❯ Anthropic files for IPO with $16.3 billion in first-half revenue and over $600 billion in compute commitments
Claude heads for public marketsAnthropic says it has publicly filed for an IPO and plans to list on Nasdaq under the symbol ANT. The number of shares and offering price have not been set. The Wall Street Journal reported, citing people familiar with the matter, that Anthropic is targeting the week of November 16 and could raise about $100 billion at a valuation of $1.5 trillion to $2 trillion. Those are reported targets, not final terms.
Six months beat last yearReuters reports that Anthropic’s revenue grew from roughly $41 million in 2023 to $386 million in 2024 and $4.59 billion in 2025. It reached $16.29 billion in the first half of 2026, about 14 times the $1.16 billion recorded a year earlier and far more than the whole of 2025. The filing provides the first systematic public view of the model developer’s business at this scale.
How Anthropic earns moneyThe company’s core technology is the Claude model family, with tiers including Fable, Opus, Sonnet and Haiku. Individuals can chat and work with documents through Claude or pay for Pro and Max subscriptions. Businesses and developers can integrate Claude into their own software through an API and pay according to the tokens processed. Reuters says consumption-based business use accounts for most revenue.
From models to applicationsAnthropic also builds products around its models, including Claude Code for reading, writing and modifying software, and Claude Cowork for enterprise tasks. Its early focus on coding created opportunities for sustained use and a route to sell other Claude services to companies. TechCrunch reported that more than half of customers spending at least $100,000 use both Claude Platform and at least one first-party app.
Large customers form tiersTechCrunch’s figures, relayed by Techmeme, show that as of September 30 Anthropic had over 11,900 customers who generated at least $100,000 in the preceding 12 months. More than 1,900 generated at least $1 million and roughly 250 generated at least $10 million. These thresholds overlap and cannot be added together. They show Claude’s reach beyond individual users into enterprise systems and team workflows.
Understanding the $80.7 billion lossAnthropic reported a net loss of approximately $80.7 billion in the first half, including a large accounting adjustment connected to financing agreements. AP reports that the adjustment accounted for about $72.4 billion, so the net loss is not equivalent to cash spent on operations during the period. Reuters puts operating loss at $693 million, down from $2.14 billion a year earlier, as rapid revenue growth offset part of the increase in expenses.
Compute is the biggest expenseTraining models and answering user requests both require computing resources. Anthropic pays for cloud capacity, data center space, equipment leases and energy. Reuters says it has committed to more than $600 billion of spending with around eight partners over roughly a decade; some agreements are non-cancellable or require payment regardless of usage. The figure covers future years and is not the amount paid in the first half.
Owned equipment and leasesThe Hum’s analysis of media and filing materials reports $482 million of property and equipment purchases in the first half. As of June 30, signed leases that had not yet begun represented roughly $321.4 billion in future payments, some subject to conditions. Culpium has put forward a figure of about $782 billion, but its calculation is not publicly visible and The Hum questions whether it consists entirely of leases. These figures have different dates and scopes and cannot simply be added.
Who controls the companyFormer OpenAI researchers founded Anthropic in 2021, including siblings Dario and Daniela Amodei. Dario is CEO; Daniela is president and board chair. Reuters says Anthropic will remain a public benefit corporation after listing. Its seven co-founders will direct a single special Class F share carrying 50.1% of voting power on important shareholder matters, while an independent long-term benefit trust participates in board governance. The structure is intended to preserve the company’s approach to AI development as it raises capital.
Major backersGoogle and Amazon are both important financial backers and compute partners. The Wall Street Journal reports that venture firm Spark Capital invested nearly $800 million across six funding rounds. Spark partner Yasmin Razavi is the only venture capitalist on Anthropic’s board.
Why go publicGoogle and Amazon supply computing as well as funding. Cloud services and leases let Anthropic obtain capacity without building every data center itself, but they also create purchase and rental obligations over many years. An IPO would add a funding channel; continued conversion of Claude usage into revenue is central to understanding those commitments.
▮ SIGNALAnthropic has demonstrated enterprise demand for models and applications, while its expansion increasingly depends on securing compute and funding years in advance.
❯ OpenAI releases GPT-6.1 Terra for faster, cheaper computer use
From seeing to doingOpenAI has released GPT-6.1 Terra, emphasizing visual understanding, computer use and natural writing. It can interpret screen content and act across applications by clicking, typing and switching between them—for example, reading an invoice and entering its details into accounting software.
Following the previous TerraOpenAI positions Terra as a model balancing speed and cost for everyday computer work such as forms, technical drawings and video editing. It follows GPT-5.6 Terra and aims to run more visual tasks at lower cost than stronger models previously required.
Speed and costOn Agents’ Last Exam, which evaluates professional work tasks, OpenAI says GPT-6.1 Terra scores 15 percentage points above GPT-5.6 Terra at medium reasoning effort and about one-third the cost per task. The evaluation includes turning architectural drawings into a 3D model. These are company-reported results under specific test conditions.
How to access itStandard API pricing is $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. The model is available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu users, and developers can access it through the API. For teams processing large numbers of forms or business-system tasks, the cost of each run helps determine whether agents can handle the work continuously.
▮ SIGNALComputer use is moving beyond whether a model can complete a demonstration: speed and cost per task now shape its place in daily workflows.
❯ OpenAI announces three AI-assisted research results and invites academic review
Three fieldsOpenAI says its models produced research results in mathematics, computer science and physics concerning local smoothing and the Kakeya conjecture, the Exponential Time Hypothesis, and the mathematical foundations of quantum chromodynamics. The company describes the manuscripts as material for researchers to understand and examine further.
Mathematics and complexityOpenAI says its model produced a proof of the local smoothing conjecture in every dimension, with implications for the related Kakeya conjecture. In computer science it claims the long-standing Exponential Time Hypothesis is false. That hypothesis concerns how quickly certain computational problems can be solved; if the argument holds, it would affect work built on that assumption. OpenAI says the result was subsequently verified in the formal proof tool Lean; the hypothesis’s original proposers will join the external group examining and explaining it.
The physics resultThe third result concerns quantum chromodynamics, the theory describing the strong interaction. Unlike the other two, OpenAI has left the timing and manner of releasing the physics materials to an external collaboration, so the details available to the public differ. The company says domain experts should interpret the results rather than treating model output as a substitute for academic discussion.
Three release approachesFollowing outside advice, OpenAI released the mathematics output directly. It also made the computer science output available while forming a group to explain it; the physics group will decide on its material. OpenAI says participants are not required to sign nondisclosure agreements and may discuss intermediate work and publish conclusions. The results remain OpenAI’s claims, not findings already accepted across the academic community.
▮ SIGNALWhen AI generates complex arguments, expert review, interpretation and reuse become part of the research workflow itself.
❯ Meta plans to raise up to $95 billion in equity for AI expansion
Two-stage fundraisingMeta, the parent of Facebook and Instagram, plans to raise up to $95 billion through equity financing for AI infrastructure and products. Reuters reports that it has launched roughly $35 billion in public offerings and plans a separate program for up to $60 billion. The timing and terms of the latter are not final.
The first saleThe launched offerings comprise approximately $22.5 billion of Class A common stock and $12.5 billion of mandatory convertible preferred stock. The second program is expected after non-voting Class C shares begin trading, allowing sales over time rather than delivering the entire $95 billion at once.
Why raise moreMeta is expanding chip purchases and data centers for model training, AI products and its existing businesses. In its latest earnings report, the company forecast 2026 capital expenditure of $135 billion to $145 billion. Third-quarter capital expenditure was about $34.9 billion and free cash flow was negative $2.12 billion. Equity would add another source alongside operating cash flow.
Suppliers show interestNvidia, AMD and Broadcom have indicated interest in buying up to $5 billion of Meta Class A shares in aggregate, Reuters says; those purchases have not been completed. The companies also supply chips to Meta, so the potential investment would connect supplier capital to a customer’s expansion.
▮ SIGNALEven a technology company with strong operating cash flow is turning to major stock sales as AI infrastructure changes its funding needs.
❯ SpaceX reaches roughly $40 billion GPU lease financing commitment with Apollo
Leased chipsSpaceX says it reached a binding commitment with an Apollo-managed entity for an approximately $40 billion master finance lease covering GPUs and related equipment. The hardware will support AI and compute services from SpaceX’s terrestrial data centers.
Who buys the equipmentUnder the arrangement, an Apollo-managed entity will purchase GPUs, servers and related equipment and lease them to SpaceX. SpaceX therefore does not have to rely entirely on its own cash to buy the equipment at delivery, but will owe payments under the lease. Apollo, Goldman Sachs and other parties have committed financing.
What the capacity is forSpaceX says the equipment will include Nvidia’s next-generation Vera Rubin chips, supporting Grok for businesses and consumers as well as existing and future cloud customers. The company has been expanding terrestrial capacity as demand for model training and user requests grows; the arrangement provides dedicated funding for additional equipment.
The stage todayThe binding financing commitment has been reached, but the lease is subject to closing conditions and is expected to close before the end of 2026. GPU and equipment deliveries are expected in the first half of 2027, when the corresponding lease schedules will begin. The full amount does not represent equipment already delivered or funding already spent.
▮ SIGNALFinancial institutions are increasingly buying AI equipment upfront and converting it into service capacity through long-term leases.
❯ Moonshot AI launches Kimi K3.1 Pro and plans to release weights by November 4
Available nowMoonshot AI has introduced Kimi K3.1 Pro for coding, 3D creation and professional tasks. Its release materials say the model and a faster Highspeed variant are available in Kimi products and through the API. Full weights are planned by November 4. Online availability and open weights are separate stages.
Understanding the specificationsK3.1 Pro has 2.8 trillion total parameters and uses a mixture-of-experts architecture that activates about 104 billion per inference. In other words, it selects part of its available capacity for each request. A one-million-token context window is intended to support long documents and multi-step tasks.
From answers to deliverablesMoonshot demonstrates game development, 3D modeling in Blender and CAD, and financial research. The company reports a 65.7% score on Terminal-Bench 4.0 for coding agents. That measures performance under specific test conditions and does not replace assessment on a user’s own project.
What open weights addPeople can already use the model through Kimi or its API; once the weights are released, developers will also be able to deploy and adapt them themselves. Long-document handling and tool use can connect research, design and development, while self-deployment offers an alternative to an online service.
▮ SIGNALModel competition increasingly turns on using professional software to deliver editable results from long tasks, alongside scale and benchmarks.
❯ Intel agrees to acquire quantum software company Classiq
Deal still pendingIntel says it signed a definitive agreement to acquire Israeli quantum computing software company Classiq. Intel did not disclose a price. The transaction is subject to regulatory and other closing conditions, is expected to close in the first half of 2027, and the companies will operate independently until then.
What Classiq sellsQuantum algorithms are expressed using quantum circuits. Classiq’s software lets developers describe a problem at a higher level and generates circuits for different hardware, reducing the need to build each circuit by hand. Classiq says it was founded in 2020 and has raised more than $200 million.
What Intel bringsIntel has spent years developing silicon spin qubits, chip manufacturing and cryogenic controls. Qubits are the basic units of information in quantum computing; Intel hopes to use semiconductor processes to support future chip production. Its existing hardware and software work would be combined with Classiq’s algorithm design platform.
After the dealIntel intends to continue supporting Classiq’s integrations across different hardware platforms. Developers could use software to abstract some hardware differences and adapt algorithms to available machines; Intel would gain a broader product stack spanning software and processors.
▮ SIGNALQuantum computing needs better qubits and tools that let developers who are not circuit specialists put them to work.
❯ Nevis raises $70 million to build AI tools for financial advisors
Funding and foundersWealth technology company Nevis announced a $70 million Series B led by EQT Ventures, with Charles Schwab, Sequoia Capital, ICONIQ, QuantumLight and Ribbit Capital participating. The company says it has raised $110 million in two years. Fortune reports, citing a person familiar with the deal, that the round valued Nevis at about $500 million; the company did not disclose a valuation. Its founders previously worked at Revolut.
Who pays for itNevis serves financial advisors and wealth management firms. In addition to advising clients, advisors prepare meetings, organize client information, open accounts and communicate with custodians. Nevis connects customer management, meeting notes, financial planning and custody workflows so AI agents can assist with repetitive work.
A practical workflowAfter a client meeting, an advisor can generate follow-up tasks and a financial plan from the conversation, then connect that work to account opening rather than re-entering information in multiple systems. EU-Startups reports that the funding will expand agent capabilities, integrations and product and sales teams.
ProgressNevis says its customers serve more than $700 billion in client assets and its agents have completed over 100,000 operational tasks. Those assets belong to customers’ clients, not to Nevis as an asset manager. Strategic investor Charles Schwab also provides a basis for connecting the product with its custody services.
▮ SIGNALFinancial AI assistants are moving from meeting summaries toward linking client information, account opening and follow-up service.
❯ ByteDance’s Jimeng launches AI short-drama studio
A single workspaceByteDance’s Jimeng AI has launched a short-drama studio that combines script preparation, character and scene management, storyboards, video creation and editing. IT Home reports that creators can manage assets around one project instead of repeatedly moving files between tools.
Start with the scriptThe studio’s DreamScript model supports original and adapted stories and can analyze a user’s uploaded text. It supports uploads of up to 100,000 Chinese characters and consecutive episode generation. Script diagnosis can review early episodes for the main plot, character identity and central conflict.
Build characters and shotsThe announced features include Anycook for images and DreamShot for storyboards. Creators can switch between a canvas and storyboard, set camera positions and preview movement in a 3D director interface. Characters, costumes and scenes can be reused throughout a project.
Revise laterThe product also supports reshooting individual shots, separating audio and generating replacement voiceovers. If one shot in an episode is unsatisfactory, a creator can adjust it without rebuilding the whole episode. Combining generation, asset management and editing is intended to support continuous production rather than a single image.
▮ SIGNALFor a series, reusable characters and scenes and individually editable shots matter alongside the quality of a single generated image.
❯ OpenRouter handles 188 trillion tokens in a week as DeepSeek leads its model chart
188 trillion tokensYicai reports that API aggregation platform OpenRouter handled 188 trillion tokens in the week of October 19–25, up about 14.6% week on week. Tokens measure information processed by models; the platform counts input and output tokens, not users or spending.
DeepSeek leadsDeepSeek V4.1-Flash-1017 processed 33.6 trillion tokens and led the weekly chart. Yicai says the model was updated on October 17 and that DeepSeek models have topped the platform’s weekly chart for three consecutive weeks. The model’s weekly volume rose 801%; the prior V4.1 Flash was second, followed by MiniMax M3.1 Flash, Zhipu GLM 5.3 Flash and Tencent Hy4 preview.
Why this platform mattersOpenRouter brings models from different companies behind a single interface, allowing developers to compare prices and switch between providers. Traffic movements provide a view into which models are being integrated and how developers distribute usage among them.
What the data coversYicai reports 93.54 trillion weekly tokens for Chinese models in the ranked set and about 22.5 trillion for US models. OpenRouter does not include providers’ own APIs, other cloud platforms, private deployments or direct consumer applications and websites. The figures describe the platform’s scope and cannot be treated as global market share.
▮ SIGNALAPI aggregators make switching easier, allowing capability, price and availability to appear more quickly in platform traffic.
❯ CXMT reportedly prioritizes domestic orders and declines to quote Apple
Supply reportJiemian, citing China-Singapore Jingwei and industry sources, reports that memory maker CXMT recently declined to quote Apple for DRAM supply while prioritizing domestic demand. CXMT had not responded directly when the report was published. A lack of a quote does not establish a signed deal or termination of an existing supply agreement.
Earlier price discussionsThe report says earlier coverage in August described talks between Apple and CXMT over prices for mobile memory including LPDDR5X, with a request for a discount not accepted. LPDDR is low-power DRAM used in devices such as phones; price and available capacity both influence a device maker’s procurement decisions.
Domestic product progressCXMT has supplied memory to domestic device manufacturers. In a September announcement, it said its LPDDR6 chip had entered commercial mass production and debuted in Xiaomi’s 18 Fold. The company cites peak transfer speeds of 12,800 Mbps and improved bandwidth and power consumption versus the prior generation in internal tests.
Allocating capacityJiemian cites market views that long-term domestic device-maker orders occupy substantial capacity, giving them priority in allocation. For CXMT, this concerns how to distribute existing output among customers; for device makers, it affects the availability of memory and procurement choices.
▮ SIGNALMemory competition is shaped by product specifications and by how suppliers allocate scarce capacity among long-term customers.
❯ XPeng wins Guangzhou L4 demonstration qualification as Volkswagen tests L3 in Hefei
XPeng moves toward demonstrationsXPeng says its robotaxi business has received qualification for L4 demonstration operations across Guangzhou. Initially, access will continue to expand gradually through invitation codes, with service widening as operations progress.
Vehicle and earlier workXPeng previously introduced a GX-based robotaxi in Guangzhou and progressively conducted road and user-experience tests. It says the vehicle uses four in-house Turing AI chips and a second-generation vision-language-action model. Demonstration operations will let it test dispatch and passenger service alongside driving capability.
Volkswagen tests L3CnEVPost reports that Volkswagen China obtained an L3 road-test permit in Hefei and began testing and collecting road data. Volkswagen previously set a goal of achieving L3 technology readiness in 2027. A road-test permit is not approval to offer the feature broadly in production vehicles.
Different levelsAn L3 system performs driving tasks under specified conditions but expects the driver to take over when requested. L4 systems handle the driving task within defined conditions and areas. Volkswagen’s passenger-car project and XPeng’s robotaxi service have different permissions and use cases.
▮ SIGNALMoving from demonstrations to deployment requires testing road performance, service operations and passenger experience together.
❯ Apple unveils iPad mini with OLED display and A20 Pro chip
A small tablet refreshApple has introduced a new iPad mini with an 8.5-inch OLED screen, A20 Pro chip, a thinner and lighter body, and IP67 water and dust resistance. US pricing starts at $699; preorders begin October 28 and availability starts November 6.
Screen and hardwareThe screen is larger than its predecessor and brings OLED to iPad mini for the first time. Apple says peak brightness is more than double the previous generation, and offers an optional nano-texture finish to reduce reflections. The front camera is now on the landscape edge for video calls; the device also supports faster wireless connectivity and fast charging.
Chip and AIA20 Pro adds neural accelerators to the GPU and dual 16-core Neural Engines for on-device AI computation. The tablet runs iPadOS 27 and supports Apple Intelligence and Siri AI features, including requests that draw on on-screen information and personal context. Availability varies by region, language and settings.
Who it servesThe iPad mini’s size makes it easy to carry and use one-handed for reading, field work, games and video calls. The update brings a stronger screen and processor to that form factor, while water and dust resistance adds durability away from a desk.
▮ SIGNALApple is bringing local AI processing, a better display and greater durability together in a portable device designed for more tasks.