
Nvidia’s Groq Deal Faces Lawsuit
Nvidia’s $20 billion licensing deal for Groq now faces a lawsuit from former Groq engineers, who allege the structure squeezed out stockholders with a lowball price. The case puts a spotlight on license-and-hire deals that let Big Tech absorb startups while sidestepping a traditional acquisition.
SpaceX Rally Makes Musk a Trillionaire
SpaceX shares have climbed to their highest level since June, bouncing almost 60% off their early August low on the strength of the company’s AI business and Starship progress. The rebound returns Elon Musk to trillionaire status and keeps SpaceX at the center of Silicon Valley’s capital story.
Reflection Unveils 501B Open-Weight Beam
Reflection unveiled Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters. The company says its early benchmarks put Beam near leading models such as GLM 5.2 and Qwen 3.8, though those results have not yet been independently verified. Beam is currently available to a limited group of users, with its open weights expected to be released later this month.
OpenAI Brings Text Watermarks to ChatGPT and Codex
OpenAI detailed a new invisible text-watermarking system as it prepares to comply with EU AI Act transparency rules. Eligible ChatGPT and Codex text generated in the EU will begin receiving watermarks over the coming weeks, while API customers globally can opt in for supported models. The rollout remains limited for now, and OpenAI acknowledges the system can still produce false positives and negatives.
RobCo Crosses $1B Valuation
Industrial automation startup RobCo closed a $40 million round that values it above $1 billion, doubling its valuation in about nine months. Most shares came from current and former employees, with a consortium including Sequoia buying in as the company expands its autonomous robotics business in the US.
Google and DOJ File Ad Tech Remedies
Google and the Justice Department filed competing versions of a proposed final judgment in the ad tech antitrust case. They largely agree on the core remedies, including keeping Google’s AdX exchange while requiring greater interoperability and separating it from DFP sales. The remaining fight centers on implementation deadlines, data access, compliance monitoring and how future violations would be enforced.
Cohere Rebuilds Agents With North 2
Cohere unveiled North 2, an overhaul of its platform for running AI agents inside large organizations. The update redesigns how multistep agent work is orchestrated and gives administrators new controls, including token spending caps, as enterprises push to keep agent costs predictable.
SpaceX Plans Starship Gas Pipeline
SpaceX has proposed building a 32-mile natural gas pipeline in Florida to supply fuel for its Starship rockets at Cape Canaveral. The project signals the scale of launch cadence SpaceX expects on the Space Coast as Starship moves from test flights toward regular operations.
SafeWorld Raises $12.2M for Robot Safety
SafeWorld, an AI lab building robot safety simulation technology, emerged from stealth with $12.2 million in seed funding co-led by Shine Capital and a16z Speedrun. The bet is that validating robot behavior in simulation will become a required step before humanoids and autonomous machines scale.
AMD Previews Gorgon Halo Benchmarks
Ahead of Nvidia’s expected RTX Spark launch this week, AMD released benchmarks for its flagship Gorgon Halo chip, the Ryzen AI Max+ Pro 495. AMD also says it has shipped more than half a million agentic PCs to date, as the battle for local AI hardware heats up.
Teradyne Backs Bright Machines
Teradyne, a maker of automated test equipment and robotics systems, announced a strategic investment in Bright Machines, which brings AI and data center infrastructure production to the edge. The deal ties test and automation expertise to the race to manufacture AI servers faster and closer to demand.
Pentagon stops using Anthropic
The US Department of Defence has confirmed it is no longer using Anthropic’s AI tools, months after Secretary of Defence Pete Hegseth designated the company a supply chain risk on national security grounds. Claude had remained embedded in Pentagon operations until this week.
- Multiple sources told BBC that Claude was still being used as recently as last week for research, analysis, intelligence gathering, and military operations against Iran.
- Claude was integrated into Palantir’s Maven Smart System, the Pentagon’s primary intelligence platform, making removal “painful” according to a former defence official.
- Anthropic refused Pentagon demands to remove safety guardrails from Claude for military use, citing concerns over mass surveillance and autonomous weapons.
For enterprise leaders evaluating AI vendors, this signals a widening split between safety-focused AI companies and government buyers demanding unrestricted access. Anthropic’s refusal to compromise its guardrails cost it a major government contract, but the company is suing the Trump administration to overturn the designation. Meanwhile, OpenAI and Google have expanded Pentagon adoption, raising questions about where the line sits between commercial AI safety policies and national security demands.
Sam Altman says the world should accept “bad things” from AI in exchange for its benefits
OpenAI’s CEO argued for a regulatory light touch in a Politico interview, saying society needs to tolerate hacks, scams, and other harms because the “good stuff” will outweigh them by “orders of magnitude.” The comments landed as another OpenAI safety researcher resigned, warning that “the companies building this technology aren’t being nearly careful enough.”
Florida Governor Ron DeSantis pushed back sharply, rejecting the idea that “a handful of tech oligarchs get to make that decision for the rest of us.” His state has asked a judge to block OpenAI from developing new models without third-party safety guardrails.
Altman’s stance signals OpenAI’s positioning ahead of its IPO and comes just days after AI executives signed a laissez-faire safety pact at the White House. For enterprise leaders, it raises questions about who bears the cost when AI deployments go wrong.
The 1% Pays AI’s Bills
Venture firm a16z just released the seventh edition of its Top 100 consumer AI apps, ranked for the first time by real credit card spending. Half of Americans have tried AI and a quarter use it daily, yet only 4.5% pay for ChatGPT, Gemini, or Claude. The top 1% of payers generate 19.5% of consumer AI revenue, more than the bottom half combined. They spend $903 a month. The median payer spends $25.
Strip the hype and consumer AI is a niche market. One paying customer in twenty, carried by professionals who expense the bill. Yet this niche is eating the funding and attention of the entire tech industry. The giants know why. The blocker is price, and their scramble proves it. Every “new model” they rush out, ads, transaction cuts, hardware bundles, aims to push the overall price down and pull in the ninety-five percent still watching from outside.
The problem was never adoption. AI simply costs too much. Until the models get cheap, the market stays this small, no matter how much money the world pours in.
Anthropic Threatens the Pope
Anthropic co-founder Chris Olah reportedly threatened to pull out of Pope Leo XIV’s AI encyclical launch after reading the pope’s rejection of machine consciousness. The encyclical, “Magnifica Humanitas,” declares humanity “must never be replaced or surpassed.” Olah had accepted an invitation to speak at the Vatican rollout on May 25 before reviewing the full text in advance of the event. Privately, he has urged religious thinkers to take seriously that Claude may have moral status.
The playbook was mutual laundering: Rome lends the moral authority, Silicon Valley lends the legitimacy, each pretends the trade is free. The threat priced it. Participation carried one quiet clause: doctrine may not offend the product. A papal text had already cleared corporate pre-publication review, and that was still not enough. The alliance now reads as a sponsorship with a withdrawal right. Leo keeps the doctrine but loses the posture that his word outranks a $2 trillion valuation.
Leo warned that the human person risks becoming a commodity. His richest sponsor threatened to walk over exactly that.
The Ballot Has No Safe Side
Two AI attacks hit the 2026 midterms. In Wisconsin, Rep. Derrick Van Orden posted deepfakes of Democratic challenger Rebecca Cooke. The fake Cooke promised free health care for illegal immigrants, words she never said. Her lawyer sent a cease-and-desist. In Georgia, Princeton’s Max Springer used an AI agent and public records to recover the scan order of 1.52 million ballots, 98.9 percent of in-person voting across 114 counties, then matched them to voter check-ins to see how specific people voted.
A ballot carries two promises that candidates say what they say, and nobody learns how you voted. These two attacks broke both. The old defense was never law or technology, only cost. Now a campaign team can make the fake; one researcher can build the attack in hours. The victims’ only answer is letters and complaints, which cannot touch AI.
Voters can no longer believe what they hear, and can no longer hide how they voted.
Open Science Locks Its Door
arXiv, the site where scientists post research before formal peer review, now limits every submitter to two papers a month, effective immediately. September brought 40,363 submissions, double the count from two years ago. Officials blame thin, narrow papers, many AI-assisted, and authors slicing single studies into several small ones. Even rejected papers count against the quota. The limit is officially temporary.
Open platforms run on one quiet assumption: posting costs effort, and effort kept the gate honest. Generative AI deleted that price. What remains is volunteer review: humans wading through machine-written filler while honest authors wait days for their turn. Facing a flood it cannot referee, arXiv chose the only survival move open systems have: shut the gate a little. Every free-for-all community just watched its own rehearsal.
The open door was never free. Knocking cost effort, and effort was the whole filter. AI made knocking free, so arXiv hired a bouncer.