❯ OpenAI tells investors annualized revenue is about $50 billion, below the $68 billion figure that had circulated, and Nvidia, Oracle and CoreWeave shares fall

Nearly $20 billion shortOpenAI said in an investor presentation that its annualized revenue was roughly $50 billion at the end of September. CNBC confirmed the figure on October 8 and said it was lower than the $68 billion widely reported late last month. The Financial Times reported it first. Annualized revenue extrapolates the current pace over a year; it is not money already booked.

Two numbers, two definitionsA person familiar with the matter told CNBC that the $68 billion figure included gross revenue from OpenAI’s partners, to help investors make a more direct comparison with Anthropic. Anthropic told investors in August that its annualized run rate hit $65 billion at the end of July. In the same presentation, OpenAI cited 77% total run-rate growth in the third quarter and 107% for its enterprise business.

The compute chain sold off togetherAfter the report, Nvidia fell 3% on Thursday, Oracle nearly 6% and CoreWeave nearly 8%; AMD and Broadcom each lost 4% and Intel 5%. All of them are chip and compute companies. The market ties their order outlook to the revenue of the leading model companies, so when OpenAI’s number moves, this supply chain is repriced first.

Valuation pressure before the IPOCNBC put OpenAI’s valuation at $852 billion and said it confidentially filed a prospectus in June, with executives signaling a 2027 debut; the company is also in early talks on a new round of about $30 billion, with no term sheet finalized. Before a listing, revenue has to be examined on one consistent basis, and the wider the gap between circulated and self-reported figures, the weaker the company’s hand on pricing.

▮ SIGNALThe two leading model companies are hard to compare because nobody agrees whether sales made through cloud providers count as their own; prospectuses will force them onto the same arithmetic.

❯ OpenAI opens Ultrafast mode for GPT-6.1 Sol, up to 8x faster at six times the standard API price

Same model, different tierOpenAI’s developer account announced on October 8 that Ultrafast mode for GPT-6.1 Sol was rolling out that day in the API, Codex and ChatGPT Work, at speeds up to 8x Sol Standard and with near-Astra intelligence. Ultrafast is not a new model. It is a service tier that makes the same model return output faster.

Speed costs five times moreAPI pricing is $12 per million input tokens and $60 per million output tokens. According to RuntimeWire, Sol’s standard rates are $2 and $10, so Ultrafast is exactly six times as much. OpenAI named three uses: debugging an outage, agents navigating apps, and live experiences where every second counts.

Subscribers need Pro 500The API tier is open to all developers. In Codex and ChatGPT Work it is limited to Pro 500, eligible usage-based Enterprise and credit-based Edu plans, and Enterprise administrators must turn it on. The same day, Codex lead Thibault Sottiaux said steering is now instant: when a user changes direction mid-task, the model adjusts right away instead of wasting effort.

Which requests deserve the fast laneDevelopment teams now have to sort their requests into tiers. Agents make successive calls to models and tools, and every wait compounds, so that is where speed pays most; background jobs nobody is watching do not need the fivefold premium. The price gap is large enough to make teams decide seriously which work has a person waiting at the screen.

▮ SIGNALAs model capabilities converge, speed has become a product with its own price tag, and OpenAI is betting developers will pay more to wait a few seconds less than to get a slightly smarter answer.

❯ Google Cloud launches an enterprise Gemini agent that works for days, has its own email address and can run on Claude

Objectives, not instructionsGoogle Cloud introduced the Gemini agent at its Gemini at Work event on October 8, positioning it as the single universal agent inside a company: answering questions, handling documents, generating images and writing code from one prompt box. Google Cloud CEO Thomas Kurian said users give it objectives, not instructions.

Still working after the laptop closesThe agent runs in the cloud with one set of memory and context, so work that takes hours or days continues after the user logs off. It can spin up temporary sub-agents to split a job, or act as a standing “coworker agent” with its own @agents company email and persistent storage. According to TechCrunch, colleagues can email it or call it into a group chat, and its actions are logged under its own name.

Claude inside GeminiGoogle separates the agent from the model: by default the system picks a model for each job, and users can also choose one themselves. The first third-party option is Anthropic’s Claude, with open-source and other private models to follow. The agent connects to Workspace, Microsoft 365, Slack, Jira, Snowflake and other systems, and to any MCP server.

Businesses first, consumers laterGoogle says nearly 90% of the Fortune 100 use Gemini Enterprise, so it is launching for businesses first and consumers later. That gives corporate IT departments another option: instead of buying a separate agent for each team, one entry point manages permissions, audit trails and spending caps, and context and skills stay in place when models change.

▮ SIGNALGoogle is willing to put a rival’s model inside its own agent, which shows it believes enterprise customers ultimately commit to the layer where the entry point and the data live, with the model as a swappable part.

❯ Anthropic launches OSS Scanner, using its strongest models to scan open-source projects for free and sending reports to maintainers without human review

Free, opt-in and recurringAnthropic launched OSS Scanner on October 8, a vulnerability-scanning service for open-source software. Once a project opts in, Anthropic scans its code periodically with its strongest models, including Claude Mythos, at no cost. Each report includes a self-contained reproducer, an explanation of the vulnerability and, where available, a candidate patch.

More findings than reviewersOver the past six months Anthropic’s models found more than 29,000 candidate vulnerabilities in major open-source projects, and humans have only been able to review about 6,000. Many maintainers, after receiving the first reports, asked for everything, validated or not, and nearly 5,000 reports have gone out that way. OSS Scanner makes that fast track permanent: reports are fully model-generated with no human triage, and Anthropic acknowledges some may be wrong.

One false positive in 97To check quality, Anthropic asked expert penetration testers to review 97 critical and high-severity findings from the scanner across 48 projects: 85 met the bar for formal disclosure, 11 were real but duplicated known issues, and only one was a false positive. A maintainer of the cryptography library wolfSSL said all but two of the 74 reports it received were valid and five became CVEs.

Fixing faster than attackersAnthropic’s reasoning is that exploits can now be written in minutes, so whoever finds a flaw first has the edge. Core maintainers enroll by submitting a request to a designated GitHub repository; eligibility follows criteria similar to Google’s OSS-Fuzz, requiring critical impact on infrastructure and user security. The pressure shifts to maintainers: reports arrive faster and in greater volume, while verifying and merging patches still takes people.

▮ SIGNALWith models pushing the cost of finding bugs toward zero, the bottleneck in open-source security has moved from nobody reading the code to nobody having time for the reports; maintainer attention is the scarce resource.

❯ Manus is reportedly valued at $4 billion after its $500 million-plus round, the highest among AI application companies with a China background

The missing valuation arrivesWhen Manus parent Butterfly Effect announced a round of more than $500 million on October 8, it disclosed no valuation. LatePost reported the same day that Manus now has a post-money valuation of $4 billion, making it the most highly valued AI application company with a China background, with the most cumulative funding and the largest cash reserves. Boyu Capital and IDG Capital led the round, with existing shareholders Tencent, HongShan (Sequoia China) and ZhenFund participating again.

An agent that finishes the jobManus launched in March 2025 and went viral on a demo of its AI agent. As TechCrunch describes it, the company offers a chatbot and tools for building apps and websites in natural language, and can also produce designs, presentations and video. When Meta announced its acquisition in December 2025, Manus was reportedly bringing in more than $100 million in annual recurring revenue.

Twice the last priceMeta’s purchase price was about $2 billion; the deal was not completed, the founding team and existing shareholders bought the shares back from Meta, and the company said on September 1 that it had resumed independent operations. On LatePost’s figure, the new valuation is double that earlier price. According to SiliconStar, it is also the largest single round to date in China’s AI application sector.

More cash, more rivalsSiliconStar notes that Manus released Manus 2.0 and a personal agent app called Cue on September 28, is developing a product for the Chinese market, and has opened 17 positions. Leading on cash reserves means it can outlast comparable application companies on hiring and compute; but an agent product has to call other companies’ large models, and the more users it has, the more it pays model providers.

▮ SIGNALAn application company’s valuation doubled after an acquisition that did not close, which shows investors no longer need a big-tech buyer to anchor the price of the agent layer.

❯ SemiAnalysis estimates China has more than 24 GW of delivered data center capacity, with ByteDance alone taking about a fifth

Counted across a thousand sitesResearch firm SemiAnalysis estimated in its China Datacenter Model, published September 25, that China has more than 24 GW of delivered data center capacity. According to FourWeekMBA, the figure comes from tracking more than 1,000 facilities across over 60 operators. The Financial Times cited the data on October 8. A gigawatt is a unit of power, used in the industry to gauge how much computing equipment a data center can run.

Another 50 GW in the queueBeyond the 24 GW, the model separately counts about 20 GW of dated pipeline and about 30 GW of announced projects. The same report puts US capacity at 56 GW as of the end of 2026, Asia-Pacific excluding China at about 15 GW, and Europe, the Middle East and Africa at about 14 GW combined. SemiAnalysis also notes that published estimates of China’s capacity differ by as much as 15x.

ByteDance is the biggest tenantUnder the model, ByteDance occupies roughly a fifth of the country’s delivered capacity and rents nearly all of it. The report says combined capital expenditure at Alibaba, Tencent and Baidu reached $20 billion in the second quarter, more than doubling year on year; with ByteDance included, the four are on a trajectory to $100 billion for the year.

Vacancy and shortage at onceThe report describes two markets: older retail racks around tier-one cities run at 30% to 50% utilization and often cannot support AI power densities, while new wholesale capacity for AI is being built elsewhere and filled quickly. For operators, old racks may not catch this demand; the orders go to those who can secure power and land and deliver a new 100 MW facility in under a year.

▮ SIGNALComparing compute now means counting powered buildings and not just chips, but building data is harder to gather, and a 15x spread between estimates is itself a piece of information.

❯ Isomorphic Labs is reportedly in talks for new funding at a $40 billion to $50 billion valuation

A valuation before a deal sizeAccording to SiliconANGLE, citing a Bloomberg report on October 8, Alphabet spinoff and AI drug discovery company Isomorphic Labs is in talks for a new funding round at a valuation of $40 billion to $50 billion. The talks are at an early stage, and neither the size of the round nor the investors have been disclosed.

Finding where a drug can landIsomorphic’s flagship product is IsoDDE, an AI platform that speeds up the manual research steps in developing a therapy. A drug usually has to attach to a specific point on a disease-causing cell, known as a binding pocket, and finding one is slow and difficult. The company said in February that IsoDDE was more than 50% better at finding binding pockets than the closest competing tool and about twice as accurate as AlphaFold at predicting how a drug and its target fit together.

Revenue from drugmaker partnershipsThe company earns money through research partnerships with drugmakers. In 2024 it signed deals worth nearly $3 billion with Eli Lilly and Novartis, and this January it launched a similar collaboration with Johnson & Johnson; all three focus on small molecules. Its last raise, a $2.1 billion round announced in March, included Alphabet, two of its venture funds and several outside backers.

Valuation running ahead of the medicineOnly a little over six months separate the $2.1 billion round in March from the tens of billions now being discussed. Investors are betting that platforms like IsoDDE can sharply cut the time needed to screen drug candidates; for partners such as Lilly, Novartis and J&J, the more capital favors the platform, the harder it becomes for latecomers to get the same terms.

▮ SIGNALAI drug discovery is now valued on the scale of early rounds at leading model companies, but its proof arrives on the timetable of clinical trials, which will test investors’ patience far longer than software does.

❯ Model leaderboard Arena raises a $200 million Series B at a $3.1 billion valuation and releases an Alignment Index measuring whether models behave

Valuation nearly doubles in ten monthsAI model evaluation platform Arena announced a $200 million Series B on October 8 at a $3.1 billion valuation. Lightspeed and Khosla Ventures co-led, with existing investors including a16z participating. According to TechCrunch, its Series A in January was $150 million at a $1.7 billion post-money valuation.

Users vote, models get rankedArena began in 2023 as a research project at UC Berkeley. Ordinary users enter prompts or request coding projects for free and judge which model did better, and those votes feed the leaderboard. In September last year it introduced a paid product, AI Evaluations, selling model labs and enterprises detailed performance analytics based on community feedback. The company says annualized revenue has passed $100 million, up from $30 million in January.

A new board for honestyReleased alongside the funding, the Alignment Index measures whether a model acts in line with what the user intended, looking at three problems: taking actions it was not asked to take, attributing statements to the wrong source, and claiming to have completed tasks it did not do. TechCrunch said a slate of OpenAI models leads the preliminary board, with Claude Opus 5.5 in sixth place.

Fixed exams no longer sufficeArena argues that static benchmarks break down once models recognize they are being tested, and that when agents write code or run analyses for people, users often cannot check the work. Enterprises buying models therefore need evaluations drawn from real use. Arena’s own difficulty is staying neutral: its paying customers are the same labs it ranks.

▮ SIGNALAn evaluation company more than tripling revenue in a year shows vendor-published scores are no longer enough to buy on, and refereeing is turning into a business of its own.

❯ Keyu Tian’s world-model lab reportedly raises close to $30 million at a $200 million post-money valuation, with no name or product yet

Money before a nameAccording to RuntimeWire, citing a Bloomberg report on October 7, researcher Keyu Tian has raised close to $30 million for his new AI lab at a $200 million post-money valuation, with backers including 5Y Capital and IDG Capital. Bloomberg said the lab has no public name or product.

First author of an award-winning paperTian completed a master’s degree at Peking University in summer 2025. He was first author of “Visual Autoregressive Modeling,” a 2024 paper by researchers affiliated with Peking University and ByteDance that NeurIPS named a Best Paper. Its method generates images from coarse to fine, starting at low resolution and stepping up.

World models must grasp physicsThe new lab is pursuing world models, AI systems that represent and predict how physical environments change. Bloomberg cited robotics, interactive video and autonomous driving as potential uses, none of them demonstrated products. Producing a convincing still image and simulating a coherent environment that changes as a camera moves or objects collide are two different levels of difficulty.

The benchmark is Fei-Fei LiBloomberg frames Tian against Fei-Fei Li. Her World Labs has released Marble, a model that generates navigable 3D worlds, said in February it raised $1 billion, and on September 28 announced a definitive agreement to join AMD. Tian’s round, by contrast, is a bet placed before any product exists; investors are buying his research record and his ability to assemble a team.

▮ SIGNALA single best-paper award can now support a $200 million valuation, which shows capital in world models is arriving at the stage where there are people but not yet a model.

❯ Quantum computing company Oratomic raises a $475 million Series B, lifting its valuation from $1.5 billion to about $5.4 billion in three months

Two rounds in three monthsCalifornia quantum computing company Oratomic announced a $475 million Series B on October 8. According to SiliconANGLE, the round comes three months after its $300 million Series A and was co-led by six firms including ARCH, Spark Capital and Khosla Ventures, bringing total funding to $775 million and lifting the valuation from $1.5 billion to about $5.4 billion.

Doing it with ten thousand qubitsFounded in 2025, Oratomic builds qubits, the basic computing units of a quantum computer, from arrays of neutral atoms. Qubits are extremely fragile: slight disturbances in heat, electromagnetism or vibration can make them lose information, so error correction is needed. The industry generally assumes about 1 million qubits for commercially meaningful fault tolerance; Oratomic believes it needs only 10,000.

No business model yetCEO Dolev Bluvstein told the Wall Street Journal the company has no idea what the business models or exact applications will be, only that it must first build a working computer that is cheap and manufacturable. The same day, UK-based Universal Quantum announced a Series A of more than $100 million for its trapped-ion approach, co-led by DCVC and Firgun Ventures.

Capital flowing to pre-product hardwareSiliconANGLE cited McKinsey data showing the quantum computing industry generated just over $1 billion in revenue worldwide in 2025. When a one-year-old company with no settled business model is valued in the billions, the most direct effect falls on quantum companies pursuing other approaches: investors will compare them on metrics such as qubits required, and the bar for raising money rises for everyone.

▮ SIGNALQuantum fundraising is starting to move at the pace of AI model companies, but without revenue to verify it, the valuation rests on an engineering hypothesis that has yet to be proven.

❯ SpaceX buys nationwide 800 MHz low-band spectrum as Starlink Mobile aims to become a carrier, and the three big US carriers fall after hours

Filling the low-band gapSpaceX announced on October 8 an agreement to acquire a nationwide low-band spectrum license portfolio. According to CNBC, the seller is Grain Management and the portfolio covers up to 14 megahertz of paired spectrum in the 800 MHz band, subject to approval by the Federal Communications Commission. SpaceX said it paves the way for Starlink Mobile to become a major mobile carrier in the US.

Low-band signals get through wallsWireless spectrum comes in higher and lower bands. According to Finimize, citing Reuters, Starlink Mobile already holds global 2 GHz mid-band spectrum; low-band travels farther and penetrates walls better, and most unmodified phones already support 800 MHz. The FCC this week also approved SpaceX’s Gen2 constellation application covering 15,000 very-low-Earth-orbit satellites.

Going around the carriersAT&T, Verizon and T-Mobile all fell in extended trading. CNBC noted the three formed a joint venture last week to expand coverage in underserved areas through satellite direct-to-device service, without Starlink. Starlink had 12 million subscribers as of June 30 and $4.29 billion in connectivity revenue in the second quarter. T-Mobile had already removed Starlink from its own satellite promotions, and SpaceX now intends to bypass carriers and serve users directly.

Cities still need towersTMF Associates analyst Tim Farrar cautioned that this is still a very limited amount of spectrum and that reliable building penetration in cities would require SpaceX to deploy towers on the ground. CNBC also said the added spectrum will combine satellite and ground coverage only once the company builds a tower network. Investment bank William Blair expects SpaceX to bundle Starlink home broadband with mobile service, and that is what carriers have to guard against now.

▮ SIGNALA satellite company buying terrestrial spectrum shows that for direct-to-phone service to go mainstream, it still has to come back to the traditional telecom resources of spectrum and base stations.

❯ Musk says “certain oligarchs” are blocking Starlink in India, and the Indian government calls claims of unfair regulation baseless

"You can guess who they are"Elon Musk posted on X on October 7 that Starlink is “being blocked by certain oligarchs” in India seeking to maintain a “monopolistic chokehold” on the Indian people, calling it “a crime against the people of India.” According to CNBC, he named no one, saying only “You can guess who they are.” India’s telecom and internet market is dominated by Jio and Airtel, which together hold more than 80% of it.

Government says licenses are even-handedThe Indian government responded on October 8, calling suggestions that its regulatory framework is unfair or discriminatory “baseless and misconceived.” It said licenses have been granted to three global satellite communication providers and that every licensee must demonstrate compliance with security conditions. Starlink received its government license last year, but the service has yet to launch.

Last year's partnerships went quietSpaceX announced deals last year with Jio and Airtel to roll out Starlink in India, with little progress since. In June, Jio Platforms announced plans for its own low-orbit satellite service, leasing capacity from global constellation operators. Lauren Dreyer, SpaceX’s vice president of Starlink business operations, said at India Mobile Congress on October 7 that 11,000 satellites are in orbit, serving more than 170 countries.

The fight is over timingCounterpoint Research partner Neil Shah said the Indian government is concerned about a foreign player operating a sensitive technology like satellite communications, and about how terminals will be geofenced and whether data will be localized. He expects Starlink to get approval around the time Indian companies develop their own satellite internet services. For Starlink, the cost of waiting is its first-mover advantage.

▮ SIGNALSatellite internet ignores borders technically, yet on the ground it must clear licensing, security and local-partner hurdles everywhere, so coverage ends up set by each country’s pace of admission.

❯ Meta bans ByteDance from advertising in the US and six other countries, also blocking third-party ads that link to TikTok

No more ads from a rivalAccording to Bloomberg, Meta began on October 8 a complete restriction on ads and paid marketing messages placed on its platforms by ByteDance. The ban takes effect in seven countries: the US, Canada, Egypt, Indonesia, Japan, Thailand and Vietnam. It also covers third-party advertisers running campaigns that link to TikTok and other ByteDance properties.

Meta calls it normal business practiceMeta spokesperson Chris Sgro said in a statement: “We don’t have to run ads from a competitor whose goal is to pull people off our apps.” He said declining promotional services to a competitor is a normal practice across industries and that Meta will keep competing on product quality and user experience.

Child safety lit the fuseBloomberg said the rivalry heated up in August, when Meta agreed to an $18 billion settlement with US states requiring new safety features for minors on Instagram and Facebook, and has since pressed TikTok and YouTube to make similar commitments. Last month TikTok removed dedicated Instagram links from user profiles and rejected ads from Meta demanding comparable concessions.

One acquisition channel fewerApps commonly win new users by advertising on other platforms. After the ban, ByteDance can no longer advertise through Facebook and Instagram in these seven markets, and its budget has to move elsewhere. Third-party advertisers driving traffic to TikTok shops and accounts are caught as well; they are not Meta’s competitors, yet they lose the same channel.

▮ SIGNALSocial platforms used to advertise on one another and trade traffic; now they are shutting their doors to each other, and the cost of moving users between apps will rise with it.

❯ Trump awards the National Medal of Science to Musk, Brin, Huang and Su, and the National Medal of Technology and Innovation to Nadella and Dell

All six honorees run tech companiesUS President Donald Trump presented two medals on October 8 at the Science: A New Golden Age Summit in Washington. The list published by the White House shows Elon Musk, Google co-founder Sergey Brin, Nvidia’s Jensen Huang and AMD’s Lisa Su received the National Medal of Science, while Microsoft’s Satya Nadella and Dell founder Michael Dell received the National Medal of Technology and Innovation.

The citations are about computeThe White House cited Huang for advancing the graphics processing unit into a general-purpose computing architecture and transforming massively parallel computation, and Su for innovations in processor architecture and chiplet-based integration that powered the world’s first exascale supercomputer. The National Medal of Science has been bestowed by the president since 1963; the National Medal of Technology and Innovation was first awarded in 1985.

Another announcement the same dayAccording to NBC News, hours before the ceremony Vice President JD Vance announced the administration would suspend Microsoft and several other companies from a program that lets foreign workers apply for green cards, even as Microsoft’s CEO was a guest of honor. A White House spokesperson thanked the recipients and said they help keep America leading in innovation.

Executives on the science podiumThe two medals are the highest US honors in science and technology, and this time all six recipients are founders or heads of technology companies, four of them directly in chips, servers or cloud computing. The White House release repeatedly uses the term “super intelligence” and recognizes computing infrastructure as a scientific achievement.

▮ SIGNALWhen the National Medal of Science goes to people who make chips and build data centers, the official definition of a scientific contribution has come to include the supply of compute.