❯ Anthropic posts $16.8 billion in Q3 revenue as growth slows, and some investors reportedly lower IPO valuation expectations to about $1.5 trillion
Growth slows to 45%At an investor meeting in San Francisco on Wednesday, Anthropic said third-quarter revenue was $16.8 billion, up 45% from the second quarter, and that it was profitable on an adjusted basis. According to CXO Digital Pulse, relaying Reuters, second-quarter revenue was $11.6 billion, more than double the first quarter, so growth has slowed markedly.
Talking down from $2 trillionReuters, citing four people familiar with the matter, said some current and prospective investors are reassessing the IPO valuation, with some discussing about $1.5 trillion, below the roughly $2 trillion previously expected for a November listing. These are discussions among investors, not a set offering price. Anthropic did not immediately respond to a request for comment.
Revenue next to costsAnthropic’s revenue grew about twelvefold last year to nearly $4.6 billion, its prospectus showed a $42 billion loss for 2025, and it has committed to spending at least $518 billion over a decade on cloud services, compute and infrastructure. Investors are weighing growth against that long-term bill. Its listing will also shape expectations for a possible OpenAI IPO next year.
▮ SIGNALLeading model companies used to be valued mainly on how fast they grew; as listings approach, investors are asking whether that growth can cover the compute bills already signed.
❯ Meta releases Muse Spark Open weights and works with Apple to run the 816-billion-parameter model locally on two Mac Studios
The top US open modelMeta released the weights for Muse Spark Open on October 15, calling it the most capable US open-weight model to date. It is a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active and a 1 million-token context window, built for coding and agentic work. The weights are downloadable from Hugging Face and other platforms under a license that allows limited commercial use.
Two Macs as one supercomputerMeta worked with Apple so the model can run on two M5 Ultra Mac Studios. Meta explains that each machine’s GPU can use up to 512 GB of unified memory, so two can hold the whole model. According to CNBC, users can pair two to four Mac Studios to code or analyze video with it. Mark Zuckerberg thanked Apple CEO John Ternus in a post on X, and Ternus replied with congratulations.
Old rivals at the same tableCNBC notes that relations between the two hit bottom after Apple’s 2021 privacy changes, which Meta estimated would cut its revenue that year by $10 billion. The collaboration took months, and neither side would say whether it will go further. For developers, the main change is that a model near the frontier can run on their own desk without the cloud, though it needs top-spec Mac Studios that are not cheap.
▮ SIGNALAs open models grow, data centers are no longer the only place that can hold them, and personal computers’ memory capacity has become a reason for model companies to pick partners.
❯ Microsoft finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model it says is 4x to 12x more compute-efficient than DeepSeek-V4-Pro-Base
A bigger, sparser baseMicrosoft AI announced that it has completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active. Microsoft says it beats every open-weight model it tested on its held-out evaluations and is 4x to 12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base, with the savings put into training a larger model.
38 days on 16,000 GB200sMicrosoft says MAI-Base-2 was trained on 16,000 GB200 GPUs over 38 days in August and September, with training progressing 89% of the time. In June it released MAI-Thinking-1, post-trained on the previous base, MAI-Base-1. The new model is still in post-training and not yet available to use.
Clean data as the pitchMicrosoft stresses that the model’s pre-training data is organic and involves no distillation from other models, giving it clear provenance. It plans to license the weights to trusted partners through Foundry. Microsoft wants to make traceable training data a reason to choose it over open models, especially for enterprise customers who worry about copyright and compliance risk.
▮ SIGNALCompanies building their own large models now pitch “we didn’t distill anyone” as a feature, adding a “where did the data come from” test to enterprise procurement that open models do not always pass.
❯ Nvidia releases the open Nemotron 3.5 Super VL, with up to 5x the agentic throughput of similar-sized models on a single GPU
One model for text and imagesNvidia released Nemotron 3.5 Super VL on October 15, an open, customizable multimodal model. Its mixture-of-experts architecture has 120 billion parameters and activates 12 billion per token, combining reasoning, tool use and understanding of text, images, documents and video in one model, so companies need not deploy separate text and vision models.
5x throughput on one GPUNvidia says that on a single GPU it delivers up to 5x the agentic throughput of other open models in its size class. Companies including Scale AI, Ramp and AI21 have post-trained it on their own data, and Nvidia cites results such as 40x lower inference cost, 10x lower latency and 40% higher accuracy; these figures are Nvidia’s own.
The chip seller makes models tooDevelopers can keep training it with Nvidia’s NeMo toolkit and open datasets on Hugging Face, and the model is optimized for inference on Nvidia platforms such as DGX Spark. The same day, Meta released the larger Muse Spark Open. Companies have more open models to choose from, and Nvidia’s approach ties its model to its own hardware and software so customers do their customization on its platform.
▮ SIGNALBy making its own open models, Nvidia is really selling a path to customization that runs fastest on its chips, turning the model into a sales tool for its compute platform.
❯ TSMC posts NT$763.1 billion in Q3 net profit, a sixth straight record, and confirms Terafab talks while insisting it must run the fab
Profit up nearly 70%TSMC reported third-quarter results on October 15. According to CNBC, revenue was NT$1.49 trillion (about $46.36 billion) and net profit NT$763.1 billion, up 68.7% year on year, both above estimates and a sixth straight quarterly profit record. Operating margin was 58.7%, above the top of its guidance.
3nm is a third of revenue3-nanometer chips accounted for 33% of wafer revenue, the largest share, and 7nm and more advanced nodes together made up 79%. By platform, high-performance computing was 68% of revenue. The company expects fourth-quarter revenue of $52.3 billion to $54.3 billion and capital spending in 2026 to 2028 significantly higher than in the previous three years. A SemiAnalysis analyst told CNBC that TSMC says demand remains well above supply.
New fabs and TerafabAccording to Finimize, CEO C.C. Wei said on the earnings call that a new 3nm plant in Tainan will reach volume production in the first half of 2027, another in the first half of 2028, and the Arizona site in the third quarter of 2027. He confirmed TSMC has sent teams to Texas to evaluate Musk’s Terafab project but repeated that any alliance requires TSMC to operate the fab. Musk has said his companies will build and run the fab, with TSMC subleasing a section at most.
▮ SIGNALThe value of leading-edge manufacturing lies in process know-how, not buildings, and TSMC would rather pass on a new fab’s capacity than hand over operational control.
❯ ASML’s equipment sales to China rose 109% quarter on quarter in Q3, and it expects China to be about 20% of full-year sales
China sales more than doubleLithography maker ASML reported third-quarter results on October 14 with record net sales of €11.7 billion. According to TechWire Asia, China’s share of its system sales, measured by shipment destination, rose from 14% in the second quarter to 23%, with China sales up 109% quarter on quarter.
Full-year outlook unchangedASML kept its expectation that China will account for about 20% of total 2026 sales. CEO Christophe Fouquet said domestic demand in China for mainstream logic and memory is very strong, with new fabs coming online. According to CNBC, ASML also raised the low end of its full-year sales forecast from €43 billion to €44 billion, for a range of €44 billion to €45 billion.
Orders out to 2030Fouquet said some customers want to place orders for 2029 and even 2030, that planned supply for 2027 and 2028 will not fully meet demand, and that the company is looking at ways to expand beyond its previously forecast 30% capacity increase. During the quarter, Samsung and TSMC both committed to ASML’s next-generation High NA EUV machines.
▮ SIGNALWith equipment orders stretching four years out, the pace of chip capacity growth now depends more on how fast upstream machines can be built than on how many fabs chipmakers want.
❯ Tim Cook says Apple Intelligence is approved for the Chinese market, with a test version due later this month
Translation and writing firstAccording to China Daily on October 15, Apple Executive Chairman Tim Cook said in an interview in Shanghai that Apple Intelligence has been approved for the Chinese market, with a test version later this month. The first features include real-time translation, writing tools and image cleanup; the AI-powered Siri is not in the initial rollout.
The filing is doneThe report says that the day before Cook’s remarks, the Ministry of Industry and Information Technology updated telecom equipment network-access records showing that the China model of Apple’s first foldable, the iPhone Duo (A3721), completed a license renewal on September 29 that added a filing number for its built-in large language model. Three models were newly filed in this batch, including Apple Intelligence and Alibaba’s Tongyi models.
A key step for China usersChina Daily calls the completed model filing a key step for Apple in advancing AI features in the Chinese market. Once the test version arrives, users of China-model iPhones will get everyday features such as translation and writing first; Apple has not said when Siri’s AI capabilities will follow.
▮ SIGNALFor a foreign phone maker bringing AI to China, the path runs through model filing first and phased rollout second, making local compliance part of the product launch timetable.
❯ DeepSeek refunds all API users seven days of spending after restoring service and says it has removed single points of dependency
Restored after nearly five hoursAccording to ITHome, DeepSeek’s web and API services failed on the morning of October 15, and the API was unavailable from 8:11 to 13:04 Beijing time, about five hours. DeepSeek’s developer platform then posted a notice apologizing and saying it had launched a full review and overhaul of service stability.
A week of usage refundedAs compensation, DeepSeek is refunding each account’s spending from October 8 to 14 in full as credits, which arrive automatically and are drawn down first for subsequent API charges. The notice says the company will comprehensively review and remove single points of dependency so that similar problems do not recur.
Developers care about uptimeAPI users are mostly developers who build the model into their own products, so when the service stops, their applications stop too. DeepSeek chose to refund actual usage in full rather than compensate only for the outage window, tying the size of the refund directly to each user’s usage. Stability after the overhaul will be proven by how the service runs from here.
▮ SIGNALWhen developers bet their products on someone else’s model API, providers compete not only on price and capability but on how they make things right when something breaks.
❯ COL Group says AI short dramas and video generation account for 55% of China’s token consumption, the largest use
Three uses take nine-tenthsA-share listed COL Group said at an investor-relations event that, according to industry token-usage data it cited, AI short dramas and video generation account for 55% of China’s AI token consumption, e-commerce and marketing 24%, and software development 15%. According to MyDrivers, the figures have drawn renewed attention in recent days. On these numbers, short dramas and video generation consume more than twice as many tokens as software development.
Why short dramas use so muchCOL Group explains that short dramas demand fast video generation and turnaround. Live-action-style AI dramas have stricter frame-rate and 1080p requirements and typically use two to three times the tokens of 720p comic-style AI dramas; production teams also generate and screen shots over and over, known as “pulling cards,” which keeps consumption high.
Market size drives demandA report by the research center of the National Radio and Television Administration and others shows China’s micro-drama market topped $14 billion in 2025, about 78% of the global market. According to NetEase, COL Group said about 1,500 new AI short dramas launched per day in China in the first quarter, and it expects the global short-drama market to exceed $100 billion around 2030, with China about half.
▮ SIGNALChina’s largest AI usage is in content production rather than coding, and the shape of compute demand follows the country’s most mature consumer market.
❯ Figma launches Figma Specs, a Markdown document that sets the design rules AI agents must follow
A design handbook for agentsDesign platform Figma launched Figma Specs at its first Config event in India, a Markdown editor that lets users write down a company’s design language in plain text. The document serves as a reference for teams and for Figma’s own and third-party AI agents, and can reference canvas elements and embed FigJam diagrams and code layers.
Making AI output more predictableFigma opened its on-canvas agent to more users in the past week. Because AI output is non-deterministic, the company is providing tools to steer agents toward more predictable results: teams can centralize design libraries, code libraries and guidelines, and agents follow that system. Code layers are also in open beta, letting users ask the agent to turn a design into code or change code to iterate on a prototype.
More code, not more releasesFigma chief design officer Loredana Crisan said 17 times more code is produced today than in the past, but releases are up only 1.3 times, because much of that code is written to discover what the product should be. Figma’s MCP now supports more than 20 new tools, including Grok Bot and Indian AI company Sarvam. The position design tools are competing for is the step where teams decide what to build.
▮ SIGNALOnce AI makes writing code cheap, the scarce part becomes deciding what to build, and design tools are using that to place themselves upstream in agent workflows.
❯ Accenture sets the company-paid Claude budget at $29 per employee per month, with overages charged to project budgets
The company pays $29According to RuntimeWire, citing 404 Media, Accenture began centrally funding $29 per employee per month for Claude Enterprise on October 1. Employees who exceed that amount must charge additional use to a project or team budget.
Not a hard capThe policy sets no universal technical cutoff: employees can seek project funding, while access and licenses remain subject to review. According to an internal memo seen by 404 Media, as relayed by Techmeme, some employees had been spending as much as $1,000 a month on Claude. Accenture had previously announced a large Claude rollout.
The AI bill goes back to projectsCharging overages to projects means every heavy use needs a business reason, moving the spending decision from individual employees to project owners. Usage-billed tools such as Claude Code will see their spread inside large companies directly constrained by budget rules like this.
▮ SIGNALCorporate enthusiasm for AI tools is meeting the finance department, and usage-billed products now have to convince not only employees but the people who hold the budget.
❯ OpenAI reportedly trains staff in counterespionage, warning of honey traps, bribes and eavesdropping in public places
A friendly stranger may be a spyAccording to Yahoo News, relaying a Wall Street Journal report citing unnamed sources, OpenAI’s internal training warns employees about “honey pot” schemes that use sex or romance to get close to a target. One example: a stranger approached an employee at a San Francisco playground with their child and steered the small talk toward work; the stranger was a spy.
Everyday secrecy rulesOpenAI’s security experts also warn about bribery and eavesdropping in public places such as cafes, and tell staff to use privacy screens on phones and laptops in public, wipe whiteboards after meetings and avoid wearing company-branded clothing in public. The report says Anthropic and Google have given employees similar warnings.
After the bugging caseThe report notes these risks took on new urgency this week after the US Justice Department charged an OpenAI engineer with installing a hidden recording device in an executive conference room. Frontier labs’ weights, training methods and product plans are enormously valuable, and the cost of protecting them is extending from cybersecurity into employees’ daily behavior.
▮ SIGNALOnce model companies’ secrets are worth the effort of stealing, AI labs’ security practices start to converge with those of defense contractors and intelligence agencies.
❯ UK AI compute company Nscale heads for an IPO at up to $35 billion as its flagship Norway site runs six months late
$3.36 billion raised before listingAccording to Techstrong IT, citing Bloomberg, UK AI infrastructure company Nscale is preparing to list at a valuation of up to $35 billion after raising about $3.36 billion in pre-IPO funding. Nvidia provided $1 billion in financing to secure demand for its own processors.
A neocloud renting out computeNscale builds data centers and rents compute to AI companies, a model known as a neocloud. The report says hyperscalers and research labs increasingly rent infrastructure from third parties to turn long-term capital risk into flexible operating expense. Nscale reports $103 billion in contracts.
Norway late, Britain short of powerIts flagship site in Glomfjord, Norway, near the Arctic Circle, is planned at 30 megawatts but is about six months behind schedule due to cooling leaks, regulatory hurdles, construction delays and a fire; only 6.4 MW is operating. Its project in Loughton, Essex, needs 90 MW of power and was still a scaffolding yard in March, with connection delays affecting progress. Ofgem said in June that 315 data centers were waiting in grid connection queues.
The question for IPO buyersThe report notes the model relies heavily on debt, shifting overcapacity risk from cloud giants to companies like Nscale. Investors must decide whether the $103 billion in contracts is a long-term revenue stream or a windfall from today’s compute shortage. Australia’s Firmus withdrew its IPO just last week.
▮ SIGNALCompute companies’ valuations are written into contracts while delivery is held up by grids and construction sites, and every delay before listing gets priced in by investors.
❯ RapidCanvas raises about $29 million in Series B funding to build custom AI applications for businesses with AI agents
About $29 million in Series BAccording to Dealroom, relaying Axios Pro, Austin-based RapidCanvas raised about $29 million in a Series B, CEO Rahul Pangam said. Axios’s report, as relayed by Techmeme, says Vertex Ventures and Hitachi Ventures led the round.
AI for companies short on talentRapidCanvas uses AI agents to help businesses build and deploy custom AI applications, with its agentic platform handling the development and deployment work. Its target customers are enterprises that want custom AI but lack the in-house talent to build it.
Next round possibly late next yearPangam said the company could raise a Series C by late next year as the platform evolves. Specific uses of the money were not disclosed.
Selling the last mileRapidCanvas is not selling a model but the step of turning models into applications a business can use. What it has to prove is that companies without technical teams need a provider that builds and deploys the application for them more than they need a stronger model API.
▮ SIGNALThe hard part of buying AI is often deployment rather than the model, and the middle layer that “builds it for you” can still raise money.
❯ Binance.US is reportedly seeking up to $100 million at about a $1 billion valuation, down from $4.5 billion in 2022
Under a quarter of the last priceAccording to Cointelegraph, citing Bloomberg, crypto exchange Binance.US is seeking to raise up to $100 million privately at about a $1 billion valuation to expand its US business. When it raised $200 million in 2022, its pre-money valuation was $4.5 billion.
Trying to win back shareThe report says the round is part of Binance.US’s effort to regain market share from larger rivals after parent Binance’s 2023 settlement with US authorities, under which Binance Holdings agreed to pay $4.3 billion in penalties and accept additional monitoring.
World Liberty sits outBinance.US informed crypto platform World Liberty Financial about the round, but it will not participate. The drop from $4.5 billion to about $1 billion reflects Binance.US’s position after losing US market share; whether it can expand again depends on whether new money brings back users and trading volume.
▮ SIGNALAn exchange still raising money after its valuation fell nearly 80% shows that in the US crypto market, a license and a compliance record remain valuable assets in themselves.
❯ Tesla unveils the next-generation Roadster, with the fastest version using cold-gas thrusters to hit 60 mph in under 0.59 seconds
Thrusters that make it jumpTesla unveiled its next-generation Roadster on the evening of October 15 in Texas. According to Electrek, the car has been “cyberfied” and fitted with cold-gas thrusters that let it jump and flip. The livestream featured impressive stunts but still no price or production date.
Three versionsAfter the event Tesla posted three configurations on X: the road-ready Performance does 0-60 mph in under 1.6 seconds; the race-ready Thrust adds rear cold-gas thrusters for under 0.9 seconds; and the single-seat Rocket, with thrusters around the car, does under 0.59 seconds with a top speed above 230 mph and is not approved for public roads. Tesla quotes range above 300 miles and says the thrusters were developed with SpaceX.
Nine years lateTesla first showed the car as a prototype in November 2017 and promised production in 2020, followed by repeated delays, including several postponements of the reveal this year alone. Electrek notes the range promise is roughly half what was claimed nine years ago and two seats are gone, while Tesla collected deposits of $50,000 to $250,000 from customers back then.
▮ SIGNALA sports car with no price or production date was staged as a technology show, displaying less a car than the engineering Tesla and SpaceX can share.
❯ TIME names its 300 Best Inventions of 2026, with new Apple CEO John Ternus on the cover
300 inventions on the listTIME published its Best Inventions of 2026, naming 300 inventions across fields including artificial intelligence, accessibility and agriculture. The AI category includes high-density cooling for multistory data centers and an autonomous AI agent “that can’t break containment.”
Ternus on the coverThe cover subject is John Ternus, who took over as Apple CEO on September 1. TIME’s interview centers on Apple’s first foldable phone, the iPhone Duo, priced from $1,999. The piece describes Ternus as a detail-oriented engineer who was little known outside the company before becoming a senior hardware executive in 2021.
AI tools on the list tooGoogle DeepMind announced that SynthID, its AI content watermarking technology, made the list, having watermarked more than 180 billion images and videos to help people tell AI-generated content from other media. TIME says all rankings are made by its editorial team; some companies may have paid an application fee to be considered, which does not affect inclusion.
▮ SIGNALOn the year’s inventions list, tools for identifying and constraining AI sit alongside AI itself, and keeping AI in check has become an innovation category of its own.