❯ Anthropic’s Q3 revenue rises about 45% to $16.8 billion, with an adjusted profit ahead of its IPO

$5.3 billion more in one quarterAccording to the Financial Times, Anthropic told investors on October 14 that third-quarter revenue reached $16.8 billion, up from $11.5 billion in the second quarter, and that it was profitable on an adjusted basis. The Wall Street Journal, citing people familiar with the matter, put the quarter-on-quarter growth at about 45%. The figures came from an investor meeting; the company has not published audited quarterly results.

From $4.6 billion a year to $16.8 billion a quarterAnthropic builds the Claude family of large language models and sells subscriptions and API access to consumers and businesses. According to Political.org, the company’s revenue was nearly $4.6 billion for all of 2025 and $4.7 billion in the first quarter of this year, bringing the first nine months to roughly $33 billion. An adjusted profit excludes items the company chooses to leave out, and is not the same as profit under standard accounting rules.

An equally large billThe disclosure comes as Anthropic prepares to go public. Yahoo Finance reports that it confidentially filed with the SEC on June 1, and its public prospectus has yet to appear. Political.org adds that the company has committed at least $518 billion over a decade to cloud services, computing and infrastructure. What prospective investors need to judge is whether revenue growth can keep outpacing those long-term commitments.

▮ SIGNALRevenue scale is no longer the open question for model companies; valuation now hinges on whether revenue grows fast enough to outrun the compute bills already signed.

❯ Modal raises a $750 million Series D at a $15.75 billion valuation as annualized revenue tops $1 billion

Valuation tripled in monthsAI infrastructure company Modal announced on its official blog a $750 million Series D at a $15.75 billion post-money valuation, led by Accel, with Kleiner Perkins joining as a new investor and existing backers including Menlo Ventures and General Catalyst participating. The company also said its annualized revenue run rate has passed $1 billion, with revenue up more than 15 times year over year. TechCrunch previously reported that Modal was valued at $4.65 billion four months earlier.

Running the servers for developersModal sells “serverless” AI compute: developers write the code, and Modal schedules GPUs, starts containers and runs the jobs, drawing on capacity from more than 20 clouds. Customers use it for model inference, image generation and training, and to give AI coding agents isolated “sandbox” environments to run in. Modal says sandboxes alone have reached about a $300 million run rate, with nearly 3 billion sandboxes launched so far.

Customers across industriesNamed customers include coding-agent maker Cognition, fintech company Ramp, music generator Suno, video generator Runway, and robotics companies Physical Intelligence and 1X. Modal has more than 200 employees in New York, San Francisco and Stockholm. The blog did not say exactly how the new money will be spent.

AI startups don't want to manage GPUsBuying and scheduling GPUs is costly and labor-intensive for many AI startups. Modal lets them pay by usage and scale up or down at will, turning compute into something closer to a utility. Agents need large numbers of short-lived, isolated environments, and the rapid growth of sandbox revenue shows that this new demand can already support a business of its own.

▮ SIGNALThe more agents there are, the bigger the business of giving them somewhere to work; sandboxes, once an overlooked piece of infrastructure, are becoming a billion-dollar revenue line.

❯ CoreWeave bond yields near 15% and its stock falls 15.5% as AI infrastructure debt rattles investors

One bond sparks a selloffAccording to SiliconANGLE, shares of AI cloud provider CoreWeave fell more than 15.5% on October 14. The trigger was a $1.25 billion batch of its bonds whose yield briefly passed 15%, meaning the bond’s price had dropped sharply. Investor Chamath Palihapitiya posted a chart of the yield on X, setting off a viral discussion, while traders bought large numbers of put options.

From 3 to 51 data centers in four yearsCoreWeave rents out GPU computing optimized for AI training and inference. The report says its data center network grew from three facilities to 51 in four years, financed largely with debt, and it reportedly has about $50 billion in outstanding obligations. In June its contract backlog stood at $104 billion, much of it tied to Microsoft, Meta and OpenAI. CEO Mike Intrator wrote in a memo to staff that he is confident the backlog “more than supports all of our debt obligations.”

The selloff spreads down the chainOn the same day, Oracle, which has borrowed heavily to build data centers, closed 8.8% lower; Dell, which counts neoclouds as customers, fell about 11.3%; AMD dropped about 5.7%; and DigitalOcean and others fell more than 10%. Rising borrowing costs eat directly into the margins of compute companies that expand on debt, and raise the bar for their next round of expansion.

▮ SIGNALHowever large the order book, a company has to survive until the cash arrives; the bond market is now using interest rates to reprice how fast AI infrastructure can expand.

❯ ASML earns €5 billion in Q3 and says supply in 2027 and 2028 won’t fully meet customer demand

Profit beats forecastsDutch lithography giant ASML released its third-quarter results on October 14: net sales of €11.66 billion, net income of €4.96 billion and a gross margin of 57.5%. Figures compiled by Techmeme from CNBC show sales slightly below the €11.72 billion analysts expected, while net income comfortably beat the €4.1 billion estimate.

Guidance raised againASML expects fourth-quarter sales of €14.2 billion to €15.2 billion, and has narrowed and lifted its full-year 2026 sales forecast from €43–45 billion to €44–45 billion. Lithography machines “print” circuit patterns onto wafers, and advanced chips cannot be made without ASML’s EUV systems.

Buyers have to queueCEO Christophe Fouquet said in the release that the supply ASML has planned for 2027 and 2028 “will not fully meet the demand of our customers.” The company is studying ways to add capacity for 2028 beyond the 30% increase it announced earlier, and some customers already want to place orders for 2029 and 2030. As a result, chipmakers’ expansion plans are increasingly constrained by tool deliveries, not just by money.

▮ SIGNALThe AI chip bottleneck is moving upstream from the fabs; when even lithography tools must be booked four or five years out, capacity itself becomes the scarcest asset.

❯ CXMT advances its 4F² DRAM architecture, with a new server memory module due by year-end

Standing transistors uprightAccording to Sina Finance, CXMT announced on October 10 a key advance in its 4F² DRAM architecture. President Cao Kanyu told the 4th Integrated Chips and Chiplets Conference that the company expects to launch a new-generation DDR5 RDIMM, a server memory module, using the technology by the end of this year. Specifications and a launch date have not been disclosed.

More cells in the same areaF is the smallest feature a chip process can make, and the area of each DRAM memory cell is measured in F². The industry standard is 6F²; 4F² means a smaller cell and more storage in the same space. The key is CXMT’s in-house VSG (vertical surrounded gate) transistor: it stands the transistor channel upright and wraps the gate around it, giving stronger control that eases leakage and short-channel effects. The company says it has more than 300 published patent families in 4F² technology.

Share and products advance togetherCXMT listed on Shanghai’s STAR Market on July 27 this year. According to Huxiu, citing an HSBC report from early October, CXMT’s share of the global DRAM market has risen to about 10%, from about 4% a year earlier. The same report says that on October 13, after the news, SK Hynix fell 10%, Samsung Electronics 7.4% and Micron about 4%. Earlier this year CXMT also announced mass production of LPDDR6 and of its fifth-generation process platform.

Server memory is the tough testChoosing a server memory module as the first 4F² product means facing data centers’ strict demands on performance, stability and compatibility head-on. AI servers keep needing more memory, and if the new architecture reaches volume production on schedule with steady supply, server makers will gain another DRAM supplier to choose from.

▮ SIGNALMemory competition is switching tracks: from shrinking line widths on an old architecture to being first to turn an entirely new architecture into a product that can be mass-produced.

❯ SpaceXAI’s Grok Voice 3.0 tops a speech-to-speech benchmark while holding its price at $0.08 a minute

No. 1 with 86.9SpaceXAI released the voice model Grok Voice 3.0 on its website on October 14. It scored 86.9 on independent evaluator Artificial Analysis’s speech-to-speech index, ahead of OpenAI’s GPT-Live-1 (83.2) and Google’s Gemini 3.8 Live (82.6). Models of this kind listen and speak directly, without first turning speech into text, and its time to first audio is under one second.

Steadier on calls and interruptionsOn a simulated support-call test with varied accents, background noise and interruptions, Grok Voice 3.0 scored 71.6%, slightly below GPT-Live-1’s 74.5%, though the company says its latency is 40% lower. On a test of handling pauses, turn-taking and interruptions it scored 97.0%. Pricing is unchanged from the previous generation at $0.08 a minute, and developers on the latest-version endpoint are switched to the new model automatically.

Tesla and Starlink firstThe company says Grok Voice already handles millions of conversations a day in Tesla vehicles, where drivers use it to check calendars, answer email and order coffee. Starlink’s call centers use it to answer about 20,000 calls a day, resolving roughly 70% without a human, and one in four sales calls ends in a purchase before the caller hangs up. Better results at the same price make it easier for call centers and other voice-heavy businesses to justify replacing some human agents with AI.

▮ SIGNALVoice AI has moved from “understanding” to “handling”: whoever reliably gets things done on noisy, interrupted real calls will win customer service, the biggest deployment market.

❯ Doubao phone assistant ends its notice period as 153 apps opt out of screen automation

Open by default, with an opt-out listAccording to the 21st Century Business Herald, the 30-day notice period for ByteDance’s Doubao phone assistant’s Screen Automation Execution Protocol ended on October 14. Third-party apps that have not filed an opt-out will now be opened to screen automation gradually by risk tier, and the opt-out list published the same day contains 153 apps. Developers can opt out at any time, and the platform promises a response within three working days.

Who is on the listBy the paper’s count, the list includes 65 apps from Tencent and its ecosystem and 23 from Alibaba and its ecosystem, both marked “including but not limited to” and together about 57.5% of the total, plus 14 from Kuaishou and 11 from Meituan. Another 33 are listed individually, including Pinduoduo, Xiaohongshu, Zhihu, Didi and several bank and carrier apps. Among AI rivals, DeepSeek and Kimi are listed individually, while Qwen, Yuanbao and Kling appear under their parent groups. No ByteDance or JD apps are named.

An assistant that taps the screen for youThe Doubao phone assistant comes preinstalled on ZTE Nubia’s M153 and NaviX Ultra phones. It reads what is on screen and simulates taps and swipes to compare prices, book restaurants or file leave requests inside other apps. After the technical preview launched last December, some apps showed login anomalies, and Doubao then disabled operations on WeChat and other apps. During the notice period, it could operate only system apps, ByteDance apps and apps that had explicitly agreed.

Super-apps don't want to be bypassedSuper-apps such as WeChat and Taobao earn much of their revenue from users browsing feeds and ads inside the app; if an assistant clicks through the steps on the user’s behalf, that exposure is skipped. The opt-out list shows that for phone agents to really work across apps, they will have to negotiate with each platform, not just prove they can tap the buttons.

▮ SIGNALThe hardest problem for phone agents is not model capability but who owns the entry point: whoever taps the screen for the user is diverting attention that used to belong to the apps.

❯ OPPO partners with Manus so Find X10 phones can carry out multi-step tasks for users

From answers to errandsAccording to IT Home, as carried by Sina Tech, OPPO announced on October 13 a partnership with AI agent platform Manus, integrating Manus’s agent capabilities into OPPO’s AI Mind Pilot as a new agent mode. It will debut on the upcoming Find X10 series running ColorOS 17. According to Sina News, the first models to get it run the international version of ColorOS 17.

Planning a trip in one sentenceAgent mode understands the user’s end goal, breaks it into steps, searches for information, picks tools and delivers a finished result, and it can work directly with files and photos on the phone. OPPO’s examples include pulling flight, hotel, restaurant and ticket details from different apps and combining them with online information into a daily itinerary, or finding trip photos with one instruction and turning local files and web material into a presentation.

Local data stays on the phoneOPPO stresses that ColorOS 17 uses on-device models to understand information locally. For local tasks such as finding, organizing and renaming files and photos, the original content stays on the device and is not shared with Manus, and cloud processing requires the user’s explicit permission first. Tang Kai, OPPO’s president of software engineering, said agentic AI marks a shift from simply answering questions to helping users get things done.

▮ SIGNALBy building agents into the operating system, phone makers hold what third-party apps cannot: control over local files and privacy permissions, the precondition for users to hand over tasks with confidence.

❯ Apple plans to launch its first Apple Intelligence features in mainland China with iOS 27.3 in December

Basic features firstBloomberg’s Mark Gurman reports that Apple is preparing to launch its first Apple Intelligence features in China in December. MacRumors, citing Bloomberg, says they will arrive with iOS 27.3, starting with basic features such as photo editing; Siri AI, Image Playground and the newer Genmoji are not included. Apple has not officially announced a launch date.

Working with Alibaba and BaiduApple Intelligence is the AI feature suite Apple introduced in iOS 18.1, and it has not been available in mainland China until now. MacRumors says Apple is working with Alibaba and Baidu in China, and that the Ministry of Industry and Information Technology has granted Apple network access permits to add Alibaba’s Qwen and Tongyi Wanxiang models to the iPhone Duo, iPhone 18 Pro and other models.

Local rivals got there firstiPhone users in mainland China will get Apple’s own AI features for the first time, though the first batch is limited and less complete than the international version. Huawei and other local brands have long offered AI features in China, and Apple will need later updates to catch up in the premium-phone AI experience.

▮ SIGNALIn China, the AI phone race is not about who releases a model first but who can safely build a local partner’s model into the system first.

❯ Apple closes above a $5 trillion market value for the first time, the second company after Nvidia

Seventh time luckyAccording to 9to5Mac, Apple shares closed at $342.96 on October 14, up 0.5%, for a market value of $5.01 trillion, its first close above $5 trillion. The Wall Street Journal noted it was the seventh time Apple had crossed $5 trillion intraday; the earlier attempts did not hold to the close. Nvidia reached that level first, in October 2025.

Rising while tech fellMost tech stocks fell that day: Oracle dropped 8.83%, Meta 3.27%, Microsoft, Amazon and Google 1.7% to 1.8%, and Nvidia 3.16%. Possible drivers cited in the report include new products such as the iPhone 18 Pro lineup and iPhone Duo, and investor confidence in John Ternus, who became CEO on September 1. Apple reports its next quarterly results on November 2.

A test aheadOn the same day AI infrastructure stocks fell together on debt worries, Apple rose. 9to5Mac notes that the November 2 results will show how Apple is handling rising component costs and recent hardware price increases, testing whether the optimism around new products and new leadership shows up in profits.

▮ SIGNALOn a day when companies borrowing to build compute were sold off and Apple, which sells the hardware, hit a record, the market was sorting out who carries AI’s costs and who collects its demand.

❯ OpenAI accuses an ex-employee of taking unreleased device files to Apple, escalating their trade secrets fight

More than a hundred filesAccording to 9to5Mac, OpenAI said in a court filing on October 14 that a former employee, on his last day on June 2, sent more than a hundred highly confidential files about at least one unreleased device to personal accounts. The files include detailed device images, engineering and manufacturing data, supplier information, and reports on cosmetic coatings and sealing. Apple said it has found no evidence supporting the allegations.

From defendant to accuserThe employee moved from Apple to OpenAI, worked there about six months and then returned to Apple. OpenAI says he also presented OpenAI’s plans to Apple’s senior leadership, including its vice president of operations in Asia. The case began when Apple sued former employees in July, accusing them of stealing trade secrets for OpenAI; TechCrunch reports that a judge had already ordered OpenAI to hand over related hiring records.

The cost of hardware talent movingOpenAI wants Apple to assure it that Apple does not possess or use OpenAI’s trade secrets, and says Apple has not done so. Both companies are developing AI hardware and engineers are moving back and forth between them, so the outcome will directly shape how hardware teams recruit and vet new hires, and could slow both sides’ device development.

▮ SIGNALThe first front in the AI hardware race is not products but people: when the same engineers shuttle between rivals, every hire can become evidence.

❯ OpenAI engineer arrested over an alleged bug in an executive meeting room, built on a $14 board

Hidden behind the TVAccording to TechCrunch, Abdulrahman Alfozan, an engineer on OpenAI’s product team, was arrested by the FBI on October 9 and charged with attempting to intercept oral communications, which carries up to five years in prison. Prosecutors say he planted a hidden recording device in the USB port of a television in an executive conference room at OpenAI’s San Francisco office. OpenAI has fired him.

A development boardThe Register, citing the charging documents, says the device was a Seeed XIAO ESP32-S3 Sense development board, planted around September 28 and found wrapped in toilet paper by staff on September 30. Tom’s Hardware puts its price at about $14. On October 5, security staff found a second recording device from the same maker in a bag he had left behind.

Using AI to find the roomsThe documents say he used an internal productivity agent to ask about office addresses, meeting rooms, calendars and where external partners met, and that a chatbot account linked to his spouse was used to ask how to build and modify a recording device. OpenAI says it has no indication he accessed or exfiltrated model weights, source code or customer data, and the FBI is still investigating.

▮ SIGNALA dev board costing a few dollars plus an internal AI tool was enough to map out executive meetings; protecting secrets at AI companies starts with governing the AI in employees’ hands.

❯ Linus Torvalds uses Claude Code to rework kernel code, roughly tripling small-read throughput

From 10 million to 30 million reads a secondAccording to PBX Science, Linux creator Linus Torvalds pushed an experimental branch called noref-filemap-read-with-llm. Claude Code produced the code, test programs, a review and the first draft of the cover letter under his direction. On his laptop, eight threads repeatedly reading the same cached file went from about 10 million reads a second to more than 30 million.

One fewer counter updatePreviously, every read of cached file data had to update a shared “reference count,” so threads reading the same file at once fought over it. The new patches use the kernel’s RCU mechanism so reads no longer need to modify that count, reducing contention. The trade-off is that reading a 16-byte range from a page not yet mapped becomes slower. The branch has not been merged and is a candidate for the Linux 7.4 merge window.

Decade-old bugs found along the wayIn separate work, Torvalds used LLMs to make sense of the kernel’s obscure associative array library, assoc_array, and built a test harness that confirmed a bug reported by developer Nils Ollrogge and uncovered three more, all dating back to the library’s first commit in 2013. In his tag message he wrote that he “did what I preach,” using modern LLM tools to understand the code.

▮ SIGNALWhen even the most demanding kernel maintainer has AI write code, the developer’s value becomes clearer: deciding what to change and judging whether it is right are harder to replace than typing it.

❯ Anthropic opens a Singapore office and signs an AI safety MOU with the Singapore government

Fifth office in Asia-PacificAnthropic announced on X on October 14 that it has opened an office in Singapore, its fifth in the Asia-Pacific region, and signed a memorandum of understanding with the Singapore government on AI safety and responsible adoption.

Signed before the ministerAccording to Tech Edition, the MOU was signed by Chng Kai Fong, Permanent Secretary of the Ministry of Digital Development and Information, and Anthropic co-founder Ben Mann, and witnessed by Minister for Digital Development and Information Josephine Teo and Michael Sellitto, Anthropic’s head of APAC policy.

Testing standards and trainingUnder the MOU, Anthropic will hold technical exchanges with Singapore’s AI safety and assurance institutions on model evaluation, testing standards and incident response, support AI skills for small businesses and public officers, and share research data including from the Anthropic Economic Index. Customers named in the report include GovTech Singapore, DBS, GIC, Grab and Sea, and more than 1,500 GovTech engineers and technical staff use Claude Code.

Second in per-capita useAnthropic says Singapore ranks second among 121 countries for Claude.ai usage relative to population, nearly six times the expected level. A local office lets it serve those government and corporate customers more closely. The MOU is a cooperation framework, and no concrete outcomes have been announced yet.

▮ SIGNALA safety MOU is both a compliance gesture and a pass into government customers, and markets like Singapore, with heavy usage and clear rules, are the first to be courted.

❯ Salience Labs raises a $125 million Series B for all-optical switch chips that speed up AI data centers

Intel and AMD both inAccording to SiliconANGLE, silicon photonics company Salience Labs raised a $125 million Series B led by Bullhound Capital, with Intel Capital and AMD Ventures as strategic investors and SBI Group, Hudson River Trading and the British Business Bank joining as new backers.

Keeping data as lightSalience Labs makes network switch chips for data centers. Today’s fabrics convert optical signals to electrical ones for switching and then back to optical, adding latency and power at every conversion; its all-optical switch routes data as light directly on the chip. In September the company showed a 64-port all-optical switch, with 128- and 256-port versions in development. Its CEO is co-founder Vaysh Kewada.

The network bottleneck in giant clustersThe report says that as GPU clusters grow to tens of thousands of chips, electronic switch fabrics cannot keep up with bandwidth needs, making latency and power consumption bottlenecks. Industry analyst Cignal AI expects the optical circuit switching market to exceed $8 billion by the end of the decade. The money will go to volume manufacturing, market expansion and R&D on higher-density switches; no customers have been disclosed.

▮ SIGNALAs AI clusters grow, the bottleneck shifts from computing fast to moving data fast, and the links between GPUs are becoming the next chip category worth betting on.

❯ Workplace chat app Zenzap raises $44 million to put AI agents into frontline teams’ group chats

Aiming to replace WhatsApp at workAccording to TechCrunch, workplace messaging app Zenzap raised $44 million from investors including Bessemer, Lightspeed and The Garage, with individuals such as Wix president Nir Zohar also taking part. The valuation was not disclosed. The company plans to use the money to expand its product and R&D teams.

Built for stores and job sitesZenzap was founded in 2022 by Guy Weiss and others and has about 50 employees across the US, Tel Aviv and Europe. It targets hospitality, construction, healthcare and fast food, helping companies communicate with frontline staff while following IT and security rules. Customers include FedEx, Dollar General, Taco Bell and Starbucks. The free tier offers group chats, task management and AI search, and paid plans start at $51 a month for up to 20 users.

Agents move into the group chatZenzap has now launched AI agents for workflow automation, along with AI transcription of voice notes and calls, billed by credits. Stores and job sites reschedule shifts often and staff rarely sit at a computer, so calling an agent inside the chat window everyone already uses is easier to roll out than deploying a separate piece of enterprise software.

▮ SIGNALThe shortest route for agents into frontline industries is not new software but the chat window workers already open every day.

❯ Berkshire Hathaway invests another $330 million in Bending Spoons, lifting its stake to 4%

10 million shares at a discountAccording to a Form 6-K Bending Spoons filed with the SEC, Berkshire Hathaway will buy 10 million shares at $33 each, about $330 million, raising its stake to 4%. The price is about 13% below the previous close. According to the Financial Times, the shares jumped more than 40% on the news. The deal is expected to close in the fourth quarter.

Buying mature software and remaking itBending Spoons is an Italian tech company whose model is to acquire established digital businesses, restructure and continuously optimize them to grow earnings, and reinvest the proceeds in further acquisitions. Its portfolio includes Evernote, Vimeo, WeTransfer, Eventbrite, Airtable and AOL. It listed on Nasdaq in July, when Berkshire invested $600 million.

Ammunition for dealsThe company says the money will go to general corporate purposes and new acquisitions. Francesco Patarnello, its head of business acquisitions, said the company continuously evaluates M&A opportunities and its pipeline remains strong. Berkshire’s long-term capital adds to Bending Spoons’ firepower, and mature software products with slowing growth now have a buyer that can act with more confidence.

▮ SIGNALMature software does not have to grow fast to make money; acquiring, integrating and steadily optimizing it can be a business too, and that is the model patient capital is backing.

❯ Shareholders approve Fox’s $22 billion Roku deal, expected to close in the first half of 2027

A near-unanimous voteAccording to Variety, shareholders of both Fox and Roku approved Fox’s roughly $22 billion acquisition of Roku. TheWrap, citing Fox’s preliminary tally, says about 197 million Class B shares voted in favor and about 72,000 against; Roku’s final count will come in a separate SEC filing.

Content plus distributionAfter closing, Roku shareholders will own about 27% of the combined company and Fox shareholders about 73%. Fox owns live sports, news and the free streaming service Tubi, which has 110 million monthly active users; Roku is a TV streaming platform reaching more than 100 million households. Fox CEO Lachlan Murdoch called it a “transformative transaction” that pairs Fox’s content with Roku’s distribution and technology platform.

One more hurdle at the DOJThe US Justice Department issued a “second request” for more information in September, extending its review. Fox expects the deal to close in the first half of 2027. The report says the combined company would become the third-largest US media distributor, behind YouTube and the merged Paramount–Warner Bros. Discovery.

▮ SIGNALStreaming competition increasingly looks like a battle over distribution: content companies no longer want just to sell shows, but to own the first screen viewers see when they switch on the TV.