Silicon Valley Daily
Broadcom has agreed to lend Anthropic up to $42 billion, according to Anthropic's IPO prospectus, about a third of a $125.2 billion TPU lease the AI lab has signed; Bloomberg says the shares could start trading before Thanksgiving. OpenAI parted ways with three researchers it says mishandled sensitive information, and researchers showed a reasoning-extraction attack kept working on Azure for weeks after OpenAI blocked it on its own API. The New York Times reports Meta cut its 2025 tax bill by $3.9 billion by classifying AI data centers as research pilots. A federal judge threw out publishers' antitrust suits over Google's AI Overviews, and Trump told TIME he now calls AI “SI.” Funding went to security and chips: Armadin raised $255.5 million and Volantis $88 million.
Anthropic's filing shows a Broadcom loan of up to $42 billion as IPO nears
Anthropic's IPO prospectus shows Broadcom agreed to lend it up to $42 billion through a convertible note, enough to cover about a third of a $125.2 billion, five-year TPU lease commitment, Reuters reported. Broadcom is therefore Anthropic's chip supplier, lessor and lender at once; Anthropic says it does not expect any notes to be sold before the IPO. Separately, Bloomberg reported that Anthropic could start formal IPO marketing as soon as the week of November 9 and begin trading before Thanksgiving.
OpenAI parts ways with three researchers over alleged leaks to a safety group
OpenAI said it parted ways with three researchers after an internal investigation found they mishandled sensitive information outside company procedures. The Wall Street Journal, citing sources, reported the information went to a third-party AI safety organization; OpenAI has not said what was shared or how. The departures come as the FTC investigates OpenAI, Anthropic and other AI companies, and a week after OpenAI held back GPT-6.1 Astra from public release over safety concerns.
Meta saved $3.9 billion in 2025 by labeling AI data centers research pilots
Meta claims a federal research tax credit by classifying its AI data centers as “pilot models” and Nvidia chips as experimental materials, the New York Times reported. The credit saved Meta $3.9 billion in 2025, up from $2 billion a year earlier and $700 million in 2023, the largest benefit among US public companies. Meta's filings warn the savings could be challenged, and its reserve for uncertain tax positions rose 45% to $18.74 billion; auditor EY, which helped set up the approach, is now pitching it to other chip buyers.
Kevin Mandia's Armadin raises $255.5 million at a valuation above $2.5 billion
Armadin, the AI security-agent company started by Mandiant founder Kevin Mandia, raised a $255.5 million Series B co-led by a16z and Accel. The company says the round values it at more than $2.5 billion and brings total funding to $445 million; Bain Capital Ventures and Redpoint joined as new investors. Armadin builds autonomous agents for cyber defense.
Judge Mehta dismisses Chegg and Penske antitrust suits over AI Overviews
US District Judge Amit Mehta on Wednesday dismissed antitrust lawsuits from Chegg and Rolling Stone owner Penske Media, which argued Google coerced publishers into supplying content to AI Overviews for free. “An expectation is not an agreement. It is simply how a general search engine works,” Mehta wrote, adding that antitrust law cannot substitute for lawmakers addressing the economics of new technology. Mehta is the judge who ruled against Google in the 2024 search monopoly case.
Microsoft's Office and Teams chief Ryan Roslansky leaves; Copilot takes over
Ryan Roslansky, the former LinkedIn CEO who took over Office and later Teams, will leave Microsoft at the end of 2026 after nearly 18 years. In his memo he said leading the next phase would require being in Redmond full time while his family is in the Bay Area. Office and Teams move to Charles Lamanna's Copilot, Agents and Platform group, a week after Nadella pitched the new Copilot as an “OS for work” with Office built in. Phil Spencer and Rajesh Jha left Microsoft earlier this year.
Reasoning-extraction trick kept working on Azure after OpenAI closed its API
OpenAI says it shut down a campaign that sent 16,000 extraction requests from more than 4,000 users on July 24–25, and ties a core group to people associated with Moonshot AI. Researchers led by Joachim Schaeffer say that when they retested on September 13, the attack was blocked on OpenAI's and Anthropic's own APIs but still worked on Microsoft Azure against every OpenAI model they tried, including GPT-6 Astra, and against Anthropic models up to Sonnet 5. By their timeline, OpenAI added safeguards to the Azure endpoint on September 27, and the Anthropic extraction stopped reproducing there on September 28.
Trump says Amodei differs from his media image and renames AI “SI”
In a TIME interview on September 28, President Trump said Anthropic CEO Dario Amodei's views are “much different” from “what is portrayed in the media,” and said he no longer calls the technology AI but “SI,” for superintelligence. He added he would not “kill a thing based on phony threats.” Axios reports Trump declared “superintelligence” the executive branch's preferred term at Tuesday's White House lunch with tech leaders. Apple was the one megacap not represented; investor Brad Gerstner called its absence a “nothing burger.”
Micron sues YMTC, saying it poached engineers and patented Micron know-how
Micron filed a US lawsuit accusing Chinese memory maker YMTC of hiring away key Micron engineers, who then patented technology tied to their Micron work. Micron alleges YMTC is using those patents to sue it for infringement, and that some of the patents rightfully belong to Micron. The two companies have been in patent litigation since YMTC first sued Micron in the US in 2023.
Proofpoint: China-aligned TA419 posed as an Anthropic employee to phish AI policy experts
Proofpoint says a group it tracks as TA419, which it assesses as China-aligned and espionage-motivated, ran credential-phishing campaigns in July against US AI policy experts, including invitations to a fake AI policy advisory committee. The group earlier impersonated a prominent Anthropic employee in February and spoofed a former White House official, according to the report. Proofpoint does not say Anthropic's own systems were breached.
Glow: AI agents exposed 13,000 internal screenshots on public GitHub
Security startup Glow says its PixelLeak research found more than 13,000 internal images from over 300 organizations exposed across 900-plus public GitHub repositories. Glow attributes many cases to coding agents that, lacking a private way to host screenshots, uploaded them to public repos as a workaround; about a third of affected organizations used a tool called gitshot. The Decoder reports the images included customer data, login credentials and unreleased products. Glow says it began notifying affected organizations on September 9.
Anthropic makes Claude for Government generally available to civilian agencies
Anthropic made Claude for Government generally available to US federal and state agencies in a FedRAMP High authorized environment. Agencies pay for usage with hard spending caps, can set department budgets and get audit logs; sensitive operations on Anthropic's side require two-person approval. Claude Code CLI and Claude for Microsoft 365 remain in early access. The Decoder notes the Pentagon still classifies Anthropic as a supply-chain risk.
DeepSeek and Huawei open-source programming tools for Ascend 950 chips
DeepSeek and Huawei released open-source DeepGEMM-Ascend and DeepEP-Ascend compute and communication libraries, plus native TileLang support, for Huawei's Ascend 950 chips. Both libraries were developed and tested on Ascend 950 hardware. DeepEP's documentation says its published performance used a specially configured proof-of-concept HDK, with the recommended commercial HDK planned for around October 15.
Volantis raises $88 million to link AI chips and memory with Face ID-style lasers
San Francisco chip startup Volantis raised $88 million in venture capital, Reuters reported. It plans to use vertical-cavity surface-emitting lasers, the same kind used in the iPhone's Face ID, to connect AI processors with memory chips. Optical links of this kind aim to let one processor reach more memory than short electrical wires allow.
Newsom vetoes California bill on secret recording with smart glasses
California Gov. Gavin Newsom vetoed SB 1130, which would have penalized secretly recording people with camera- and microphone-equipped wearables. In his veto letter Newsom said the bill defined wearable recording devices “too broadly or imprecisely” and duplicated protections already in state law. The bill would have made California the first state to regulate smart glasses; TechCrunch notes Meta sold more than 7 million wearable glasses last year.
Google's first orbital TPUs ride SpaceX's Transporter-18 with Planet satellites
Alphabet scheduled its TPUs to launch Thursday on SpaceX's Transporter-18 mission from Vandenberg, aboard a solar-powered Planet Labs prototype. It is the first in-orbit test of Project Suncatcher, Google's effort to run machine learning infrastructure in space. Satlyt, a Sunnyvale and Nairobi startup that just raised an $8 million seed round, has software on the same launch; it wants to let many companies' satellites share computing jobs. Alphabet's stake in SpaceX is worth more than $82 billion.
AWS and Cloudflare release open decision models for agents
AWS's Strands team released Strands Decider 2B, a decision model built on Qwen3.5-2B that picks among specified answers rather than generating free text; code, weights, training data and scripts are open and it runs locally on CPU or GPU. Cloudflare released Clef and Clef-flash, decision models with Apache 2.0 weights, image input and a 64K context window, hosted on Workers AI. Cloudflare's new RL fine-tuning service starts through its forward-deployed engineers, with self-service planned later.
Samsung raises US Galaxy S26 prices by $100, the 1TB Ultra by $200
Samsung raised US prices on most of its Galaxy S26 lineup by $100 and on the 1TB Galaxy S26 Ultra by $200, about six months after launch. Prices of the Galaxy Z Fold 8 and Z Flip 8 are unchanged. The increase was reported by PhoneArena and confirmed by more than a dozen outlets.
Barrett Lyon's Doxx.net raises $38 million Series A led by a16z
Miami-based Doxx.net, which offers serverless peer-to-peer calling, messaging and file transfer for people and AI agents, raised a $38 million Series A led by Andreessen Horowitz as its platform enters open beta, Axios reported. In an a16z interview, founder Barrett Lyon, who sold his DDoS-defense company Prolexic to Akamai, described the service as a way to communicate without leaving a record. The company's name drew criticism online.
University AI beats Stratego's top player after under $8,000 of training
Researchers from Carnegie Mellon, NYU, Stanford and MIT report in Nature that their Ataraxos system beat Pim Niemeijer, considered Stratego's best player, 15 wins to one loss with four draws over 20 games. Training took one week on 16 Nvidia H100 GPUs plus four days on four GPUs, under $8,000 at 2025 prices. The authors estimate DeepMind's DeepNash, which lost to top players in 2023, used $3 million to $4.5 million of compute.
Shopify launches Canvas, an AI chat workspace for designing online stores
Shopify introduced Canvas, a desktop workspace where merchants design their online store by chatting with its Sidekick agent while an interactive preview renders the store's actual theme code. Sidekick edits theme files and checks screenshots as it works. Shopify says Canvas rolls out over the coming days; TechCrunch reports initial gaps in support for third-party themes, app blocks, markets and translations.
Blue Cross links AI-assisted hospital coding to $942 million in extra charges
The Blue Cross Blue Shield Association estimates hospitals' use of AI-assisted medical coding contributed to $942 million in extra costs for its plans between 2023 and 2025. About 70%, or $653 million, came from added diagnoses that were not accompanied by any change in care, the group told CNBC; 60% of hospital systems adopted AI coding tools in that period. The American Hospital Association said the analysis lacks context and that patients are older and more complex.
Benchmark leads a round in UK chip startup Tendrils Compute, Newcomer reports
Benchmark won a competitive process to lead a round in Tendrils Compute, a Cambridge, UK chip startup led by Nils Cremer, Newcomer reported, citing people familiar with the matter. The company is already discussing a follow-on raise that could value it above $1 billion, sources said. Tendrils is building general-purpose chips based on interaction nets, a graph-based computing model, aimed at CPU bottlenecks in multi-step agent workloads.
AWS open-sources Dogwood, a local rules engine for agent tool calls
AWS released the Dogwood Local Engine under Apache 2.0, a local Rust library that returns allow or deny verdicts on an agent's tool calls by checking them against time-ordered rules over a durable history of earlier calls. Policies can change mid-session. The engine does not block calls itself: the agent harness must enforce verdicts, report events accurately and isolate the engine.
Sony brings AI upscaling to the base PS5 with QSSR
Sony launched Quick Spectral Super Resolution, an AI upscaling tier for the standard PS5 developed through its Project Amethyst work with AMD. Marvel's Wolverine and Ghost of Yōtei support it first. Because the base PS5 lacks dedicated machine-learning silicon, The Verge says players should expect output closer to 1440p, versus 4K from PSSR on the PS5 Pro.
Amazon refreshes Kindles, adds a $35 page-turn remote and AI chats in Audible
Amazon refreshed the Kindle, Kindle Paperwhite and Kindle Colorsoft with a flush front display, thinner and lighter designs, new colors and up to 32GB of storage, and introduced the $35 Kindle Click Bluetooth page-turning remote. Audible is rolling out features to track characters and places in a book and to talk with characters using generative AI.
New York City's click-to-cancel rule takes effect
New York City's click-to-cancel rule took effect Thursday, requiring businesses to make canceling a subscription as easy as signing up. It is the first municipality to have such a rule in force after an appeals court struck down the federal FTC version. Fines start at $525 per violation, and the city estimates savings for residents of $21.5 million to $162.5 million a year.
Google caps Chromebook updates at mid-2034, short of its 10-year pledge
A Google support document says ChromeOS devices will receive updates and security patches through mid-2034. Devices whose promised 10-year support would run past that date will instead get help moving to Googlebook OS, the Android-based platform Google launched a week ago. OSNews argues this breaks the 10-year update promise for recent Chromebook buyers.
Study: Opus 5.5 says “this matters” 116 times more often than humans
Marketing firm Graphite had frontier models rewrite 10,000 pre-ChatGPT articles and found 13,000 phrases at least twice as common in AI text as in human writing. Claude Opus 5.5 uses “dependable” 23 times more often than humans and “this matters” 116 times more often; OpenAI's Astra overuses “not simply X” framings more than 100-fold. Opus 5.5 uses em dashes 99% less than Opus 5, but Graphite says the total number of tells is holding steady.
Ideogram 4.5 edits only the selected part of an image
Ideogram released Ideogram 4.5, which it says changes only the areas a user asks to edit while leaving the rest of the image intact. It outputs native 2K images across four quality tiers priced from 0.8 to 22 cents per image, and is available on Ideogram's platform and API. Partners include Picsart, Runway, Pika and Leonardo AI, and Ideogram says an open-weight release is coming.
Ai2's Olmo-core 3 benchmarks training infrastructure for trillion-parameter MoEs
The Allen Institute for AI released Olmo-core 3, open training infrastructure for large mixture-of-experts models. It reports system benchmarks for a 1.2-trillion-parameter configuration on 512 Nvidia B300 GPUs and a separate short capacity test at 2.38 trillion parameters. The tests used random routing to measure system performance; Ai2 describes its next Olmo model as future work.
Weizmann team's AI reconstructs what people see from brain scans
Researchers at Israel's Weizmann Institute led by Michal Irani built an AI tool that reconstructs images a person is viewing from fMRI scans, and can also predict brain activity from an image. Irani hopes it could help locked-in patients communicate. UC Santa Barbara neuroscientist Tommy Sprague told MIT Technology Review that a similar approach could be used to extract mental imagery without consent.
Leaked video: GrayKey tool keeps locked iPhones from auto-rebooting
A leaked law-enforcement training video obtained by 404 Media shows Magnet Forensics, maker of the GrayKey phone-unlocking tool, pitching GrayKey Preserve and an Evidence Preservation Mode designed to defeat the iPhone's inactivity reboot, which locks police out of seized phones. In the video, a Magnet employee says the product disables Bluetooth, Wi-Fi and cellular radios on capture. The video appears to date from early 2025.
Most large Dutch data centers do not report power and water use
An investigation by Lighthouse Reports with Trouw and other European outlets found most data centers withhold energy and water data despite an EU reporting duty for sites of 500 kW or more. The Netherlands has 186 such commercial data centers, but public figures exist for the electricity use of only 44 and the water use of 47. Dutch data centers used 5.1 billion kWh in 2024, 4.6% of national electricity consumption.
Chesky says agents need an OS; Photon raises $4.5M to replace apps with agents
Airbnb CEO Brian Chesky, whose company rolled out AI-powered search this week, told TechCrunch chatbots are poor for browsing and shopping and that AI agents need an operating system of their own. Photon, which helps developers build agents that work over iMessage, SMS and WhatsApp, raised $4.5 million in seed funding; it says it has signed up more than 40,000 developers and grew revenue tenfold in four months. Photon held a mock funeral for mobile apps in San Francisco on September 17.