Funding & Investment Story 1 of 12
DeepSeek Is Set to Raise at Least $12 Billion With Tencent and CATL Behind the Round
DeepSeek, the Chinese lab known for its low cost models, is on the verge of a financing of remarkable size. Bloomberg reported on Tuesday that the company is set to raise at least $12 billion, about 80 billion yuan, in a round backed by Tencent and the battery maker CATL.
The size of the round is itself the story. According to Bloomberg, DeepSeek initially sought about 50 billion yuan, but investor demand was far bigger than expected. A lab that built its reputation on doing more with less is now being offered far more capital than it asked for, and the investors lining up are among the most strategically important companies in China.
The pairing of Tencent and CATL tells executives something about how Chinese capital is reading the AI race. These are not passive financial backers. Both have reasons to want a domestic frontier lab that can stand beside American model providers, and both have the balance sheets to keep funding one through the expensive years ahead.
DeepSeek released its V4 models earlier in 2026, and the new capital arrives as the gap between Chinese and American systems has become a regular subject of boardroom conversation. The company's pitch has always been efficiency: competitive capability at a fraction of the training and serving cost. Twelve billion dollars changes the strategic question from whether DeepSeek can keep up on a budget to what it will build when the budget is no longer the constraint.
For Western enterprises the implications are practical rather than abstract. DeepSeek's open releases have become a default option for teams that want to self host capable models, and a better funded DeepSeek is likely to keep that pressure on pricing across the industry. At the same time, procurement and security teams in regulated sectors will keep asking hard questions about provenance, data handling and export controls before any Chinese model touches sensitive workloads.
The broader signal is that the capital intensity of frontier AI is no longer a uniquely American phenomenon. When a Chinese lab can raise a round of this scale from strategic domestic investors, the competitive map for the next generation of models gets more crowded, and the price of staying at the frontier keeps rising for everyone.
DeepSeekTencentCATLChina AI
Policy & Regulation Story 2 of 12
The Pentagon Says It Has Stopped Using Anthropic's Claude, but BBC Sources Dispute the Timeline
The United States Department of Defense has told the BBC that it has ceased the use of Anthropic products, closing a chapter in one of the most consequential disputes yet between an AI company and the government that once relied on it. The statement marks the formal end of a phaseout that began in February, when Defense Secretary Pete Hegseth moved to designate Anthropic a supply chain risk to national security and the Pentagon was given six months to remove the company's products.
The official line, however, does not fully match what people close to the work are saying. BBC sources said Claude was still being used by the US military as recently as the week before the Pentagon's statement, including in operations involving Iran. If accurate, that means a model the department had formally designated a risk continued to support live operations well past the point it was meant to be gone.
The reason is not hard to find. Claude was integrated into Palantir's Maven Smart System, the platform the military uses to process large volumes of intelligence. Once a model is embedded inside a mission system at that depth, removing it is not a matter of canceling a subscription. Workflows, integrations, validation and operator habits all have to be rebuilt around a replacement, and that work competes with the operational tempo the system exists to support.
For enterprise leaders, this is the most useful part of the story, because it is not really about defense. It is a case study in what vendor exit looks like once a foundation model sits inside critical infrastructure. Contracts can be terminated on a date. Dependencies cannot. A six month deadline turned out to be an aspiration rather than an engineering plan, and the gap between the two was filled by continued use of the very tool being removed.
The lesson for any organization running models inside core processes is to design for substitution before it is needed. That means abstraction layers between applications and model providers, documented evaluation suites that let a replacement be qualified quickly, and a realistic inventory of where each model actually runs. Most companies do not have that inventory today.
The episode also underlines how quickly the political environment around AI suppliers can shift. A provider that was embedded in national security systems became, within a matter of months, a designated risk. Boards that treat model selection as a purely technical decision are underpricing that kind of exposure, and the Pentagon's experience suggests the cost of unwinding it arrives long after the decision to leave.
AnthropicPentagonDefenseClaude
Industry Dynamics Story 3 of 12
Meta and Microsoft Are Working to Cut Their Employees' Use of Claude
Two of Anthropic's most prominent corporate customers are pulling back. The Information reported on Monday that Meta and Microsoft are working to cut their employees' use of Claude, as both companies push staff toward tools they build themselves.
The numbers in the report are striking. According to The Information, the number of Meta employees using Claude Code fell from about 60,000 to about 30,000 this year. Microsoft, the report said, had been on pace to spend at least $1 billion a year internally on Anthropic technology, and that estimate has been cut by more than a third. Meta's own internal coding tool, MetaCode, now has more than 30,000 users, The Information reported.
The shift is easy to understand from the inside. Both companies build AI models of their own, and both have a strong incentive to keep internal engineering work on their own stacks. Every hour a Meta or Microsoft engineer spends inside a rival's coding assistant is money flowing to a competitor and usage data that does not improve the home product. Once internal tools are good enough, the business case for paying an outside vendor at that scale weakens quickly.
For Anthropic the story cuts two ways. Losing seats at two of the largest technology companies in the world is a visible blow, and it raises a fair question about how durable enterprise coding revenue is when the customer also happens to be a model builder. At the same time, the fact that Claude reached this depth of adoption inside Meta and Microsoft in the first place is a striking endorsement of the product, one that most of its customers, who do not build their own frontier models, have no equivalent alternative to.
For other enterprises the more relevant lesson is about leverage. AI coding tools have become one of the fastest growing line items in engineering budgets, and many organizations adopted them team by team without a central strategy. Meta and Microsoft are now doing what any large buyer eventually does: consolidating spend, steering usage toward preferred platforms, and treating developer tooling as a cost to be managed rather than an experiment to be encouraged.
Most companies cannot replace a frontier coding assistant with one they built themselves. But they can measure actual usage, negotiate on the basis of it, and keep their workflows portable enough that switching providers remains a real option rather than a theoretical one. That leverage is worth more each quarter as these tools become embedded in how software gets written.
AnthropicMetaMicrosoftAI Coding
AI Business Models Story 4 of 12
OpenAI Will Test Visual Ads Inside ChatGPT Image Generation
OpenAI is taking its advertising business into one of ChatGPT's most popular features. The company said on Monday that it will begin testing a visual ad format during image generation in ChatGPT in the United States later in October. The ads will be clearly labeled and kept separate from the image a user is creating.
The audience is enormous. OpenAI said ChatGPT now reaches 1.2 billion people weekly, which puts its potential advertising reach in the same conversation as the largest consumer platforms in the world. Image generation is a natural place to start. A user asking for a living room makeover, an outfit, or a product concept is often already in a shopping frame of mind, and a visual ad that offers inspiration alongside the request fits that moment more naturally than a text link.
OpenAI is also making the case to advertisers with early results. The company said WeightWatchers saw a 15.3% lower cost per acquisition from ChatGPT ads compared with search benchmarks. It is a single customer data point published by the seller, and marketers will want to see how it holds up at scale, but it signals exactly where OpenAI intends to compete: for the performance budgets that today flow to search.
To win those budgets, OpenAI is building the measurement plumbing that large advertisers require. The company named DoubleVerify and Integral Ad Science as brand safety partners, two names that every major media buyer already uses to vet placements. That choice matters. It tells agencies and brand teams that ChatGPT ads are meant to slot into existing buying and verification workflows rather than demand new ones.
For marketing leaders, the practical question is no longer whether conversational AI will carry advertising, but how quickly it becomes a line item worth planning for. Search spending has long been justified by intent: people tell the search box what they want. A chat assistant captures intent in far richer detail, and image generation adds a visual layer on top. The brands that learn how to show up usefully inside that moment, without feeling intrusive, will have an advantage when these tests turn into a full ad market.
There is a trust question underneath all of this. Users come to ChatGPT for help, not for promotions, and OpenAI's emphasis on labeling and separation reflects how carefully it has to manage that line. How the company balances revenue against the experience will shape whether its advertising business scales or stalls.
OpenAIChatGPTAdvertisingMonetization
Policy & Regulation Story 5 of 12
OpenAI Will Watermark ChatGPT Text for EU Users to Meet AI Act Rules
OpenAI said on Monday that it will add an invisible watermark, called textGrain, to eligible ChatGPT and Codex text output for users in the European Union over the coming weeks. The company tied the move directly to the EU AI Act, which requires providers of generative AI to make generated content identifiable in a machine readable way.
The rollout is not limited to Europe. OpenAI said API customers globally can opt in to text watermarking, which remains off by default for the API. That design leaves developers outside the EU free to choose, while giving companies that serve European users, or that simply want provenance signals for their own reasons, a way to turn it on.
What stands out in the announcement is how candid OpenAI is about the limits. The company said that at a 1% false positive rate, its detector identified the watermark in about 80% of 200 token passages and about 95% of 400 token passages. It also said replacing 10% of words with synonyms reduced detection from about 92% to 66%. In other words, the signal is reasonably strong on longer, untouched text and degrades quickly once someone edits it.
That candor is useful for executives, because it frames what watermarking can and cannot do. It is a compliance mechanism and a provenance signal. It is not a reliable way to prove whether a person wrote something, and it is not a tool that HR, academic integrity offices, or legal teams should treat as definitive evidence. Any policy that relies on text watermark detection as proof will run into exactly the editing problem OpenAI describes.
For companies operating in Europe, the more immediate issue is operational. Text produced in ChatGPT and Codex by EU users will soon carry a signal that did not exist before. Legal and compliance teams should understand how that interacts with their own disclosure obligations, and product teams that build on the API should decide deliberately whether to opt in rather than discovering later that a regulator expected them to.
The bigger picture is that the AI Act is now shaping product behavior in concrete ways. Requirements that once lived in policy papers are turning into features, defaults and settings. Watermarking will not be the last of them, and organizations that track these changes as they happen will spend far less effort catching up when enforcement arrives.
OpenAIEU AI ActWatermarkingCompliance
Funding & Investment Story 6 of 12
Moonshot AI Closes a Round at About $50 Billion and Eyes a Hong Kong IPO
Moonshot AI, one of China's leading AI labs, is preparing to test public markets. Bloomberg reported on Tuesday that Moonshot closed a final private funding round at a valuation of about $50 billion and is targeting a Hong Kong IPO in the first quarter of 2027 that could raise up to $5 billion.
The financial trajectory behind that valuation is steep. According to Bloomberg, Moonshot's annualized recurring revenue of about $1 billion is expected to rise to $2 billion by December. That kind of growth, if delivered, would give public market investors something rare in this sector: a frontier model company with real revenue scaling quickly enough to justify the price.
The deal is already taking institutional shape. Bloomberg reported that Bank of America is the overall coordinator for the planned listing, with China International Capital Corp, Deutsche Bank and Goldman Sachs as sponsor banks. That roster of global and Chinese banks signals an offering designed to reach international investors, not just domestic ones.
Moonshot's timing lands in the same week as DeepSeek's reported mega round, and the two stories together say a great deal about the Chinese AI market. Capital is concentrating in a small number of labs that have proven they can build competitive models, and those labs are now choosing between private strategic money and public listings. For global investors, a Hong Kong IPO offers a way to own a Chinese frontier lab directly, something that has been almost impossible until now.
For enterprise leaders, the significance is less about the stock and more about the durability of the competition. A publicly listed Moonshot would face disclosure requirements and quarterly scrutiny, but it would also gain a deep, permanent source of funding. That makes it more likely that Chinese models remain a serious, well funded alternative in global markets, continuing to pressure prices and pushing American providers to keep moving.
Timing for listings of this size can shift, and Bloomberg noted that deliberations are ongoing. But the direction is clear. The Chinese AI sector is maturing from a set of venture backed challengers into companies preparing to answer to public shareholders, and the first of those offerings will set the benchmark for every lab that follows.
Moonshot AIIPOHong KongChina AI
AI Models Story 7 of 12
Reflection AI Unveils Beam, a 501 Billion Parameter Open Weight Model
Reflection AI has unveiled the model it was built to deliver. The company announced Beam on Monday, a sparse mixture of experts open weight model with 501 billion total parameters and 23 billion active parameters. Reflection said Beam's weights, technical report and model card will be released under the Apache 2.0 license later in October.
The scale of the training effort puts Reflection firmly among serious model builders. The company said Beam was pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 NVL72 GPUs. For a company founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, and backed by Nvidia, that is a fast path from founding to a frontier class open release.
Reflection's headline claim is about efficiency. The company says Beam achieves reasoning scores comparable to Z.ai's GLM-5.2 while using 3 to 4 times less inference compute. That comparison is deliberate. Many of the strongest open weight models available to enterprises today are Chinese, and Reflection is positioning Beam as an American alternative that can match them on capability while costing less to run.
For enterprise and government buyers, that positioning addresses a real gap. Many organizations want the control that open weights provide, including the ability to run models in their own environments, fine tune them on proprietary data, and avoid dependence on a single API provider. But security reviews and procurement rules in regulated sectors increasingly make Chinese models difficult to approve. A capable, permissively licensed American open model gives those buyers an option that clears both bars.
The Apache 2.0 license matters as much as the benchmarks. It is one of the most permissive licenses in common use, which means companies can deploy, modify and build commercial products on Beam without the usage restrictions attached to some other open releases. Legal teams that have spent months reviewing custom model licenses will appreciate the simplicity.
The caution for buyers is that the weights are not yet in hand. Until the model and its technical report are public, the efficiency and benchmark claims remain the company's own. Teams that are interested should plan evaluations now, using their own workloads, so they are ready to test Beam the day it ships rather than relying on launch materials to make the decision for them.
Reflection AIOpen WeightBeamNvidia
AI Infrastructure Story 8 of 12
Etched Is Fielding Funding Offers at a $40 Billion to $50 Billion Valuation
Investor appetite for alternatives to Nvidia is producing valuations that would have seemed impossible a year ago. TechCrunch reported on Monday that Etched, the startup building chips designed specifically for AI inference, is fielding funding offers at a valuation of $40 billion to $50 billion.
The climb has been extraordinarily fast. Etched raised $700 million at a $21 billion valuation earlier this year, TechCrunch reported. A new round at the reported range would roughly double that figure in a matter of months, a pace that reflects how urgently the market is searching for ways to make serving AI models cheaper.
The bet behind Etched is specialization. General purpose GPUs are extraordinarily flexible, which is why they dominate both training and inference. But as AI workloads shift toward serving enormous volumes of requests in production, a chip designed to do one job extremely well can, in principle, deliver more output per dollar and per watt. Inference is where the recurring costs of AI live, and every enterprise running models at scale feels them.
There are early signs of commercial traction. According to TechCrunch, quantitative trading firm Jane Street took delivery of an early Etched system. A firm known for its rigorous approach to technology and its sensitivity to latency and cost is a meaningful early customer, even if a single deployment is a long way from proving the model across the market.
For technology leaders, the rise of companies like Etched is worth watching for a practical reason: inference cost is quickly becoming one of the largest variables in the business case for AI. If specialized silicon delivers on its promise, the price of running agents, copilots and customer facing models could fall sharply, which would change the economics of projects that look marginal today. Diversity in the supply of compute also reduces dependence on a single vendor, which procurement teams have been asking for since the GPU shortages began.
The risk is concentration of a different kind. Chips built around particular model architectures can lose their edge if the architecture of leading models changes. Investors offering these valuations are betting that today's dominant designs will last long enough for specialized hardware to pay off. That is a reasonable bet, but it is still a bet, and buyers should evaluate any specialized platform with that in mind.
EtchedAI ChipsInferenceValuation
Enterprise AI Story 9 of 12
Utah Lets Nolla Health's AI Issue Initial Prescriptions for Acne
A line that many in medicine assumed would hold for years has just been crossed in Utah. Nolla Health announced on Monday that its AI has been authorized in the state to issue initial prescriptions for acne treatments, which the company called the first such program in the United States. The service costs $4.99 a month for eligible Utah residents.
The program is built around staged oversight. Under the Utah pilot, a physician approves every AI prescription before it is issued for the first 100 patients. After that, physicians review prescriptions after they are issued rather than before. The structure is designed to let regulators and the company build evidence step by step, rather than granting broad autonomy on day one.
Nolla is making a case on both quality and access. The company said licensed clinicians agreed with its AI's treatment recommendations in more than 96% of real world cases. It also says the average wait to see a dermatologist in Utah is 61 days. Acne is a common, often chronic condition, and the argument is that a narrow, well bounded AI service can relieve pressure on a specialty with long queues while keeping physicians in the loop.
The choice of condition is shrewd. Acne treatment involves a relatively small set of well understood medications, clear clinical guidelines, and a low risk profile compared with most areas of medicine. That makes it a sensible test case for autonomous prescribing, and a template other states and companies are likely to study closely.
For healthcare executives, the significance goes beyond dermatology. Utah has shown that a state regulator is willing to grant an AI system real clinical authority when it comes paired with structured oversight and a narrow scope. That combination is likely to become the playbook: start with low risk, high volume conditions, prove safety with data, and expand gradually. Health systems that want to shape those rules rather than react to them should be building their own evidence now.
Expect pushback. Medical associations and patient safety advocates are likely to press concerns about AI making clinical decisions, and those concerns will intensify as programs like this expand. The outcome of Utah's pilot, including how it handles edge cases and adverse events, will carry weight far beyond a single state and a single condition.
HealthcareUtahRegulationNolla Health
AI Infrastructure Story 10 of 12
DayOne Data Centers Files for a Nasdaq IPO as AI Demand Lifts Revenue
The AI infrastructure boom is heading to Wall Street from an unexpected direction. Singapore based DayOne Data Centers has filed for a US IPO on Nasdaq under the ticker DODC, offering public investors a direct way to own a fast growing data center developer that builds outside the United States and China.
The growth figures in the filing are dramatic. DayOne reported revenue of $512 million in the first half of 2026, up from $151.5 million a year earlier. That more than threefold increase reflects how quickly demand for AI and cloud capacity is spilling beyond the traditional data center hubs of the United States and China.
The scale of what DayOne has lined up is just as notable. The company has 2.3 gigawatts of customer bookings and 962 megawatts of operational capacity, according to its IPO filing. The gap between those numbers is the investment thesis and the risk in one: there is far more committed demand than built capacity, which promises strong growth if the company can finance and deliver the buildout on schedule.
DayOne's footprint sits outside the traditional centers of the industry. Land, power and proximity to growing digital markets are drawing AI capacity to new regions, and developers that secured sites and power early are in a strong negotiating position.
For executives, the filing is a reminder that the physical constraints of AI are reshaping global geography. Compute is increasingly located wherever power and permits are available, and that is pulling investment toward markets that were secondary a few years ago. Companies planning AI deployments that depend on regional capacity, data residency, or latency to customers abroad should be paying attention to who is building there and how quickly.
For investors, DayOne will be an important test of public market appetite for pure play AI infrastructure outside the United States. Data center developers require enormous, sustained capital, and the public markets will scrutinize execution, financing and customer commitments closely. How this offering is received will influence whether other international developers follow it to New York.
Data CentersIPOSoutheast AsiaDayOne
Enterprise AI Story 11 of 12
Cohere Launches North 2 With Agent Memory and Budget Controls
Cohere is betting that the next phase of enterprise AI will be won on control rather than raw capability. The company launched North 2 on Monday, an upgrade to its agent platform that, Cohere said, adds agent memory that persists across sessions along with organization level budget caps, user quotas and rate limits.
Those two additions address the two complaints enterprise teams raise most often about AI agents. The first is that agents forget. Without memory, every session starts cold, and users have to re explain context that the system should already know. The second is cost. Agents that call models repeatedly can run up bills quickly and unpredictably, and finance teams have had few tools to cap that spending before it happens.
North 2 also leans into Cohere's longstanding strength with regulated industries. The platform can be deployed on premises and in fully air gapped environments. For banks, government agencies, defense contractors and healthcare organizations that cannot send sensitive data to a public cloud, that flexibility is often the deciding factor, more important than which model tops a benchmark.
The positioning is a clear response to the market Cohere competes in. The largest model providers have focused on consumer reach and frontier capability. Cohere has carved out a role as the enterprise specialist that meets customers where their data and their compliance obligations already live. North 2 extends that strategy from models into the agent layer, where governance questions are becoming more pressing.
For technology leaders, the release is a useful checklist of what a production agent platform should offer. Memory, spending controls, quotas and deployment flexibility are not exotic features. They are the basic governance that any system with autonomy over company resources should have. Teams evaluating agent platforms from any vendor should ask for each of them explicitly, and should be wary of pilots that skip these controls because they slow down a demo.
The broader shift is that AI agents are moving from experiments run by enthusiastic teams to infrastructure that finance, security and compliance functions must sign off on. Vendors that make that sign off easy will find enterprise adoption faster, and Cohere is clearly building for that buyer.
CohereAI AgentsGovernanceEnterprise
Generative AI Story 12 of 12
At 19, Ghost's Founder Raises $11 Million to Build a $3,499 Computer for Personal AI
A new hardware startup is betting that people will want their AI agents to live at home rather than in someone else's cloud. TechCrunch reported on Monday that Ghost, a startup whose chief executive Zain Javaid is 19, has raised an $11 million seed round led by Andreessen Horowitz.
The product is called Core, and it is a dedicated computer for running personal AI. Core costs $3,499 and uses an Nvidia RTX Pro 4000 SFF Blackwell GPU, giving it enough local compute to run capable open models without sending data to an outside provider. Ghost expects to ship the device in the last week of October.
The idea taps into a growing tension in consumer AI. The more useful an assistant becomes, the more it needs to know: your messages, your calendar, your health data, your documents. Many people are uneasy about handing all of that to a cloud service, and a box that keeps the data and the models on hardware the user owns offers a different trade. Privacy becomes a feature of the architecture rather than a promise in a policy.
The economics are also shifting in Ghost's favor. Open models have improved quickly, and hardware capable of running them locally has become more compact and affordable. A few years ago, running a strong model at home required a workstation few consumers would buy. Today it fits in a small form factor device priced like a high end laptop.
For enterprise leaders, Ghost is a small company, but it points at a trend worth watching. Local and edge AI is gaining momentum for the same reasons Ghost cites: privacy, control, and freedom from per token pricing. The same logic applies inside companies that are wary of sending sensitive data to external APIs. Expect more vendors to offer on premises appliances that bundle hardware, models and management into a single product.
The challenge for Ghost is the one every consumer hardware startup faces: convincing people to buy and maintain a new device when their phone and laptop already do so much. Backing from a firm like Andreessen Horowitz buys time to find out whether the personal AI computer becomes a category or remains a niche for enthusiasts.
GhostHardwareAI AgentsAndreessen Horowitz