Nvidia lines up as an anchor investor in Anthropic's IPO
Reuters reported on September 11 that Nvidia is in talks to anchor Anthropic's initial public offering with a commitment of up to $10 billion. Bloomberg followed with the same figure and added that Anthropic is seeking to raise as much as $100 billion at a valuation near $2 trillion. If it prices there, the listing would be the largest IPO on record, ahead of SpaceX's roughly $1.77 trillion debut in June. Anthropic declined to comment, and Nvidia did not respond to requests for comment.
The two companies are already tied together. In November 2025 Nvidia said it would invest up to $10 billion in Anthropic as part of a wider agreement, and Anthropic committed to buying $30 billion of Microsoft Azure capacity running Nvidia chips. Amazon and Google are both existing backers and compute suppliers: in April Anthropic said it would spend more than $100 billion on AWS over ten years for over a million Trainium2 chips, and it has deals with Google and Broadcom for gigawatt-scale TPU capacity. An anchor commitment lands before the offering is marketed to the public, which is why it carries weight for the rest of the book. It sets a reference price that the market then argues with.
The numbers behind the raise have moved quickly. Anthropic raised $65 billion in May at a $965 billion valuation, and its annualized revenue run rate passed $65 billion by the end of July, up from about $9 billion at the end of 2025. The company wants to complete the listing before the US midterm elections in November, with a roadshow expected to begin in mid-October. Two things are worth separating. The valuation is a target under negotiation, not a settled number, and Nvidia's participation is a statement about demand for its own chips as much as an endorsement of Anthropic's economics.
— Reuters · Bloomberg
🔗 HuggingNews: Anthropic targets $100 billion raise in largest IPO with $10 billion Nvidia investment · Reuters via 新浪财经: 英伟达考虑出资100亿美元参与Anthropic 1000亿美元IPO · 环球网: Anthropic 冲刺超级 IPO，英伟达拟最高投资 100 亿美元
OpenAI opens the Codex harness to everyone and ships a voice model that talks over you
Two releases on September 11 move OpenAI further into infrastructure. The Agents API entered public beta, exposing the managed harness that runs its Codex-style agents through a single API call. The four concepts are agent, environment, session and events. OpenAI handles session orchestration, context compaction across long tasks, sub-agent coordination, lazy tool loading and recovery after a crash, and billing adds no platform fee on top of model tokens, tool usage and any hosted sandbox compute.
The practical change is that the agent loop stops being something every team rebuilds. Sessions survive long tasks, context gets compacted instead of truncated, and sub-agents can be spawned and reconciled without the caller writing the scheduler. That removes a category of engineering work that has kept smaller teams out of production agents, and it also moves the durable parts of those systems onto OpenAI's runtime, which is the trade to weigh before committing.
GPT-Live-1 is the second piece. It is a full-duplex speech model: it listens and speaks at the same time, treats interruptions as normal rather than as errors, and can hand deeper reasoning to a backend model such as GPT-6 Astra while the conversation keeps moving. The voice layer is priced at $0.05 per minute. The older way to build a voice assistant chained speech recognition, a language model and synthesis, so every boundary added latency and a place for things to break. HeyGen open-sourced a real-time avatar framework built on GPT-Live-1 the same day, pairing it with its LiveAvatar and HyperFrames work.
— OpenAI (official) · AI Native Foundation
🔗 OpenAI: Introducing GPT-Live-1 in the API · AI Native Foundation: Global AI Native Industry Insights 20260911 · GitHub: HeyGen liveavatar-gpt-live demos
Cursor's coordinator agent runs thousands of subagents, and the code no longer needs GitHub
Cursor shipped roughly fifteen updates between September 4 and 11, and the center of them is Projects. Each project runs a coordinator agent that does not write code. It splits the work into tasks, writes the plan, and delegates to subagents that can run in the thousands in parallel. The project keeps the codebase, research notes and user preferences for months, so a task can resume long after it started. Jobs run in the cloud by default and keep going when the laptop closes; when something needs local testing, a local agent is pulled up.
The rest of the release is about where the work runs. Origin now hosts repositories natively with two-way GitHub sync, pull-request tools and links into Vercel and Buildkite for previews and pipelines, so a team can start a project without GitHub at all. Agents can subscribe to pull-request and Slack events and start work without anyone writing a new prompt, and they will fix CI on their own. For companies that will not let code leave the building, execution can run on self-hosted machines inside the network, with dynamic team pools and hibernation, and secrets stay internal. Pre-warmed build snapshots cut cloud-agent boot times by up to 10x.
Read against OpenAI ending its model supply to Cursor in November, the shape makes sense. Rather than stay an AI editor on top of someone else's model, Cursor is pulling hosting, CI, preview and publish into its own product. The cost question the release does not answer is who pays for thousands of parallel subagents and the coordinator overhead sitting on top of them. Beta usage data will settle that.
— Cursor (official)
🔗 Cursor: Projects · My AI Guide: Builders weekly, Cursor Origin hosting and agent swarms · AI 中文社区: AI 动态日报 2026-09-12
A team of Devin agents factored a 260-digit RSA number
Cognition said its engineers used multiple Devin agents to factor RSA-260, a 260-digit semiprime, breaking the RSA-250 record that had stood since February 2020. The work is not the model doing arithmetic. The agents were pointed at building a high-performance GPU lattice sieve, the classical algorithm behind integer-factorization records, and the project ran for about three weeks. Cognition published the process and the bill, including which steps needed a human and which the agents handled.
The result matters less as a cryptography event than as a data point about orchestration. Factoring a number of this size with a lattice sieve is a long engineering project: writing CUDA, tuning memory access, running sieves, managing a large linear algebra step and checking results. That is exactly the kind of multi-week task where an agent has to hold a goal across many sessions, recover from failed runs and know when to ask for help. RSA-260 is far from any key size in use, so there is no security consequence. The useful part is the public record of where human intervention was still required.
— Cognition (official)
🔗 Cognition · AI 中文社区: AI 动态日报 2026-09-12
A robot foundation model that learns a new factory task from one video
Skild AI released S1, a robot foundation model that learns previously unseen long-horizon tasks from a single video demonstration. The company built it on NVIDIA's Blackwell, Isaac and Omniverse stack, trained on a mix of synthetic and real-world data. The target is the part of manufacturing and warehousing that changes often: when a line is retooled or a product changes, the usual answer is a reprogramming cycle, and S1 is aimed at skipping it.
The funding around the same problem is moving just as fast. Mecka AI, a two-year-old startup that provides robot training data infrastructure, is closing a Sequoia-led round at a valuation near $500 million, months after its Series A. Maven Robotics came out of stealth with a $100 million Series A and robots already running in customer facilities, selling hardware, software and deployment as one package. The three sit at different layers of one stack, and the money going into the data layer is the tell. One-shot learning is a claim about a model, but the raw material it needs is hours of labeled robot experience that almost nobody has at scale.
— NVIDIA (official) · TechCrunch
🔗 NVIDIA Blog · Skild AI · TechCrunch
Positron raises $875 million to bet that inference does not need HBM
Positron AI closed about $875 million across a Series C and C-1 at a post-money valuation near $5 billion, up from roughly $1 billion at its February Series B. The round splits into about $375 million of Series C co-led by NEA, Andra, Atreides, Valor and SemiAnalysis Capital, and up to $500 million of C-1 led by NEA and Jim Clark with Qatar's QIA taking part. Reuters reported the raise on September 10.
The money goes to the Asimov inference ASIC, built on TSMC's N3P process with tapeout targeted for late 2026 and production in the second half of 2027, plus the Titan system that packages it. The architectural bet is the interesting part. Asimov uses LPDDR5X memory instead of HBM, which sidesteps the HBM and CoWoS packaging bottlenecks that have constrained supply and pushed up prices for every other accelerator. Whether that works comes down to memory bandwidth per dollar on the workloads buyers actually run, and only shipping silicon can settle it.
— Positron AI (announcement) · Reuters · TechTimes
🔗 Reuters: AI chip startup Positron's valuation skyrockets in latest funding round · TechTimes: Positron AI Raises $875M to Prove Commodity Memory Can Beat HBM in Inference · Positron AI
DeepSeek's V4.1-Flash activates 8B of 552B parameters and decodes above 400 tokens a second
DeepSeek released V4.1-Flash, a sparse mixture-of-experts model with native multimodal vision, a 1M-token context window and decoding above 400 tokens per second. Only about 8 billion of its 552 billion parameters are active for any given token, and the KV cache is 890 bytes per token, which is the number that decides how many long sessions fit inside a fixed pool of memory. DeepSeek claims it beats its own flagship V4-Pro on coding and agentic tasks at lower cost.
The release lands inside a pricing pattern rather than a capability race. DeepSeek's 2026 has been a run of cuts, and a Flash model that is cheaper to serve than the flagship is the same move applied to architecture: fewer active parameters, a smaller KV cache and a 1M-token window mean long agent sessions cost less to hold open. The company has separately retained CITIC Securities to prepare a STAR Market listing.
— DeepSeek (official) · Hugging Face
🔗 DeepSeek on Hugging Face · AI Briefing: DeepSeek releases V4.1-Flash with 1M-token context · My AI Guide: daily roundup 20260910
AI Daily Digest is published every morning by KD Agentic. Sources are linked inline; aggregator coverage is used for discovery only and is not cited as a primary reference.
Top comments (0)