❯ DeepSeek’s annualized revenue run rate reaches $1 billion as it seeks roughly $7.5 billion

Citing people familiar with the matter, The Information says DeepSeek’s annualized revenue run rate has reached $1 billion, up from less than $500 million a few months ago. A run rate projects the current pace over a year; it is not revenue already earned over a full year. DeepSeek has not published financial statements showing which products or customers account for the increase.

DeepSeek develops large language models. It offers a chat product for consumers and APIs that let developers add coding, question-answering and document processing to their own software. Its official pricing page charges for input and output tokens—roughly, the amount of text processed. Usage therefore affects API revenue, while the company’s full revenue mix remains undisclosed.

Alongside that growth, DeepSeek hopes to finalize roughly $7.5 billion in funding by late October. The final valuation, investors and specific use of proceeds have not been disclosed. Model development and serving user requests both require compute; operating income and external funding together influence how much the company can invest.

For developers using DeepSeek, steady revenue and new capital could help sustain its models and services. For investors, growth provides evidence of demand, but heavier usage can also raise serving costs. Assessing commercialization therefore requires both recurring customer spending and a view of how much extra compute is needed to generate each additional dollar of revenue.

▪ SIGNALA model business must show both that customers will keep paying and that each extra dollar of revenue does not require an equivalent increase in compute spending.

❯ Anthropic seeks majority voting control for seven founders ahead of a potential IPO

The Information reports that Anthropic is asking shareholders to approve a voting arrangement under which its seven co-founders could retain 50.1% of voting power as long as at least three maintain minimum stakes. Shareholders have not approved the proposal, and the full terms have not been disclosed.

Anthropic develops Claude, offering an AI assistant, the Claude Code programming tool and model access for developers building their own products. It needs ongoing investment in research and serving capacity. Funding and governance are consequently linked: raising more money can dilute founders’ ordinary shareholdings.

The proposal would separate ownership from decision-making power. The founders could retain a voting majority even without owning most of the shares, provided the conditions are met. The report connects the arrangement to a potential IPO and compares it with Palantir’s structure, but Anthropic has not announced a listing timetable. Minimum stakes, covered decisions and expiry conditions depend on the final approved documents.

For the founders, the arrangement could reduce pressure from outside shareholders to change direction for short-term returns. For future investors, contributing capital would not necessarily bring proportionate influence. Board oversight, shareholder protections and conditions for ending the special rights therefore matter. The structure will help determine who controls long-term research spending as well as ordinary shareholder votes.

▪ SIGNALAnthropic wants room to maintain its research direction after raising more capital; investors will want the checks on that control defined just as clearly.

❯ TypeSafe seeks over $1 billion soon after its seed round, targeting a valuation above $10 billion

TypeSafe announced a $40 million seed round last week and is already discussing a new raise of more than $1 billion. The Information says the new round targets a valuation above $10 billion, compared with the roughly $200 million seed valuation reported by PitchBook. Talks are ongoing, so the target is not a completed deal price.

TypeSafe supplies AI decision models to software developers, beginning with Jev. Its product introduction describes a service that accepts a question, possible answers and the current situation, then returns a judgment and probabilities that software can use directly. A support application might decide which team should handle a request; another might check whether content meets a requirement. Customers pay to call the model repeatedly, and TypeSafe has published input-token pricing.

Jev is available in early access. It is designed for fast, structured decisions rather than long-form text generation. TypeSafe is addressing software tasks that need a little judgment without invoking a full conversational assistant each time. Moving from an early product to a proposed valuation above $10 billion puts the focus on converting developer experiments into sustained enterprise use.

If such models can make many small decisions cheaply, developers could incorporate AI into more ranking, support-routing and content-checking tasks. The test is whether those decisions are reliable enough for production systems and produce repeat usage. A completed funding round would provide more development resources, although the company has not disclosed specific uses for the proposed proceeds.

▪ SIGNALJev’s opportunity is to make AI judgment as routinely callable as other software functions; genuine business demand behind those calls must ultimately support the valuation.

❯ Modal Labs seeks funding at a roughly $15 billion valuation, triple its level four months ago

Bloomberg reports that Modal Labs is discussing funding at a valuation of about $15 billion, roughly triple its valuation four months ago. Modal declined to comment, and the planned raise and final terms have not been disclosed. The figure remains a negotiating target.

Founded in 2021, Modal provides AI cloud infrastructure for developers. Its website offers model serving, training, batch processing and code execution environments. Developers submit work, while the platform allocates GPUs or conventional processors and adjusts capacity with demand. Customers buy compute and the service of managing it, reducing the work of preparing servers themselves.

The funding discussions come as more companies put AI functions into actual use. After a model is trained, every answer, generated image or agent task still consumes compute. Bloomberg notes that more open models are also encouraging companies to deploy models themselves, creating demand for platforms that operate them. Meeting that demand requires capital for chips and computing capacity.

For customers, on-demand resources can reduce the burden of building and managing infrastructure. For Modal, scaling speed, reliability and pricing help determine retention. The higher valuation reflects investor expectations, but the company must avoid preparing expensive capacity that customer usage fails to fill. Matching resource purchases to actual customer usage will directly affect the return on expansion.

▪ SIGNALInference demand brings both revenue and compute bills; Modal’s advantage must come from pooling customer workloads while keeping service fast and costs under control.

❯ Baseten seeks funding at a roughly $26 billion valuation, double its June level

Bloomberg reports that Baseten is discussing funding at a target valuation of about $26 billion, twice its $13 billion valuation in June. The company declined to comment, and the round size and terms have not been disclosed. The funding round remains under negotiation.

Founded in 2019, Baseten supplies software and compute for other companies to run AI models. Its customers may have a model or application in mind without the infrastructure team of a major AI lab. A startup putting an open model into customer support, for example, needs it to answer continuously, handle traffic peaks and recover from failures. Baseten provides deployment and operating capabilities for such needs.

More open models mean more choice, but obtaining a model and turning it into a reliable service remain different engineering tasks. Baseten can benefit from this demand, and financing can help it secure more compute. It also competes with cloud providers and other serving platforms, so industry growth does not guarantee customer growth. Specific uses of the proposed proceeds have not been disclosed.

For businesses adopting AI, a specialist platform can shorten deployment time and reduce the need to build an internal team. Customers still have to compare price, response speed and reliability. Their growth can increase Baseten’s usage revenue, but it also raises compute spending. Supporting the new valuation over time requires both repeat customer spending and efficient request processing.

▪ SIGNALAs models become easier to obtain, companies need help running them reliably; Baseten’s value depends on turning that engineering work into a sustainable service.

❯ Xiaomi unveils MiMo-V3’s core architecture, claiming about one-fifth as much compute to read one million tokens

Xiaomi MiMo lead Luo Fuli introduced HySparse 2, the core architecture planned for MiMo-V3, in an X post. The team says that, compared with MiMo-V2.6’s architecture, it reduces the compute needed to read a one-million-token input to about one-fifth (1/5.02) of the previous amount and reduces the KV cache used while generating an answer by 4.5 times, while improving long-context retrieval benchmarks. These are team-reported tests; MiMo-V3 itself has not launched.

The architecture targets long conversations and agent tasks. Agents accumulate files, dialogue and tool results, then may need to read that history repeatedly to continue working. The computation needed to ingest the material is called prefill; intermediate information retained during answer generation sits in the KV cache. These affect waiting time and memory usage, helping explain why million-token tasks become expensive.

HySparse 2 attempts to select and reuse information more precisely. The team describes moving from block-level to token-level selection and changing how recent context is retained so local and global information can share a cache. It also uses KV bridging and reuse. In practical terms, it aims to reduce repeated computation without losing information needed for the task. The reported benchmark gains do not establish equivalent savings on every user’s bill.

If the improvements hold in real workloads, long-document analysis and agents that repeatedly call tools could use less memory and finish sooner. Saving resources must not come at the expense of missing crucial details, however. Task accuracy, elapsed time and total cost need to be assessed together to establish whether users benefit.

▪ SIGNALLong-context competition now includes affordability and reliability: storing more information matters only if a model can find the useful parts at a reasonable cost.

❯ Google plans to send four TPUs into orbit on October 1 to test space-based AI compute

According to The New York Times, Google plans to launch a Project Suncatcher test satellite on a SpaceX rocket on October 1. It will carry four TPUs for roughly a year of in-orbit testing. The schedule may change, and this is not an operating space data center.

TPUs are Google’s specialized AI processors. On Earth, they rely on data-center power, cooling and networking; all of those conditions need verification in orbit. Radiation, temperature changes, continuous power, heat removal and data transmission affect whether useful computation can happen. Four TPUs make a small experiment, but can provide orbital evidence that ground simulations cannot fully replace.

The current milestone is preparation for a test launch, far from large-scale deployment. After launch, the question is whether the equipment works reliably for the intended period. Even a successful single-satellite test would leave questions about coordinating more devices, connecting networks and dealing with failures.

For AI infrastructure providers, Suncatcher opens an experimental computing environment, but cannot yet substitute for terrestrial data centers. Any commercial service would have to include launch, communication and maintenance in its cost comparison. Its immediate contribution is technical evidence, rather than compute capacity customers can buy.

▪ SIGNALSpace compute is first a reliability and cost problem; getting a satellite into orbit is the start, while dependable long-term computation is the next source of value.

❯ Google DeepMind says Gemini 4 is in post-training and could launch before year-end

In a post relayed by The Information, Google DeepMind executive Koray Kavukcuoglu said Gemini 4 is in post-training and could launch before year-end. Google has not given a firm date, specifications or release scope. “Before year-end” is a possibility, not a commitment.

Post-training follows the initial large-scale training stage. After a model has learned broad knowledge and language patterns, the team continues adjusting how it answers, reasons, uses tools and handles unsafe requests. Evaluation and product integration also remain. Entering this stage shows progress, but does not establish readiness for a broad release.

Developers can use the timing as a reference for planning tests, but it is too early to decide whether to migrate existing systems. That choice requires comparing the released model’s task performance, speed, price and API availability. Rival providers may face another entrant before year-end, while actual rankings and market effects depend on the product that arrives.

▪ SIGNALDevelopment milestones can shift expectations, but measurable performance gains and migration benefits are what persuade customers to change models.

❯ Anthropic uses an internal AI model to make common Claude actions about three times faster in two weeks

In an engineering retrospective, Anthropic says it spent two weeks in August improving common actions in Claude.ai and the desktop app. In company tests, time from opening the website to being able to type fell from 3.1 to 0.55 seconds; starting a Claude Code session fell from 0.8 to 0.3 seconds; loading a Cowork cloud session fell from 2.6 to 0.73 seconds. These are interface and loading times, not the model’s answer-generation speed.

The work began with frequent points of waiting rather than new features. The team broke actions into measurable tasks and used its internal model to help diagnose bottlenecks, change code and create tests, with engineers deciding which changes to keep. For long answers, frame-by-frame tests also helped identify rendering stalls. AI participated in finding, fixing and checking problems.

Anthropic says four priority paths account for about 95% of user activity. The overall “three times faster” result concerns selected actions; networks, devices and conversation length still affect individual experiences. Other teams can learn from turning a vague complaint about slowness into a measurable engineering problem. For users, keeping the product smooth over time matters more than a one-off improvement report.

▪ SIGNALAI coding tools should be judged not only by code produced, but also by whether they keep reducing the waits, freezes and failures that users actually encounter.

❯ Bezos has invested $30 billion in Blue Origin as its first outside funding values it at $140 billion

Citing documents, The Wall Street Journal says Jeff Bezos has invested about $30 billion in Blue Origin since 2000. In its first outside funding round, the company raised about $10 billion at a valuation near $140 billion; Bezos contributed $2 billion to that round. His cumulative investment and his latest contribution are different measures and should not simply be added together.

Blue Origin is a commercial space company developing rockets, engines and related infrastructure. Its potential business rests on customers needing launch services, space equipment and mission capabilities. Research, manufacturing and testing require large upfront spending; more dependable revenue becomes possible once the company can deliver regularly. Investment and repayment can be separated by years.

Bezos’s long-term backing has enabled the company to keep pursuing development. The first outside round adds investors and diversifies funding sources, while his continued participation puts founder and external capital in the same financing. Specific uses of the proceeds have not been disclosed.

More capital can help the company absorb long project timelines, while bringing clearer expectations of commercial returns. New investors need to judge whether technical achievements can become regular launches, completed orders and revenue. The $140 billion figure is a transaction valuation; sustained business delivery will determine whether that price is justified.

▪ SIGNALTo attract lasting capital, a space company ultimately has to turn technical achievements into work it can repeatedly deliver and collect payment for.

❯ Tencent launches TenPayGo so visitors to China can pay WeChat merchants with international cards

Tencent has launched TenPayGo, a standalone payment app for visitors to China. Citing a Tencent statement, Bloomberg says visitors can use Visa, Mastercard and Apple Pay at tens of millions of merchants on the WeChat Pay network. Merchants keep their existing payment setup; the new entry point is mainly for visitors.

QR payments are common in everyday Chinese commerce. Visitors carrying international cards still need a way to connect those payment methods with local merchants. TenPayGo bridges familiar overseas payment tools and the WeChat network. Reusing existing merchant acceptance can make expansion easier than requiring new terminals in every store, while eligible countries, cards and fees depend on product rules.

For visitors, the practical test is smooth registration, card linking, verification and payment, plus clear handling of failures or refunds. Easier payment could help merchants avoid losing purchases. For Tencent, the standalone app creates another channel to attract visitor transactions and compete with Ant Group’s services. Actual payment experience will determine whether users stay.

▪ SIGNALVisitor payments succeed when the first purchase goes through smoothly; a large merchant network still needs a simple, reliable path for the customer.

❯ Zuckerberg says Meta began developing camera-free Ray-Ban glasses before the privacy backlash, with no launch date set

In an interview with Joanna Stern, Mark Zuckerberg said Meta had begun working on camera-free Ray-Ban glasses before the privacy backlash against glasses with cameras. He described a development direction, not a product launch; price, capabilities and timing have not been announced.

Cameras are an important input for smart glasses. They can show AI what the wearer sees, enabling questions about objects, text and surroundings, as well as recording. People nearby may not know when they are in view, so the hardware affects both capabilities and acceptance. Zuckerberg’s point about work predating the backlash explains the project’s history; the interview did not provide full specifications.

If a camera-free version launches, it could suit places where recording is unwelcome and ease some privacy concerns. It would also have different visual capabilities, so it should not be treated as an all-around upgrade over existing glasses. For Meta and its competitors, providing convenience while making the device’s behavior understandable to others will affect how widely it can be used.

▪ SIGNALMass-market smart glasses must work for both wearers and bystanders; removing one sensor could sometimes open more settings for use.

❯ US considers promoting dollar stablecoins abroad through joint ventures with companies

Bloomberg reports that the Trump administration is considering ways to promote dollar stablecoins abroad, including possible joint ventures with private companies. This is still a policy discussion. No partners, target markets or implementation documents have been announced.

Dollar stablecoins generally seek to maintain a fixed relationship with the dollar, supported by reserves such as cash and Treasuries, while moving on blockchain networks. They offer another way to transfer dollar value across borders, although settlement, conversion costs and availability depend on the service. The report links the policy discussion with expanding overseas dollar use and Treasury demand.

If overseas usage grows, issuers may need more reserves and payment companies may gain transaction demand. Those effects require real adoption; a policy proposal does not produce them automatically. Foreign governments will consider local-currency use, capital flows and regulatory responsibilities, while consumers need confidence in redemption. The partnership structure and the willingness of target markets to accept it will shape the outcome.

▪ SIGNALStablecoins can extend digital dollar payments, but cross-border adoption still requires both local regulatory acceptance and trust in the reserves.

❯ Moonshot’s Kimi K3.1 is tipped for October, but the company has not confirmed a launch

Recent rumors say Moonshot may launch Kimi K3.1 in October with Low, High and Max reasoning levels, plus Agent and Swarm modes. Moonshot has not formally announced K3.1, and the release date and feature set remain unconfirmed.

Moonshot develops Kimi. Its official documentation says the existing Kimi K3 already supports a one-million-token context and native visual understanding for tasks including long-running coding and knowledge work. Context capacity allows more material in a request, while vision allows images as input. These capabilities are the foundation that a later release would need to improve upon.

If the rumored reasoning levels are implemented, they could offer choices between reasoning effort, waiting time and results; Swarm concerns multiple agents dividing work. Names alone do not establish task quality. Developers considering Kimi should base current decisions on released products, then assess migration once official K3.1 pricing, interfaces and test results are available.

▪ SIGNALMore model modes do not automatically create more value; an upgrade must help the same job get done better, faster or at lower cost.

❯ AI short-drama producers face a price war as competition shifts toward scripts and IP

Cheaper generation tools and increasing supply are putting pressure on AI short-drama production prices. Industry commentary describes sharp declines for some standardized projects over six months, alongside moves toward customized work and longer-term IP ownership. Projects vary widely in length, quality and revision requirements, so standardized and customized production face different pricing pressures.

Traditional production assigns people to filming, performance, voice work and editing; AI tools allow smaller teams to handle some of those tasks. Lower barriers let more companies deliver similar finished videos, making price comparisons easier. Fees once supported by labor and equipment requirements may shrink, with some of the technology savings passed to customers through lower quotes.

A finished video is only part of the business. Audience appeal, rights, revisions, customer acquisition and platform shares all affect earnings. Teams selling only production work face easier price comparisons; distinctive subjects, repeat customers or reusable IP can offer more room to negotiate. That helps explain moves from contract production into operating content businesses rather than simply increasing output.

Platforms and viewers gain more content choices, while finding good work becomes more important. Producers need to combine tool efficiency with scripts, taste, client needs and distribution; more output alone need not bring profit. The same tools can reduce production costs without ensuring that each title attracts viewers or recovers its investment.

▪ SIGNALAs making a video becomes easier, a studio’s pricing power increasingly depends on creating content audiences want to watch and customers want to buy again.