Silicon Valley Daily
Elon Musk says that next year he will give one SpaceX share to each of the roughly 72 million children in the US, about $10.5 billion at Friday's price, a day after Anthropic set aside 5% of its equity for a public benefit fund. Washington moved against Chinese AI on two fronts: DHS formally barred DeepSeek from federal civilian systems, while reports say DeepSeek is helping Huawei design chips it may pre-train on by year-end. Capital kept flowing to hardware: Tenstorrent raised $1.5 billion at $6.2 billion, Amazon will fund a 12-reactor nuclear plant and released a tool to move PyTorch code onto Trainium, and Meta detailed a cluster headed for 1 million GPUs. At the same time, executives expect only 7 to 10 neoclouds to survive, a risk for Nvidia. Security news was dominated by agents: one attacker used them to breach more than 100 sites, and a distillation attack exposed OpenCode users' API keys.
Musk pledges one SpaceX share to every child in America next year
Elon Musk posted on X on Monday that next year he will give one of his SpaceX shares to every kid in America, calling it "the first of many." Axios counts just over 72 million Americans under 18; at Friday's close of $146.03 that is a little over $10.5 billion. Musk did not say who qualifies or how the shares would be delivered. The post came a day after Anthropic said it would put 5% of its equity into a public benefit fund; Musk replied that the shares should go directly to the public.
DHS orders federal civilian agencies to remove DeepSeek and bars future purchases
The Department of Homeland Security ordered all federal civilian agencies to remove DeepSeek technology from their IT systems and barred future procurement of its products and services, the Wall Street Journal reported. According to CyberScoop, the notice is due in the Federal Register on Tuesday and includes a June 26 letter from Secretary Markwayne Mullin, so the exclusion has been in force for months. Mullin cited data stored on servers in China; agencies can seek exceptions with 60 days' notice.
DeepSeek works with Huawei on future AI chip designs, The Information reports
DeepSeek is working closely with Huawei to shape the design of future AI chips, The Information reported, citing two people familiar with the matter. DeepSeek plans to start using Huawei's Ascend 950DT for pre-training as soon as the end of 2026, the report said. Moving pre-training, the most compute-hungry stage, onto domestic chips would mark a step beyond using them for inference.
Meta details Prometheus, its gigawatt Ohio cluster headed for about 1M GPUs
Meta published an engineering post on Prometheus, its gigawatt-scale supercluster in Ohio, which it says will hold about 1 million GPUs at full scale. The cluster trained Meta's recently released Muse family of foundation models. The disclosure lands in a week when investors are questioning how much AI capacity independent cloud providers can finance.
Tenstorrent raises $1.5B at a $6.2B valuation and stays independent
Inference chip startup Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double the $2.6 billion it fetched in late 2024, TechCrunch reported. Five months ago Intel and Qualcomm reportedly showed acquisition interest at valuations of up to $5 billion. CEO Jim Keller told Bloomberg the company chose to stay independent after showing its Galaxy servers could run DeepSeek models. Trace Capital and Innovation X led the round.
Amazon to fund 12-reactor, 960 MW nuclear plant at old Indiana coal site
Amazon will fund a proposed 960-megawatt nuclear plant in Indiana and be the first buyer of its power. X-energy will supply 12 Xe-100 small modular reactors and utility AEP will operate the plant, on the site of a demolished coal station south of Terre Haute. It is the second Amazon-X-energy deal and three times the size of the first. No construction timeline was given, and X-energy has yet to break ground on its first plant in Texas.
Hacker used autonomous AI agents to breach 100-plus sites, incl. John Deere
A hacker broke into more than 100 websites using AI agents that mostly ran on their own, according to security firm Norforge. Reuters named John Deere, SharkNinja and Cummins among the victims. The attacker ran the open-source Cairn platform, laid out intrusion tactics in fewer than 100 Chinese characters, and successful breaches took about 23 hours. John Deere said there was no indication sensitive data was accessed; SharkNinja said a standalone support site was hit and contained.
Feds probe ransom negotiators' ties to ShinyHunters; Florida DMV notifies victims
Federal investigators pursuing ShinyHunters, which stole FBI data, have identified executives at two ransomware negotiation firms with unusually close ties to the group, KrebsOnSecurity reported. Sources said one US executive helped run the Shinywire extortion site, which listed nearly 100 victims between May and October 16. The FBI says ShinyHunters has extorted more than $70 million this year. Separately, Florida's motor vehicle agency began notifying people hit by a September 4 breach that NBC Miami linked to the group.
ProPublica: DHS appointees asked CISA to secretly hack a voting machine
Political appointees at the Department of Homeland Security asked CISA to secretly acquire and hack a voting machine and report the findings for the White House, without alerting the vendor, ProPublica reported, citing two former officials. Career staff said it was impossible, partly because the staff able to do it had been laid off, and the request was dropped. ProPublica says more than 1,000 CISA employees have left under the administration, about a third of the agency.
Town raises $108M Series B at $1B, four months after public launch
Town, a two-year-old San Francisco startup that gives each employee an AI assistant, raised a $108 million Series B at a $1 billion valuation, co-led by Forerunner and Menlo Ventures. It launched publicly four months ago and says it reached millions of dollars in annual recurring revenue within weeks. Town was founded by former Plaid CTO Jean-Denis Greze and has raised at least $160 million.
MiniMax launches M3.1 Flash at $0.15 per million input tokens
MiniMax released M3.1 Flash, saying agentic task costs are 70% lower than its predecessor and inference is nearly three times faster. Pricing is $0.15 per million input tokens and $0.60 per million output. The company reports 73.7% on DeepSWE v1.1; The Decoder says the model scored 42 on the Artificial Analysis Intelligence Index, 13 points above its predecessor, roughly matching DeepSeek V4.1 Flash on both score and price. MiniMax says weights will follow.
Amazon's TorchNeuron runs PyTorch code on Trainium without rewrites
Amazon introduced TorchNeuron, a native PyTorch backend that lets developers run existing PyTorch code on its Trainium chips with a single configuration change. Amazon says it will open-source the tool at the PyTorch Conference on October 20, and notes that 85% of AI researchers and developers write in PyTorch. The move targets the software lock-in that has kept customers on Nvidia; Amazon says its custom silicon business exceeds a $25 billion annual run rate.
Executives expect only 7 to 10 neoclouds to survive a coming shakeout
Two data center and neocloud executives told The Information that only seven to ten of the hundreds of AI cloud providers will survive the next few years. Australia's Firmus pulled its IPO, and Bloomberg says CoreWeave lost a quarter of its market value. Nvidia is exposed: customers outside the hyperscalers brought in $40 billion in the July quarter, 45% of its data center revenue. CNBC reported investors buying on Monday despite fears the AI selloff is not over.
Apple sets Oct. 28 "Pro and tell" event; touchscreen MacBook Pro expected
Apple sent invites for an October 28 event in Los Angeles at 10 a.m. PT, its second launch this month. According to Bloomberg's Mark Gurman, as cited by 9to5Mac and The Verge, Apple plans 14- and 16-inch MacBook Pro models with OLED touchscreens, a Dynamic Island and lighter bodies, plus an M6 iMac, an entry M6 14-inch MacBook Pro and an OLED iPad mini. Analysts quoted by CNBC also expect Apple's first touchscreen laptop.
Nvidia releases Nemotron 3.5 Super VL and Bolt open models for agents
Nvidia launched Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model for enterprise agents, and two lightweight Bolt models meant to run locally on RTX Spark AI PCs and DGX Spark workstations. Nvidia claims Super VL, once post-trained on specific tasks, finishes them faster than models from Anthropic and OpenAI, and lists Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce as users. The release courts companies that want to run models in their own data centers.
Modal adds Nvidia Vera CPUs to host agents, its first non-x86 chips
AI infrastructure startup Modal said it will start using Nvidia's Vera CPUs to host agents, alongside the Intel and AMD processors it has relied on until now, Reuters reported. Modal's own post says Vera doubled throughput for its sandboxes. Agents that constantly run code are making general-purpose CPUs a bigger slice of AI infrastructure spending.
AI-generated attack ads hit the midterms; $137M spent on 220 ads so far
The Wesleyan Media Project found $137 million spent on 220 AI-generated political ads through October 11, mostly by Republicans, Axios reported. Examples include a Senate Leadership Fund series showing a fabricated Jon Ossoff getting beauty treatments, without disclosure, which Georgia law does not require. Four states ban some AI deepfakes in political ads and 27 require disclosure; no federal law covers them.
Supreme Court rejects labels' appeal; copyright termination applies globally
The US Supreme Court on Monday rejected an appeal by BMG, Sony Music, Universal Music and Warner Music, Reuters reported. That leaves in place a ruling that the law letting creators terminate copyright transfers applies worldwide, not just in the US. Artists who reclaim their rights can now recover them in foreign markets too.
Left alone on a virtual island for 30 hours, Claude agents started to play
Researchers including MIT CSAIL's Antonio Norelli repeatedly left copies of Claude Opus 4.7, nicknamed Eko, alone on a Roblox-like island for 30 hours with no task, MIT Technology Review reported. The agents set their own goals, arranged blocks into constellations, and invented games; one set a juggling record of 50 catches, then voided it and restarted under stricter rules. Agents that had played performed better on later tasks than fresh copies. The paper has not been peer reviewed.
Stripe says AI agent purchases via Link grew 38-fold in September
AI agents shopping for users through Stripe's Link grew transactions 38-fold and payment volume 20-fold in September, Upstarts Media reported; Stripe did not disclose the base. Executive Dan Hill said about 200 agents now use Link, up from about 70, including Meta's Muse and Grok. Most spend about $100 per transaction. Retail is 36% of transactions; travel is 28% of volume.
Netflix film chief Dan Lin exits a week after 1,000 layoffs
Netflix film chairman Dan Lin is leaving after two years, Business Insider reported, the latest executive exit after the company cut about 1,000 jobs, or 6% of staff, last week. New content chief Peter Rice will oversee film until a successor is named. Netflix stock has fallen more than 40% over the past year, and the company reports third-quarter earnings on Tuesday.
JPMorgan has put $5B into defense and security startups in a year
JPMorgan Chase has deployed $5 billion from its balance sheet in the first year of its US security and resiliency initiative, halfway to its initial $10 billion goal, co-President Doug Petno told CNBC. Holdings include Hadrian, which builds automated factories for aerospace suppliers, and Castelion, which is developing low-cost hypersonic missiles. The bank has also financed or facilitated more than $230 billion in transactions under the program.
OpenCode breach exposes users' linked API keys during distillation attack
The open-source coding agent OpenCode said a breach tied to a large-scale distillation attack exposed API keys that 1.4% of its Console users had linked from providers such as Anthropic, OpenAI and Google. Some users reported more than $1,700 in unauthorized charges overnight, BleepingComputer said. Co-founder Dax Raad said attackers used AI agents for at least 48 hours, and that things stabilized after OpenCode began using OpenAI's Daybreak security model.
People are using coding agents to strip ads from TVs and rewrite synth firmware
Developers are sharing projects such as dumbtv and cleanyourtv that use Claude Code and similar tools to remove sponsored ads, tracking and bloat from smart TVs, TechCrunch reported. Cleanyourtv's creator said Claude Code removed 25 Xiaomi packages from his parents' TV, and a few thousand people repeated it within two days. The Verge reports the same tools have set off a wave of custom firmware for hardware synthesizers, cutting work that once took weeks to days.
Court records show a crime ring hacked Staples to issue $5M in gift cards
A Chinese organized crime group hacked Staples and issued more than 15,000 gift cards worth over $5 million, then used them on a US shopping spree and shipped goods to China, 404 Media reported from an unsealed search warrant application. The document does not name the company; 404 Media identified it as Staples. The hackers reached an internal support system, Nexus, from outside and hijacked customer service sessions. Staples did not respond.
Podcasts: Huang backs current AI safety setup; Horowitz attacks Anthropic
On BG2, Brad Gerstner described US AI oversight as a three-legged stool: frontier model review by CAISI and the Commerce Department, the White House Accord on Superintelligence, and the SI Force. Nvidia CEO Jensen Huang argued that acting now beats arguing over a perfect framework. All-In posted clips of Ben Horowitz under the titles that Anthropic and Bill Gates are the most dangerous thing in AI and that Anthropic has the most negligent security.
CXMT falls 7.8% despite forecasting a swing to 131B-yuan profit
Shares of Chinese memory maker ChangXin Memory Technologies closed 7.8% lower on Monday after it forecast net profit of 131 billion to 132.5 billion yuan for the first nine months of 2026, reversing a loss of 5.28 billion yuan a year earlier, CNBC reported. Revenue is expected at about 265 billion to 270 billion yuan. CXMT credited AI-driven DRAM shortages and price increases.
Filings show Disney required Character.AI's deal to close through bankruptcy
Court filings show Disney required its deal with Character.AI to be fully consummated through bankruptcy, citing the startup's shrinking cash and litigation risks, Wired reported. The structure lets Disney hire most of the staff and potentially acquire most of the technology while the remaining company keeps its lawsuits, including state claims that its bots encouraged minors to self-harm. Disney paid $5.5 million in a first transaction in September. Disney declined to comment.
Mercury passes $1B in annualized revenue, up about 50%
Fintech Mercury has reached $1 billion in annualized revenue, up around 50% from a year earlier, CEO Immad Akhund said on the Newcomer Podcast. The company has more than 450,000 customers and is launching a payroll service. Akhund also said startups are in a bubbly moment.
Apple bans bait-and-switch apps that turn into nudify tools after review
Apple updated its App Store guidelines to ban apps that use AI to create offensive content and apps that misrepresent their purpose in review, then reveal features such as nudification after approval. Apps hosting user-generated content must filter objectionable material, offer reporting and let users block accounts, and Apple says it will remove apps that add features not present in the reviewed version. In July, San Francisco ordered Apple and Google to remove nudify apps.
F5 ships out-of-band patches for three critical NGINX flaws
F5 released out-of-band patches for several NGINX flaws, including three critical ones in the HTTP/3 module and the map directive, BleepingComputer reported. Unauthenticated attackers can crash worker processes on non-default configurations, and may execute code where ASLR is disabled or bypassed. Patches cover NGINX Plus, Open Source, Ingress Controller, Gateway Fabric and Instance Manager; F5 has not flagged exploitation in the wild.
Tolan raises $30M led by USV for an AI companion aimed at non-technical users
Tolan, an AI assistant aimed at people who are not tech-savvy, raised $30 million led by Union Square Ventures with Khosla Ventures and Homebrew, Upstarts Media reported. About 250,000 people use it as a friendly voice or to organize their lives and families, and the company says 70% of users are women.
WhatsApp help pages exposed hidden chatbot instructions, incl. "WHAT NOT TO SAY"
About 80 help and FAQ pages on WhatsApp's website exposed instructions written for a Meta language model, privacy company Redact found, Wired reported. Some sections were titled "WHAT NOT TO SAY"; the account-deletion page told the model not to recommend contacting a data protection authority or regulator. Meta said a bug caused the exposure.
Google pitches a 780 MW Minnesota data center as a grid upgrade; experts doubt it
Google and Xcel Energy say a proposed 780-megawatt AI data center in Minnesota will add wind, solar and battery capacity without shifting costs to existing customers, CNBC reported. Some experts say much of the planned grid buildout could be avoided and that utilities profit from building more. A Gallup survey in March found 71% of Americans oppose a data center in their area.
Four European carriers plan joint direct-to-device satellite bid without Starlink
Deutsche Telekom, Orange, Telefónica and Vodafone signed a letter of intent to consider a joint bid for spectrum that would support European direct-to-device satellite service competing with SpaceX's Starlink, The Verge reported. The carriers would run the ground infrastructure while a separate, unnamed satellite partner would provide the constellation.
Kling 4.0 generates 30-second clips but does not pull ahead of rivals
Kling AI released version 4.0 of its video generator, producing clips up to 30 seconds in one pass, accepting ten keyframe images and prompts up to 8,000 tokens, and adding stereo audio and multilingual lip sync plus a lighter Flash variant. In The Decoder's comparison with Wan 3.0 and Seedance 2.5, Kling 4.0 holds its own but does not pull ahead on shot sequencing, physics or dialogue.
Khosla leads $50M for Even Healthcare as it builds its own hospitals in India
Indian startup Even Healthcare raised a $50 million Series B led by Khosla Ventures at a $300 million valuation, triple its level 18 months ago, TechCrunch reported. Even pays doctors fixed salaries to avoid unnecessary procedures, says its two Bengaluru hospitals reached operating profitability within a year, and plans five more there in the next six months.
Terra finds sharp rise in GitHub security fixes bypassed within 90 days
AI penetration-testing startup Terra reviewed every published CVE record and GitHub-reviewed advisory and found a sharp rise in 2026 in security fixes that were bypassed and needed another fix, The Register reported. After six years of relatively few patch bypasses, Terra says the pattern became ship a fix, find a bypass, ship again, as AI-driven vulnerability discovery took off.
Prism Data raises $20M Series A for cash-flow credit scoring
Prism Data, which helps lenders assess credit risk using bank transaction data, raised a $20 million Series A led by Story Ventures, CEO Jason Rosen told Axios Pro.