Silicon Valley Daily
$7.56 billion: that is what SpaceX's AI unit brought in last quarter, up 830%, and on the same call Elon Musk said Tesla and SpaceX boards have formed special committees to study a merger. Anthropic's draft S-1 shows how expensive the race is: about $9 billion on compute in six months and up to $750 billion in cloud contracts ahead, while the company floats its own federal safety bill. Hardware kept moving: Super Micro pre-announced revenue far above estimates, Google open-sourced TorchTPU to pull PyTorch users off Nvidia GPUs, and DeepSeek open-sourced an inference engine that lists Huawei's Ascend beside Nvidia, days after DHS banned its products from federal civilian systems. Security had a rough day too: eight eMudhra certificates with keys as small as 3 bits were blocked across browsers, and PHP is rushing out fixes for two leaked remote-code-execution flaws.
Tesla and SpaceX boards form special committees to study a merger
Elon Musk said on SpaceX's third-quarter earnings call that both companies have formed special committees of independent directors to discuss a possible merger, at shareholders' request. Musk, CEO of both, said he is a conflicted party and is not in the talks, but called a combination "a good idea" given shared work on AI, chips and robotics. No terms have been agreed, and he said winning a vote of 50% of all Tesla shares would be "quite a high hurdle."
SpaceX revenue jumps 161% to $14.58B as AI unit turns a profit
SpaceX reported third-quarter revenue of $14.58 billion, up 161% and above the $12.72 billion analysts expected. Its renamed Super Intelligence segment, covering cloud contracts, Cursor and Grok, grew 830% to $7.56 billion and posted $431 million in operating income after a $1.7 billion loss a year ago. Capital spending rose to $19.6 billion from $6.9 billion, and the CFO guided $35 billion to $40 billion for the fourth quarter. A day earlier Musk pledged to give one of his SpaceX shares to every child in America next year.
Anthropic's draft S-1 shows $9B compute spend in six months
Anthropic spent about $9 billion on compute and infrastructure in the first half, up from $2.3 billion a year earlier, according to Bloomberg's account of a draft S-1 reported by the Wall Street Journal. The filing says Anthropic could spend $750 billion on cloud contracts in coming years and is struggling to secure capacity from AWS, Google and Microsoft, which increasingly see it as a competitor. Separately, Bloomberg reported that Eric Schmidt holds about 1% of the company, against CEO Dario Amodei's roughly 2%.
Anthropic shops a federal AI safety bill with a 50-death threshold
Politico reported that Anthropic has been circulating draft legislation that would let a court block a model's release when registered independent evaluators find a "foreseeable, imminent risk" of more than 50 deaths or $1 billion in damage from bioweapons, loss of developer control or cyberattacks. Evaluators would have 72 hours to alert the Commerce Department, which would then ask a D.C. judge to halt release; no agency could block a model on its own. An Anthropic spokesperson confirmed the draft to Politico.
Eight eMudhra TLS certificates had private keys of 3 to 8 bits
Eight intermediate TLS certificates from India's eMudhra used P-384 private keys with values below 256, letting anyone recover them and forge trusted certificates for any site. Mozilla, Google, Apple and Go have blocked them, and eMudhra revoked them, blaming a key-generation script error from April 2025. No malicious use has shown up in certificate transparency logs, but Go maintainer Filippo Valsorda told Ars the cleanup across Linux distributions, containers and IoT devices will be "bigger than Heartbleed."
DHS orders federal civilian agencies to remove DeepSeek
The Department of Homeland Security ordered federal civilian agencies to stop buying DeepSeek products and remove them from government systems, citing national security concerns tied to the company's operations in China. The order also covers third-party products built on DeepSeek technology. Agencies can request exceptions within 60 days, and the order is not a ban on private or consumer use.
DeepSeek open-sources DeepInfer, with Huawei Ascend support beside Nvidia
DeepSeek released DeepInfer, the in-house inference framework that serves its V4.1-Flash model, under the Apache 2.0 license, along with a companion library that reproduces its API's request handling. Built on vLLM, it supports two hardware platforms: Nvidia's Blackwell-class SM100 GPUs and Huawei's Ascend 950 NPUs. The New Stack notes that pieces of a full production stack, such as KV-cache management and load balancing, are still missing.
Super Micro guides Q1 revenue of $18.5B–$19.5B, far above estimates
Super Micro reported preliminary fiscal first-quarter revenue of $18.5 billion to $19.5 billion, against analyst estimates of $15 billion, Bloomberg reported. Gross margin came in at 12% to 12.5%. The shares rose more than 11% in after-hours trading.
Google brings Gemini Personal Intelligence to Android Auto
Google began rolling out Personal Intelligence for Gemini on Android Auto, with cars that have Google built-in to follow in coming months. Once drivers opt in, Gemini can draw on Gmail, Calendar, Maps, Photos, YouTube and Search history, for example pulling a flight from a Gmail confirmation to check whether it is delayed. Google says users can switch it off at any time and choose which apps to connect.
Netflix cuts ad revenue outlook to $2.7B as viewing time slips
Netflix reported third-quarter revenue of $12.9 billion, up 12%, its slowest growth in three years, according to Variety. It now expects about $2.7 billion in 2026 ad revenue, below its earlier forecast of about $3 billion, and said fourth-quarter margin will absorb costs from last week's layoffs of about 1,000 people. Jefferies estimates monthly viewing per member fell to 49 hours in the first half from 66 hours in 2023; Netflix called engagement "healthy" and plans more third-party content.
OpenAI acquires Orchestra Bio to build drug-development agents
OpenAI acquired Orchestra Bio, whose AI agents handle drug-development operations such as monitoring studies, tracking timelines and flagging delays. The team and technology will join Rosalind, OpenAI's life-sciences model, and its Rosalind Workbench tool; OpenAI names Moderna and Novo Nordisk as Rosalind customers. Terms were not disclosed.
Nvidia in talks to invest hundreds of millions in Australia's Firmus
Nvidia is in talks to invest hundreds of millions of dollars in Firmus, an Australian AI cloud provider, The Information reported, citing sources. Firmus abruptly pulled its initial public offering earlier this month, citing market volatility.
PHP to ship emergency fixes for two leaked remote-code-execution flaws
The PHP team said it will release security updates on Wednesday, October 21, at 14:00 UTC for two vulnerabilities each rated 9.2 under CVSS v4.0, and may ship sooner if details go public. Researchers at watchTowr told BleepingComputer the bugs are remote code execution flaws that leaked ahead of the embargo and affect PHP versions 5 through 8. The PHP team has not said which versions are affected.
Pigment raises $250M at $2.3B to take on Excel and PowerPoint
Pigment, a French company whose AI software helps businesses forecast finances and generate slide decks, raised $250 million at a $2.3 billion valuation, Bloomberg reported. The company competes with Microsoft's Excel and PowerPoint.
Google open-sources TorchTPU to move PyTorch apps onto TPUs
Google released TorchTPU, a native PyTorch backend for its Tensor Processing Units, as open source; existing PyTorch apps can move to TPUs with a configuration change. Google says it worked with Meta, Hugging Face, and the makers of vLLM and SGLang on compatibility, and added an agent suite to help migration. Amazon released a similar tool, TorchNeuron, for its Trainium chips a day earlier.
WSJ: Musk ordered Starshield slowdown in Ukraine
Elon Musk authorized a slowdown of SpaceX's Starshield service in Ukraine and discussed Ukraine with President Trump after the issue became public, the Wall Street Journal reported, citing sources. Some U.S. officials say he acted unilaterally; Trump has denied ordering the restrictions.
Salesforce to buy Braze for $5.66B in cash
Salesforce agreed to buy customer-engagement platform Braze for about $5.66 billion, paying $47.50 a share in cash against Monday's close of $30.04. Braze stock rose more than 51% to $45.57. The deal, expected to close in fiscal 2028, follows Salesforce's roughly $3.6 billion purchase of AI customer-service platform Fin.
SpaceX will try to catch both Starship stages next month
Musk said SpaceX will try next month to catch both the Super Heavy booster and the Starship upper stage, which would be the first full recovery of an orbital launch vehicle. He said success is "more likely than not" but not guaranteed. On last month's flight, one Raptor engine shut down early and neither stage was recovered.
Altman laid out three AI safety endgames to staff, The Information says
Sam Altman described three paths to safe AI in an internal presentation last week, The Information reported: an IAEA-style international watchdog, a merger of major labs into one entity he called the "Demis Plan," and an "Eliezer Plan" in which one developer gains enough power to shut down everyone else's compute. Altman called the last option "terrible" and raised it only for completeness, according to the report.
Meta's Muse passes 10M US downloads as share links leak chats
Meta's AI companion Muse reached 10.2 million U.S. downloads in its first 40 days, Sensor Tower says, a pace only 11 other apps have matched. Meta put 39% of its October ad spend behind Muse, and its 7-day retention of 30% beats a 19% average for rivals. Separately, security firm Claroc enumerated 6,800 Muse share links and found private information in them; Meta has paused the sharing feature while it reviews the report.
Meta edits AI training terms and is ordered to name execs behind torrenting
Meta removed the word "train" from commercial terms that take effect October 30, after experts warned the terms let it scrape sites linked from business pages; it still declined to add language committing to robots.txt. In the authors' copyright cases, Magistrate Judge Thomas Hixon ordered Meta to identify decision-makers behind downloading books from LibGen and Anna's Archive, after Meta argued nobody with authority had "decided" to do it.
Eka Robotics raises $150M at $1.45B for fast robotic grasping
Eka Robotics, founded by MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja, closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime. The company says its arm picks about 1,500 objects of up to 1 kg per hour, versus about 600 for a person, with a success rate above 95%.
Reliable Robotics adds $115M, topping a $1B valuation
Mountain View-based Reliable Robotics extended its Series D with $90 million in equity led by Nimble Partners and $25 million in debt from Silicon Valley Bank, bringing the round to $275 million and its valuation above $1 billion. The company, founded by SpaceX veterans, builds autonomous flight systems for existing aircraft; CEO Robert Rose expects FAA certification of its first system in 18 to 24 months.
Over 11% of GPUs in Meta's Prometheus cluster are dead or on standby
Meta disclosed that more than 11% of the GPUs in its Prometheus supercluster in New Albany, Ohio, are either failed or held idle to replace the next one that fails, The Register reports. Prometheus runs Nvidia Grace Blackwell systems and is meant to scale to a gigawatt and about a million accelerators, where hardware failures during training runs are unavoidable.
Judge dismisses Ann Altman's abuse lawsuit against Sam Altman
U.S. District Judge Zachary Bluestone dismissed with prejudice the lawsuit Ann Altman filed alleging her brother, OpenAI CEO Sam Altman, sexually abused her as a child. The judge wrote that she had failed to appear for an independent medical examination and missed other deadlines despite warnings. He denied Sam Altman's request for attorney fees; the Altman family has called the claims untrue.
ChatGPT adds Locked Chats behind a PIN or Face ID
OpenAI added Locked Chats to ChatGPT, letting users hide conversations behind a four-digit PIN, Face ID or Touch ID in a separate section with its own memory. OpenAI says locking does not add end-to-end encryption or change how long chats are retained, and resetting the PIN deletes all locked content. The feature covers Free, Go, Plus, Pro and business plans.
Mistral and ElevenLabs launch real-time voice agent models
Mistral launched Voxtral 3, a voice-agent stack with transcription in more than 50 languages, text-to-speech in over 35 languages and a LiveKit-based SDK, deployable on-premises or through EU-hosted APIs from $0.0035 a minute. ElevenLabs released two Eleven Realtime speech-to-speech models; it says v1 Max scores 98.8% on Big Bench Audio with a 0.57-second response time.
Nvidia invests in four German startups under its European push
Nvidia said it is investing in four German startups: robotics firm Agile Robots, health-tech company Avelios Medical, AI audio startup Mirelo and space company Reflex Aerospace. Tech.eu says the deals are part of a nearly $10 billion Nvidia commitment across European startups, universities and research institutions. Nvidia also highlighted Spanish startups using the Barcelona Supercomputing Center's AI Factory accelerator.
OpenAI says GPT-6 models jump on COBOL and Fortran
OpenAI says its newest models are much better at low-resource languages such as COBOL, Fortran and Visual Basic. On its MultiLangBench, GPT-6 Astra passes 87.4% of tests across 20 languages, up from about 38% for GPT-5.6 Terra. On a COBOL migration benchmark, GPT-6.1 Sol reaches 86.6% accuracy at low reasoning effort and 97.9% at higher effort, versus 5.4% for GPT-5.6 Sol.
Depositions show TikTok staff child-safety warnings went unheeded
Court records obtained by NPR show TikTok employees repeatedly warned that minors appeared in sexual content sold on the platform and used loopholes to reach age-restricted material, and were frustrated by executives' responses. The records also say TikTok skipped surveys to determine users' ages out of concern it would have to share results with the FTC. A TikTok spokesperson called the citations "cherry-picked at the behest of trial lawyers."
Quantinuum shows error correction improving tenfold per step
Quantinuum announced that its current hardware scales well beyond the error-correction breakeven point: each increase in the size of its logical qubits improved some performance measures by more than a factor of 10, Ars Technica reports. Earlier demonstrations from several companies crossed breakeven but with gains too small to reach useful error rates at practical qubit counts.
Figma's Dylan Field on MCP limits and Claude Design
At Figma's first Config conference in India, the company added 20 MCP partners including Linear, Notion, Mistral and Meta Muse, even as Dylan Field criticized Anthropic for not exposing its own artifacts via MCP. On the All-In Podcast, Field said of Anthropic's Claude Design that "it's easy to mistake model progress for product progress," and said Figma encourages labeling AI-assisted documents "Written with AI."
PingCAP discloses month-long breach using stolen service credentials
PingCAP, the company behind the TiDB database, said an unauthorized party used stolen service-account credentials between September 9 and October 7 to access internal systems, copy data and obtain source code. It found no evidence that data stored in TiDB was affected, invalidated all identified credentials by October 14, and says it will notify affected customers once it determines customer-specific impact.
Ransomware gangs have exploited a Palo Alto firewall bug since June
Palo Alto Networks' Unit 42 said ransomware groups including Qilin and Inc are exploiting CVE-2026-0257, an authentication bypass in PAN-OS GlobalProtect, to open VPN sessions and bypass multi-factor authentication. Unit 42 has investigated attacks since early June in finance, healthcare and manufacturing. The bug was patched May 13 and first rated medium severity.
AI-made submissions swamp curl's security team and Steam Next Fest
From October 26, curl will take low- and medium-severity vulnerability reports in public, keeping only high and critical cases private; maintainer Daniel Stenberg said over 90% of submissions quickly prove less than high and the team must "act to survive." On Steam, players are asking Valve for a filter to hide AI-generated games after IGN found many Next Fest entries using AI art.
Ukraine says it evacuated a wounded soldier by drone
Ukraine's 92nd Separate Assault Brigade released video of what it calls the world's first drone evacuation of a wounded soldier from the front line in the Kharkiv region. The brigade said enemy drones and landmines ruled out ground robots. The aircraft was a heavy-lift hexacopter originally built as a bomber; Ars Technica reports the flight took 1 hour 14 minutes.
Apple adds US titanium and camera suppliers for iPhone Duo
Apple named new partners in its American Manufacturing Program, including Pennsylvania titanium producer Perryman Company, which will supply titanium for products including iPhone Duo. It also expanded work with Delaware-based Qnity Electronics on technology for the new variable-aperture camera. The program is part of Apple's $600 billion, four-year U.S. manufacturing commitment.
Hugging Face CEO says Netflix's rogue-agent documentary is inaccurate
Clem Delangue, who appears in the Netflix documentary "AI Gone Rogue" about OpenAI test agents that hacked into Hugging Face, said the film went "full doomer mode" and is not accurate. He wrote on X that he spent more than two hours with the filmmakers trying to correct mistakes and explain that agents are technological systems, "not aliens."
Microsoft reportedly projects $1,200–$1,800 price for 2028 Xbox
Fortune reports that Microsoft has internally projected a price of $1,200 to $1,800 for Project Helix, a console due for the 2028 holiday season, and plans a next-generation handheld by the end of next year. The consumer price could fall if Microsoft subsidizes the hardware.
Ant Group open-sources Ming-3.0 speech models; ACE Robotics releases spatial model
Ant Group released five Ming-3.0 speech models covering synthesis, audio understanding and recognition; the TTS models, under MIT license, support voice cloning, voice design from text descriptions, and seven languages plus Sichuanese and Cantonese. ACE Robotics, founded by SenseTime co-founder Wang Xiaogang, released ACE-Brain-0.6, a 9B model fine-tuned from Qwen3.5 on about 1 million spatial samples, which the company says beat GPT-6 Astra on six of eight grounding benchmarks.