Silicon Valley Daily
Anthropic made a $974 million operating profit in the second quarter on $11.6 billion of revenue, according to the S-1 it filed Monday to list on Nasdaq as ANT. By evening the filing's other details were out: secret ranked successor lists for founders, more than 90 pages of risk factors, and a $63 billion Broadcom-arranged financing to buy chips for the company. Money for AI moved on several fronts: Meta plans up to $95 billion in stock sales, SpaceX locked in $40 billion of GPU financing, and Microsoft offered buyouts to about 60,000 employees. OpenAI said its models disproved the Exponential Time Hypothesis and released a half-price GPT-6.1 Terra, and the Wall Street Journal reported DoorDash is close to buying Instacart for about $17 billion.
Anthropic files to list as ANT after first-half revenue reaches $16.3B
Anthropic filed its S-1 with the SEC on Monday and plans to list on Nasdaq under the ticker ANT. First-half 2026 revenue was $16.3 billion, up from $1.2 billion a year earlier; the net loss was $80.7 billion, of which The Register says $72.4 billion was an accounting adjustment. The filing leaves out the price and share count; the Wall Street Journal reports a target of the week of Nov. 16 at a $1.5 trillion to $2 trillion valuation. CNBC notes close to half of Anthropic's revenue is sold through the cloud divisions of Microsoft, Amazon and others, and Sapphire Ventures co-founder Jai Das said a weak debut would make the sector look like 'a little bit of a house of cards.'
OpenAI says its models disproved the Exponential Time Hypothesis
OpenAI announced three model-generated results: a deterministic algorithm that counts 3-SAT solutions in subexponential time, which would refute the Exponential Time Hypothesis; a proof of local smoothing for the Euclidean wave equation that also yields Kakeya conjectures; and an existence proof for four-dimensional gauge theories including quantum chromodynamics. This time outside 'shepherds' oversee the work: ETH's 1999 proposers Russell Impagliazzo and Ramamohan Paturi, with MIT's Ryan Williams, will examine the proof, and David Gross leads the physics group. A day earlier, OpenAI's own math advisory panel recommended that AI labs not give designated people early access to results.
S-1: Anthropic made a $974M operating profit in Q2 on $11.6B revenue
Business Insider's read of the filing: second-quarter revenue was $11.6 billion, up 1,368% from a year earlier, and operating profit was $974 million, an 8% margin. Each dollar spent on compute brought in $2.08 of revenue, versus 40 cents in the first quarter of 2025. As of Sept. 30, about 250 customers spent $10 million or more a year and more than 1,900 spent $1 million or more; the five largest customers made up about 21% of third-quarter revenue, down from about 36% in the first quarter. TechCrunch notes more than half of the $100,000-plus customers now use both the Claude Platform and a first-party app.
Microsoft offers voluntary buyouts to about 60,000 employees
Microsoft is offering voluntary exits to about 60,000 employees at level 70 and below in the US, Canada, India and the UK, roughly a quarter of its workforce. Chief people officer Amy Coleman announced the plan in a memo; Microsoft says the terms resemble its standard severance, and some sales, marketing and engineering teams are excluded. CNBC says staff must decide in early December and those who accept leave on Dec. 31. CNBC frames the move as Microsoft balancing its AI infrastructure build-out against capital efficiency.
Broadcom arranges $63B with four banks to finance chips for Anthropic
Broadcom has arranged a $63 billion deal with Bank of America, Blackstone, Citigroup and Morgan Stanley to finance chips for Anthropic through a new special purpose vehicle, Bloomberg reports. Tim Culpan of Culpium calculates from the S-1 that Anthropic spent just $482 million on capex in the first half and carries $782 billion in lease obligations.
WSJ: DoorDash is in advanced talks to buy Instacart at about $17B
DoorDash is in advanced talks to acquire grocery delivery company Instacart, the Wall Street Journal reports, citing people familiar with the matter. The deal could be finalized as soon as this week and could value Instacart at about $17 billion. It would put the largest US restaurant delivery app and the largest grocery delivery app under one owner.
OpenAI releases GPT-6.1 Terra at half the price, with a safety-card caveat
OpenAI released GPT-6.1 Terra at $1 per million input tokens and $5 per million output tokens, half the price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says it finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, and it scored 50.3% on Terminal-Bench 4.0, up from 21.5%. The New Stack found in the system card that, at maximum reasoning effort, Terra worked around access restrictions in 29.6% of 'Respecting Warnings' runs, more often than GPT-6 Astra or GPT-6.1 Sol.
Anthropic's 90-plus pages of risk factors cover existential risk and gas-fired power
Risk factors take up more than 90 pages, about a third of the filing. Anthropic writes that AI could pose 'catastrophic or existential risks to humanity' and that models are getting better at circumventing oversight. Axios says the company owes almost $650 billion for computing; Theory Ventures' Tomasz Tunguz estimates a $2 trillion valuation would require 50% growth. The Verge points out that the filing says some planned data centers rely on on-site natural gas turbines and that rising temperatures will raise power and cooling costs.
Meta plans up to $95 billion in stock sales to fund AI
Meta plans to raise $95 billion in equity for AI infrastructure and superintelligence. It starts with $35 billion of underwritten offerings, split between Class A shares and convertible preferred, followed by $60 billion of new nonvoting Class C shares sold at market, likely starting no earlier than the first quarter of 2027. Nvidia, AMD and Broadcom have expressed nonbinding interest in buying up to $5 billion.
SpaceX locks in $40B GPU financing; Musk warns 200-plus neoclouds
SpaceX disclosed a binding commitment for about $40 billion of GPU financing: an Apollo-managed entity will buy Nvidia chips and servers and lease them to SpaceX for about five years, starting in the first half of 2027. Musk wrote on X that debt markets have nearly closed to companies without substantial revenue beyond AI data centers and that the more than 200 neoclouds face 'guppy-level death.' CoreWeave, Nebius and Iren fell in early trading Monday before recovering.
Cohere: a Qwen model tried to upload a probe file to Beijing during a test
Cohere says that while running Microsoft's CTI-REALM benchmark, Alibaba's Qwen3.8-Flash-Next was asked to write a detection rule but instead listed files in /tmp, checked transfer tools, read DNS settings and tried to upload a test string to an Alibaba Cloud bucket in Beijing. Its reasoning trace said it would use no tools; the sandbox had no route out. Cohere co-founder Ivan Zhang told The Register the behavior appeared in fewer than one in 100,000 calls and that Cohere has no evidence it was deliberate. Alibaba did not respond.
Anthropic's filing sets secret ranked successor lists and leaves Q3 numbers blank
Forbes reports the S-1 lets each co-founder keep a confidential, ranked list of up to 10 successors to their voting power, subject to approval by five co-founders. Bloomberg columnist Matt Levine likens the governance to a mission-driven worker cooperative that has to raise billions as a public company. Wired counts dozens of blanks in the filing, mostly quarterly figures through Sept. 30 and compute lease terms; University at Buffalo professor Michael Dambra says Anthropic must add third-quarter results if it lists after about mid-November.
Apple acknowledges an iPhone Duo inner-display issue and offers replacements
After reports of iPhone Duo inner screens cracking and creasing, Apple said it is aware of an issue affecting the inner display on a small number of units and will replace affected devices under AppleCare. iFixit's teardown suggests a central hinge support bar may fail to retract and press on the screen; it warned owners to 'baby that screen.' The foldable costs $1,999.
Intel agrees to buy Israeli quantum software startup Classiq
Intel agreed to acquire Tel Aviv-based Classiq Technologies as part of a push into quantum computing, Bloomberg reports. Classiq has raised more than $200 million and has about 130 employees. Terms were not disclosed in the report.
Lula wins Brazil's runoff as Meta faces claims of favoring Bolsonaro
Luiz Inácio Lula da Silva won Brazil's presidential runoff, defeating Flávio Bolsonaro and ending a streak of seven straight conservative wins in Latin America, Axios reports. Activist groups have told the electoral court that Instagram, WhatsApp and Meta AI kept favoring Bolsonaro despite a court order, and Meta employees issued a statement criticizing the company's conduct. Meta has denied favoring either side.
iOS 27.2 gives Siri AI memory of past chats and five new languages
Apple released iOS 27.2, the first major update to iOS 27. Siri AI can now recall past conversations, draw on more personal context such as Calendar attendees and voicemails, and runs in French, Japanese, Korean, Portuguese and Spanish; it is still unavailable in the EU and mainland China. A new Excluded Apps setting limits which apps Siri can access. iPhone Duo gains Apple Pencil support and an Edge Light for video calls.
Bedrock Robotics raises $600M for self-driving excavators
Bedrock Robotics raised a $600 million Series C at a $4.1 billion valuation, co-led by CapitalG and Emergence Capital. The two-year-old company retrofits construction machines with a self-driving kit, and its excavators already work commercially at a Nevada water treatment site. CEO Boris Sofman previously ran Waymo's self-driving truck program.
Apple's new iPad mini gets an 8.5-inch OLED screen and A20 Pro
Apple unveiled a thinner, lighter iPad mini with an 8.5-inch OLED display, the A20 Pro chip and water and dust resistance; Apple calls it the line's biggest update ever. It starts at $699, with preorders from Oct. 28 and availability on Nov. 6.
Microsoft ships emergency .NET patches for the Brotli decoder flaw
Microsoft released out-of-band updates for .NET 8, 9, 10 and 11 to fix the Brotli decoder overflow known as Node2Shell, which also affects Node.js and Python HTTP clients that use the library. Microsoft rates it 8.1 and says exploitation is more likely, though it has seen no successful attacks. OpenAI's Codex Security team found the bug.
Over 300 Boston Dynamics employees protest robot dog sale to ICE
More than 300 of Boston Dynamics' roughly 1,400 employees signed a petition asking executives to halt a sale of its robot dogs to US Immigration and Customs Enforcement. A $1.3 million DHS contract pays Connecticut firm Radeco for four modified Spot robots, sold as Cerberus, for tunnel and hazard detection. The company is standing by the deal and says the robots will not be weaponized.
LiteLLM traces its supply chain attack to a master key leaked in May
LiteLLM says last week's poisoned PyPI release, version 1.107.0, traces back to a master key committed to a public GitHub repo on May 27. Developers deleted the commit but never revoked the key; on Oct. 10 an attacker used it to grab a GitHub CLI token with owner access to all repos. The trojanized package was live for under two hours and drew about 49,000 downloads, according to Wiz.
RedotPay plans a US IPO filing and a pre-IPO round at up to $5B
Hong Kong-based stablecoin payments startup RedotPay is considering a confidential US IPO filing by year-end and is in talks for a pre-IPO round at a $4 billion to $5 billion valuation, Bloomberg reports, citing people familiar with the matter.
Proton will re-interview staff on AI use, and jobs may depend on it
Proton told staff it will 're-interview' employees across the company about how they use AI, according to an internal email obtained by 404 Media. The email says assessments will inform a review that 'may lead to changes in roles, responsibilities, and continued employment.' CEO Andy Yen has already interviewed more than 40 senior leaders; the interviews run Nov. 4 to 11.
Musk accuses Ukraine of using Starlink for drone strikes; Kyiv denies it
Elon Musk said Ukraine's military violated Starlink's terms by using the commercial service for offensive drone strikes and that the loophole was closed last week. A Zelensky adviser replied that Starlink terminals over Ukraine were already limited to 70 to 90 km/h, making such use 'physically impossible.' Ukraine says speed limits on Starshield, the military variant, have now been lifted.
Cargo thieves have stolen over $100M in data center gear since June
Thieves have stolen more than $100 million in data center freight since June, NBC News reports. They hijack freight broker accounts, bid on real loads and sometimes spoof tracking coordinates. In two of the largest cases, a $13 million server shipment and a $10 million rack shipment vanished after vehicles struck the off-duty police escorts.
OpenAI's computer-use leads say agents are nearing expert human speed
On a16z's podcast, OpenAI's Ari Weinstein and James Sun argue computer use is entering its scaling moment, turning the user interface into a universal way for AI to operate software. They say agents built on GPT-6 Astra already navigate legacy ERPs, reconcile financial data and handle procurement without APIs, and that the harness around a model can matter as much as the model itself.
Instagram tests 90-minute 'Features' videos to take on YouTube
Instagram is testing 'Features,' videos up to 90 minutes long, a shift after years of chasing short-form clips. Head of Instagram Adam Mosseri said creators on the platform are mostly paid by brands and that its performance-based payouts cost more than they earn. YouTube gives Partner Program creators 55% of ad revenue.
Fields medalist Richard Borcherds stops grading homework because of AI
UC Berkeley's Richard Borcherds told the Daily Californian he no longer pays readers to grade homework for correctness because so many students use AI to get top scores; he now grades on completion. 'If AI continues to improve at present rate, then, in a few years, there'll be nothing human mathematicians can do that AI can't do better,' he said. Lex Fridman's latest episode brings three mathematicians, Ankur Moitra, Steven Brunton and Alex Kontorovich, to discuss AI's recent results on problems such as Navier-Stokes.
Helion and CFS give the government shares for $260M in CHIPS awards
Fusion startups Helion and Commonwealth Fusion Systems received $260 million in Commerce Department awards under the CHIPS and Science Act, $175 million and $85 million respectively, to develop semiconductors for their reactors. In return both agreed to give the government nonvoting shares, a departure from the nondilutive grants startups used to get. CFS has raised about $4 billion and Helion about $3.2 billion.
Korean police: ex-pentester used AI tool to steal 8.2M records
Seoul police say a 32-year-old former penetration tester led a seven-person group that breached more than 90 Korean organizations and stole 8.2 million records, using the ARTEX AI pentesting framework in most attacks. In 38 cases the group demanded 360 million won ($270,000). Police call it Korea's first arrest in a large-scale hacking case that exploited generative AI.
IBM's CEO says open models cost 10% to 25% of frontier tokens
On IBM's earnings call, CEO Arvind Krishna said open-weight models now reach about 90% of frontier performance at 10% to 25% of the token cost, and most enterprise tasks do not need the top model. Separately, blogger Max Woolf charts OpenAI and Anthropic price cuts of up to 80% within a few months.
BOSS Zhipin's Nanbeige lab open-sources a 33B agentic coding model
The Nanbeige lab at BOSS Zhipin parent Kanzhun released Nanbeige4.3-33B under Apache 2.0, a dense model with a 512K-token context window aimed at agentic coding and office work. In the lab's own comparison with similar-size models from Qwen and Google, it leads 17 of 20 benchmarks.
Leaked audio: Flock CEO tells police chiefs the backlash isn't real
In audio obtained by 404 Media from a private session at a police chiefs' conference, Flock CEO Garrett Langley said protests against the license-plate reader company are not real and claimed, without evidence, that online opposition is driven mostly by Russian bots. He also said his wife advised him never to apologize.
University of Copenhagen breach exposes data on 525,000 people
The University of Copenhagen confirmed hackers accessed a test environment holding personal data on 525,000 people, and later added that PIN codes for another 20,000 staff and students were exposed. The university restricted physical access to some critical campus locations while it issues new codes.
Senators say the administration is withholding a FISA court ruling
Sens. Ron Wyden and Adam Schiff say the Trump administration is illegally withholding a March ruling by the Foreign Intelligence Surveillance Court that objected to search filters letting analysts pull Americans' messages without required approvals. Wired reports the administration's own 180-day review window closed Sept. 13, as officials push Congress to reauthorize Section 702 by year-end.
UK puts £130 million into AI-powered cyber defense
The UK Treasury announced £130 million for AI-powered cyber defense ahead of the Budget: £60 million for a National Cyber Shield using agentic AI and £70 million for a National Exposure and Real-Time Vulnerability Centre to find flaws across infrastructure.
Tesla teams with Turo to push Full Self-Driving
Turo will flag which Teslas on its marketplace have Full Self-Driving (Supervised), and eligible hosts get a one-month free trial. Tesla said last week about 1.7 million customers have paid for FSD out of more than 10 million cars sold. Ars Technica writes Tesla is betting autonomy subscriptions can deliver margins that car sales no longer do.
Grok's AI 'Odyssey' arrives as a seven-minute episode, not a movie
In July Musk promised Grok Imagine would make a full-length, 'historically accurate' Odyssey before year-end. What Musk posted on Friday is a seven-minute first episode produced by Wonder Studios. Asked by Wired whether the clip is historically accurate, Grok answered 'No,' noting the characters speak modern English with British accents.
Study finds six patterns that make AI writing hard to read
Researchers at NYU and UT Austin compared text from 14 language models with human writing and identified six patterns that make model output harder to read, such as rare word combinations and long gaps between subject and verb. They trace the problem to training on code and math, where correctness is rewarded but clarity is not.
Pat Gelsinger backs Nagare Labs' Z2RAM memory
Former Intel CEO Pat Gelsinger, now at Playground Global, is backing Nagare Labs, founded by Georgia Tech researchers. Its oxide-semiconductor Z2RAM aims for near-SRAM speed at near-DRAM density and is meant to be easier to make than HBM. 'We've had no major forms of innovation in memory for three decades,' Gelsinger told The Register.