❯ Anthropic begins IPO roadshow for 825 million shares alongside a $26 billion private placement
An offering rangeAnthropic began its IPO roadshow on November 10. It plans to offer 825 million Class A shares at $70 to $78 each and list on Nasdaq under ANT. At the top of the range, the base offering would total about $64.4 billion, including roughly $62.5 billion in newly issued shares from Anthropic, both before fees. The registration statement had yet to take effect at the launch of the roadshow, and the final price remains open. Underwriters also have a 30-day option to buy up to 123.75 million more shares.
The seven founders sell sharesExisting holders are offering about 23.65 million shares, with proceeds going to the sellers. At the $74 midpoint, each of the seven co-founders stands to sell roughly $250 million, or $1.75 billion combined. The amended prospectus also says a special class of stock held collectively by the founders will carry 50.1% of the voting power on matters submitted to shareholders generally.
Suppliers join the private placementNvidia, Google, AMD, Broadcom, Micron, Microsoft and TSMC have agreed to buy a combined $26 billion of Class N common stock at the IPO price. The placement is scheduled to close immediately after the IPO, subject to closing conditions. At the $74 midpoint, Nvidia would invest about $10 billion, roughly a tenth of the new financing expected that month. Class N shares generally carry no vote, and Google’s purchase funds part of a previous investment commitment.
Goldman's growth scenarioAnthropic sells Claude to individuals and businesses and charges for API usage. Goldman Sachs, one of its lead IPO underwriters, has modeled a path for that revenue to grow. Techmeme’s summary of Financial Times reporting says Goldman projects about $52 billion in 2026 revenue and more than $600 billion in 2030, with operating profit in 2029 and positive free cash flow from 2030.
▮ SIGNALThe offering lets Anthropic raise money from public investors and key compute suppliers while its voting structure gives the founders a way to retain control as the shareholder base expands.
❯ OpenAI’s quarterly revenue reaches about $10 billion and adjusted operating margin improves to negative 109%, reports say
Quarterly growth picks up againThe Information reported that OpenAI told investors on November 10 that third-quarter revenue rose nearly 50% from the previous quarter to about $10 billion. The outlet reported 17% sequential growth in the second quarter. Demand for newer models helped drive the rebound, and customers can use the API to integrate those models into their own applications.
The loss per revenue dollar narrowsThe Wall Street Journal’s Berber Jin reported, citing internal materials, that adjusted operating margin improved from negative 150% in the second quarter to negative 109% in the third. On that measure, the operating loss fell from about $150 to $109 for every $100 of revenue. With sales on a larger base, the quarterly adjusted operating loss was still about $10.9 billion.
Usage brings sales and costsOpenAI sells ChatGPT services to individuals and businesses and charges developers for API usage. More API calls generate revenue and require additional compute to answer questions or carry out tasks. Sustaining that pace of API growth will depend partly on whether developers keep embedding newer models in frequently used workflows.
▮ SIGNALOpenAI lost less per dollar of revenue, while the size of the overall loss still depends on how much additional sales can cover training and model-serving costs.
❯ Google opens Gemini 4 Carbon to Ultra subscribers and paid API users for complex work
Access expandsGoogle opened access to Gemini 4 Carbon on November 10. Developers can use it through the paid Gemini API, and enterprises can access it through Gemini for business. Google AI Ultra subscribers can use it in the Gemini app, AI Mode in Search and the Antigravity development environment.
A different job from ArgonArgon remains Google’s choice for most coding and knowledge work. Carbon uses more tokens to explore, execute and verify harder tasks over multiple rounds. In one Google demonstration, several Carbon agents extracted engineering drawings from seven public planning documents and iterated ten times to build an interactive 3D model of an office building.
Two price tiers over timeIntroductory public API pricing runs through November 30 at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward. Tokens are the units billed for the content a model reads or generates, so repeated checks add to usage. Antigravity is also previewing a collaboration feature that can give Carbon work such as architecture and coordination while assigning other steps to lighter models.
▮ SIGNALCarbon gives teams a way to reserve their strongest model for steps that need repeated verification. That division of work will directly affect the cost of the whole job once the introductory rates double.
❯ CoreWeave’s quarterly revenue rises 168% as its backlog reaches $132.5 billion
Revenue beats expectationsCoreWeave reported third-quarter results on November 10. Revenue rose 168% year over year to $3.661 billion, above the $3.55 billion consensus reported by CNBC. Its net loss widened to $709 million from $110 million a year earlier as quarterly interest expense climbed to $846 million.
Backlog depends on deliveryCoreWeave rents GPU and data center capacity to customers including AI developers that train and run models. Its backlog rose from $104 billion at the end of June to $132.5 billion at the end of September. The measure includes revenue CoreWeave expects to recognize subject to delivery and service availability, as equipment comes online and contracts are fulfilled.
Funding moves closer to projectsThe company raised its 2026 revenue outlook to $13 billion to $13.3 billion and still expects $37 billion to $39 billion of capital expenditures. Management said customer prepayments and hardware partner financing would supply most of next year’s funding, and plans to raise new debt tied to a set of contracts. At prevailing yields, it does not intend to issue more unsecured parent-level debt through 2027.
▮ SIGNALLarge customer prepayments can shorten the period CoreWeave has to fund GPUs and data centers itself, easing the dependence of rapid expansion on parent-company borrowing.
❯ TSMC approves investments of up to $500 million each in Anthropic and OpenAI
Two investments approvedTSMC’s board approved two investments on November 10. The foundry plans to buy up to $500 million of Anthropic common stock at its IPO price and up to $500 million of OpenAI preferred stock, for a combined maximum of $1 billion.
Two routes into the companiesAnthropic’s prospectus names TSMC subsidiary TSMC Partners as a participant in its private placement. That placement is priced at the public offering price and involves Class N stock, which generally has no voting rights, with closing expected immediately after the IPO. OpenAI is privately held, and TSMC’s board resolution gives a cap for the preferred-stock purchase. It did not disclose pricing, the specific rights attached to those shares or a closing date.
From manufacturing to ownershipTSMC’s core business is manufacturing chips designed by customers. At the same meeting, the board also authorized about $40.9 billion in capital appropriations for areas such as process capacity, packaging and facilities. The two AI investments are equity allocations, with the disclosed terms focused on share class, pricing or spending caps.
▮ SIGNALTaking stakes in both model companies gives TSMC potential equity upside alongside its chipmaking business, with a defined cap on each investment.
❯ Mercor raises $750 million at a $20 billion valuation to expand enterprise AI agent evaluation
Valuation doubles in a yearMercor announced a $750 million Series D on November 10 at a $20 billion valuation, twice its October 2025 valuation. Benchmark, General Catalyst, Nvidia and Salesforce were among the participants. The round included an employee tender, and Mercor plans to invest more in evaluation, post-training and simulated work environments.
Turning expertise into evaluationsMercor initially organized specialists for frontier AI labs, turning their practical judgment into training examples and evaluations. It now also designs tests for enterprises, using their workflows and quality standards to measure whether agents finish tasks and how much each run costs. Companies can use the results to pick models and decide where additional training is needed.
Practice in simulated workplacesMercor acquired Deeptune this year. The team recreated hundreds of enterprise applications so agents could practice using them. Simulated environments allow companies to test the same workflow repeatedly and see whether a newer model completes it more reliably. Mercor aims to sell the expert network and testing capabilities it built for AI labs to a broader set of enterprises.
▮ SIGNALOnce agents are wired into internal workflows, general-purpose benchmarks reveal less about production performance. Mercor can fill that gap with tests based on the customer’s own tasks and acceptance standards.
❯ Oura agrees to buy Counsel Health for $625 million to bring physician visits into its app
An all-cash acquisitionOura signed an agreement on November 10 to buy Counsel Health for $625 million in cash, subject to purchase-price adjustments. Counsel combines medical AI with an in-house physician team to provide on-demand primary care. The parties expect to close by year-end, subject to customary conditions.
From ring data to a consultationThe two companies have partnered since May 2026. Members can ask Counsel’s AI in the Oura app about changes tracked by the ring and bring a licensed physician into the conversation when needed. The physician can use longer-term biometric trends and medical records to diagnose, prescribe or refer as appropriate.
Consumer and enterprise channelsOura’s revenue rose 153% year over year to $533 million in its fiscal fourth quarter ended September 30. Fiscal 2026 revenue rose 92% to about $1.75 billion. Oura has more than six million members, and Counsel plans to continue expanding through enterprise channels such as health plans, health systems and employers after the deal closes.
▮ SIGNALCounsel can turn a change detected by Oura into a service that involves a physician, while gaining access to users who already collect health data over time.
❯ MiniMax releases H3-Image weights with native 4K output and up to 16 reference images
An open-weight image modelMiniMax released H3-Image and made its weights available on November 11. The H3-based still-image model can generate pictures from text or reference images and edit existing pictures. Developers can download the weights from Hugging Face or use the model through the MiniMax API, MiniMax Code or Hailuo AI.
4K and multiple referencesThe model generates square images at 4,096 by 4,096 pixels and can work with up to 16 reference images. MiniMax says a 4K image took under 10 seconds in its testing. The company attributes the speed to fewer sampling steps during generation and changes to training and deployment.
From preview to finished assetHigher resolution preserves more detail when an image is cropped or enlarged. Multiple references let a team supply examples for the subject, product and composition in the same request. MiniMax is offering half off standard pricing on hosted generation at 4K or above during launch week. Teams with existing production workflows can also use the downloaded weights to integrate generation and editing into their own tools.
▮ SIGNALWhen high-resolution output arrives quickly enough, teams can review composition and detail at the final size and spend less time moving between low-resolution previews and a separate upscaling step.
❯ Meituan releases LongCat-2.5 under MIT for visual and multistep coding tasks
Downloadable weightsMeituan published the LongCat-2.5 weights on November 9 after introducing the model on November 2. The MIT license permits commercial use, modification and redistribution subject to conditions such as retaining the license notice. The model accepts text and visual inputs and is designed for jobs that take multiple steps.
From images to actionThe model can read screenshots, charts and video. A developer could give it a screenshot of a broken page, have it locate the relevant code and then inspect the rendered result after a change. In Meituan’s tests, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, up from 70.8 for its predecessor, and 64.6 on DeepSWE v1.1, up from 9.7.
Hosted access or self-deploymentLongCat-2.5 has 1.6 trillion total parameters and activates about 48 billion per token on average. It uses only part of the model for a given input, though a self-hosted deployment still needs hardware to hold the full weights. Meituan’s hosted API lets teams trial the model by usage before committing hardware to a private deployment.
▮ SIGNALOpen weights give teams handling large volumes of screenshots and internal code a self-hosting option, with the 1.6-trillion-parameter footprint setting the hardware cost of that control.
❯ Alibaba discloses T-Head’s listing application and expects to retain control after the spin-off
The application becomes publicAlibaba disclosed on November 10 that T-Head, its chip subsidiary, confidentially submitted a listing application to the Hong Kong exchange on September 11. The exchange has cleared Alibaba to proceed with the proposed spin-off. Approval for the shares to list and trade is still pending.
Staying within the groupT-Head is currently wholly owned by Alibaba and develops data center chips and related computing systems. Alibaba expects T-Head to remain a subsidiary after the spin-off. A separate listing would give the market a way to value the chip business on its own and provide T-Head with direct access to capital markets.
Offering terms remain openThe proposed offering includes a public sale and a placement for institutional and professional investors. Alibaba has yet to publish an offering size, price or listing date. Completion depends on exchange approval and market conditions, and Alibaba shareholders would mainly retain exposure to the business indirectly through the group’s stake.
▮ SIGNALA separate listing would give T-Head its own financing channel, while Alibaba’s plan to retain control preserves an organizational basis for working with the group’s data centers.
❯ SoftBank Group posts a $7 billion quarterly loss after revising US data center power plans
Revenue rises as profit turns to lossSoftBank Group released its fiscal second-quarter results on November 10. It recorded a net loss of ¥1.12 trillion, or about $7 billion, for the three months ended September, compared with a profit of about ¥2.5 trillion a year earlier. Revenue of about ¥2.22 trillion beat the ¥2.02 trillion consensus cited by CNBC.
Power equipment bought ahead of timeSoftBank paid in advance to secure US data center power equipment as it anticipated shortages and long lead times. Payments reached about $9.48 billion by the end of September. After revising its procurement plans, the group recorded a related loss of ¥597.2 billion, or about $3.7 billion, in the six months through September. Declines in some listed holdings also affected results.
Data center projects are still movingSoftBank continued to highlight its Ohio project at the earnings briefing, pointing to a 20-year lease with OpenAI covering eight gigawatts of IT capacity. AI data centers need a power supply before facilities and chips can come online. Paying for equipment in advance can save delivery time, while changes to project plans can affect how much of that spending is recovered.
▮ SIGNALWhen power equipment has to be ordered far in advance, a data center’s exposure starts with the deposit, and a change in plans can produce a loss before the project ever comes online.
❯ Former Dali Education CEO Chen Lin leaves ByteDance and builds an AI classroom product
A new education ventureEarly ByteDance employee and former Toutiao and Dali Education CEO Chen Lin addressed his departure from ByteDance publicly for the first time in an essay about returning to education as a founder. Citing corporate records on November 10, The Paper reported that Beijing Galaxy Intelligence Education was registered in late 2024 and is wholly owned by Chen. Its school-facing product is called Xiaolu Aixue.
Stepping back from DaliChen oversaw product work in ByteDance’s early years and took charge of Dali Education when it became a standalone brand in 2020. In his account, the business contracted in 2022 and health issues led him to step back. He spent about two and a half years recovering, during which he invested and advised some ByteDance departments. In the account reported by IT Home, Chen traced the idea to a rainy night at Sayram Lake in June 2023. As generative AI began reshaping professional work, he started thinking about how children could practice solving problems independently.
Schools and teachers are the entry pointGalaxy Intelligence Education positions Xiaolu Aixue as a classroom product. Students can ask questions, practice and correct mistakes at their own pace. The system connects their answers, mistakes and feedback so teachers can plan explanations or individual support. Adoption in classes, study periods and practice sessions depends on whether teachers can use that feedback in their regular work.
▮ SIGNALXiaolu Aixue puts AI-generated feedback in teachers’ hands. Its value depends on whether teachers can use it to see sooner which student is stuck and where.
❯ SpaceX unveils Starlink Mini 2 with a hot-swappable battery rated for up to 4.5 hours
A battery inside the terminalOn November 10, SpaceX unveiled the Starlink Mini 2, combining a satellite dish, Wi-Fi router and battery in a laptop-sized terminal. SpaceX says a full charge provides up to 4.5 hours of use, and separately sold spare batteries can be swapped while the device is running.
Thinner and slightly heavierThe first Mini already combined the dish and router and required external power. The Mini 2 is 23.1 millimeters thick, down from 38.5 millimeters. It weighs about 1.6 kilograms with the battery, slightly more than its predecessor, retains an IP67 rating and can also use a 60-watt USB-C power supply.
Invitation-only preorders in the USPreorders are initially open to selected US customers by invitation. The Verge reported that screenshots shared by recipients showed a $999 offer that includes a year of Roam Unlimited service. The bundle is aimed at users who take the terminal with them, with broader availability dates still to be announced.
▮ SIGNALBuilt-in power reduces the equipment needed to set up a temporary connection, and hot-swappable batteries let that connection outlast a single charge.
❯ Trump Media posts about $410 million in quarterly profit as digital asset gains lift earnings
A profitable quarterTrump Media & Technology Group reported third-quarter results on November 9. Net income was $409.7 million, compared with a $54.8 million loss a year earlier. About $378 million in net realized and unrealized gains on digital assets and pledged digital assets accounted for most of the quarter’s profit.
Subscriptions bring in additional salesRevenue rose 205% year over year to $3 million. Media subscription revenue from Truth API and the Truth+ video service increased by a combined $1.8 million. Truth API went live on August 1 as the company’s first data licensing product, giving institutional customers a low-latency feed of public posts they can connect to their own systems.
The media business is still smallThe company recorded an adjusted EBITDA loss of about $12.3 million. That measure excludes items such as gains on digital assets. Data licensing adds a subscription stream to the media business, while changes in asset prices continue to move net income.
▮ SIGNALInstitutional subscription growth is a better way to track Truth API’s progress, since changes in digital asset prices can move quarterly net income by far more than the subscription business does.