The 30-year Treasury yield hit 5.44%, its highest since 2004, as inflation fears revived. In IPOs, British fintech Revolut is exploring a New York-London dual listing, Nscale filed in the U.S., and Oura aims to raise up to $2.2B, while SB Energy delayed its offering after valuation pushback. Data-center expansion also hit resistance: local opposition stalled $68B of U.S. projects, and Gavin Newsom signed seven California oversight bills. In tech, Microsoft cut 268 Xbox jobs and moved Halo and two other games to Activision. Meta shares rose 12% as Muse downloads surged; Amazon blocked its shopping access, while Meta unveiled a December-bound Muse Charm. Apple is testing a screenless Whoop rival. In cybersecurity, ShinyHunters claims it stole 2TB-3TB of FBI and DOJ personnel and applicant data, which the FBI is investigating. Finally, Warren Buffett stepped down as Berkshire chairman but remains a director.
Top 5 AI Highlights
[1] ⚖️ OpenAI and Anthropic reportedly neared a legally binding agreement to stress-test each other’s models. Separately, security researchers reportedly used Claude to access OpenAI’s private code repository during a bug-bounty investigation. More from The Information and Wall Street Journal.
[2] ☁️ Anthropic signed an $11.6B, seven-year cloud services commitment with Akamai to support growing CPU workloads. Akamai issued a warrant tied to spending milestones that could give Anthropic up to 5% of its stock. Separately, Anthropic released Claude Opus 5.5, which it says cuts task costs by 40% versus Opus 5. More from Akamai and Anthropic.
[3] ⚙️ OpenAI released GPT-6 Sol and GPT-6 Luna. Both accept text and image inputs; Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs one-twentieth as much for high-volume work. OpenAI also added plugin access to ChatGPT Voice, letting users speak to connected apps and continue unfinished Work tasks in text. More from OpenAI’s API changelog and ChatGPT release notes.
[4] 🏛️ OpenAI, Anthropic, and Google reportedly plan to launch an industry-run AI safety standards body without government oversight in late 2026 or early 2027. Separately, the White House asked OpenAI and Anthropic to withhold new models from the UK’s AI Security Institute until the U.S. completes its own review. More from The Information and Politico.
[5] 🛡️ Google confirmed Gemini accessed three real companies during a cybersecurity test after unintended internet access; the model stopped after recognizing the targets were real. Separately, Australia said an OpenAI research agent gained unauthorized access to public and non-public files on a Medicare statistics portal in June; no personal information is believed to have been accessed. More from Axios and ABC News.
✨Open Source AI Development
[1] 🔓 Harvey returned its gross margin to positive territory after shifting work to a model post-trained on Moonshot’s open-weight Kimi K3. Abridge is building on Nvidia open weights, Decagon routes 80% of queries through its own models, and Ramp and Rogo are exploring similar moves. Together, they show open weights moving into the economics of high-value vertical AI products. More from The Next Web.
[2] 🧠 Alibaba said it plans to build a 5T-10T-parameter AI model and unveiled its Zhenwu V900 accelerator, which the company says is three times faster than its predecessor and can scale to clusters of 500,000 chips. It also plans its first cloud regions in Turkey, Finland, and the Netherlands over the next 12 months, expanding its cloud and AI infrastructure footprint beyond China. More from Reuters, Bloomberg, and Alibaba Cloud.
[3] 🌐 DeepSeek’s annualized revenue run rate reportedly reached $1B, more than double its level a few months ago, as the company aims to finalize a ~$7.5B fundraise by late October. DeepSeek and Moonshot also briefed the UN Security Council on AI risks, while China’s regulator is reportedly investigating both developers over alleged data leaks to Anthropic. More from The Information on DeepSeek, Reuters, and The Information on the probe.
[4] 🐉 Abacus.AI released three open-weight Smaug models built on Kimi K3, DeepSeek V4 Flash, and Qwen3.8 27B for different agent workloads. Their relevance is deployment control: enterprises can self-host specialized models for long agent loops, though licenses differ by base model and the performance claims are vendor-run. More from Abacus.AI.
[5] ⚡ Xiaomi released the open-weight MiMo-V2.6 Pro and Flash models. VentureBeat reports that Pro tied Grok 4.7 xHigh and led open-weight models on Artificial Analysis’ Intelligence Index at launch, making the release more notable than a routine model refresh. More from VentureBeat and Xiaomi.
Founder’s Corner
OpenAI vs. Anthropic Wallet Share before launches of Opus 5.5 and GPT-6
Source: X
What We Read This Week
- Firebird is building a 300MW AI data center in Armenia and expects to install about 70,000 Nvidia chips by 2027, illustrating how smaller markets are trying to secure a place in the AI infrastructure buildout. (Bloomberg)
- TypeSafe AI’s Jev became the fastest-adopted model in Vercel’s AI Gateway history; Vercel said one workflow ran five to 18 times faster after replacing an OpenAI model. The startup is now reportedly in talks to raise $1B+ at a $10B+ valuation, underscoring investor demand for narrow models that return structured decisions instead of prose. (TechCrunch; Vercel; The Information)
- The UK’s AI Security Institute is contending with staff stress and burnout as employees face pressure to assess rapidly advancing frontier models on tight timelines. (Financial Times)
- Some universities are restricting AI detectors or dropping writing assignments after false-positive concerns made it harder to distinguish student work from machine-generated text. (The Atlantic)
- Chinese economists warn that the country’s roughly $295B data-center buildout may do little to lift weak consumer demand or employment, underscoring the gap between AI investment and near-term economic growth. (New York Times)
- Google plans to launch a four-TPU Project Suncatcher satellite on a SpaceX rocket on October 1 and operate it for a year, testing whether orbital computing can become a practical extension of AI infrastructure. (New York Times)
- Nscale filings show ByteDance accounted for nearly 75% of its 2025 sales and used the company’s Norway facility to access Nvidia AI chips. (Financial Times)
- Meta plans to support Muse purchases through Shopify’s Shop Pay and patched a Mac-app flaw that exposed authentication tokens. (Wall Street Journal; Ars Technica)
- Six major US banks warned that agentic shopping could increase scams, fraud, privacy risks, and payment disputes as AI agents transact for customers. (Gizmodo)
- SK Hynix launched SK Hynix Ventures in Silicon Valley, broadening its startup investments to AI computing, data centers, and optical connections between chips. (The Korea Herald)
AI Fundraising News (Sep 18 — Sep 24)
- Adlib Software: Agreed to acquire Paperbox, an agentic AI insurance platform, for undisclosed terms.
- Ande: An AI-native corporate-entertainment platform, raised $52M across seed and Series A rounds.
- Angle Health: An AI-powered health insurance platform, raised $600M at a $2.7B valuation.
- Basecamp Research: An AI life-sciences company, raised a $140M Series C.
- Baselayer: An AI platform for business verification and fraud risk, raised a $35M Series A.
- Baseten: An AI inference company, is in talks to raise Series G funding at a $26B valuation; the amount is undisclosed.
- Beagle Labs: An AI-assisted commercial-insurance risk assessment company, raised a $4.1M pre-seed round.
- BigHat Biosciences: An AI-native protein therapeutics company, raised a $75M Series C.
- Biolevate: An AI scientific and regulatory workflow company, raised a €30M Series A.
- Brahma AI: An AI audiovisual-content platform, raised $150M.
- Chamelio: An AI platform for in-house legal teams, raised a $26M Series A.
- Cognex: Agreed to acquire RealSense, which makes vision technology for robotic perception and physical AI, in a ~$600M deal, including $500M in cash plus retention earnouts.
- Confido: An AI-powered back-office platform for consumer brands, raised a $55M Series B.
- Corridor: An AI home-health documentation platform, raised a $25M Series B.
- Cyera: An AI-agent and data-security company, raised a $400M Series G extension.
- DeepSeek: An open-weight AI model developer, aims to finalize a ~$7.5B fundraise by late October; sources say its annualized revenue run rate reached $1B.
- Ekai: An AI business-context platform, raised a $1.7M pre-seed round.
- Ema: An enterprise AI-agent company, raised a $77M Series B.
- Enveda: An AI drug-discovery company, raised a $311M Series E at a $2B valuation.
- F13: An AI model developer for precise, editable vector graphics, raised a $5M pre-seed round.
- Fearn: An AI-native patent law platform, raised a $5.5M seed round.
- Firecrawl: A web-data library for AI agents, raised a $75M Series B.
- Go.AI: An AI hardware-appliance provider for banks, raised an $85M Series A.
- Heidi: An AI medical-scribe and care platform, raised $340M, including a $100M Series C and $240M in non-dilutive financing.
- Iambic Therapeutics: An AI drug-discovery company, filed for a Nasdaq IPO.
- Micro1: An AI training-data company, raised $100M+ at a $4B valuation.
- Mika: An AI bookkeeping and tax platform for small businesses, raised a €6M seed round.
- Mirendil: A self-improving AI model developer, is in talks to raise up to $1B at a $5B post-money valuation.
- Modal Labs: An AI infrastructure company, is in funding talks at a ~$15B valuation; the amount is undisclosed.
- Morphotonics: A nanoimprint-lithography company serving AI glasses and data-center optics, raised a €40M Series B.
- Noxtua: An AI legal-document technology company, sold a majority stake to C.H.BECK; terms were undisclosed.
- Nscale: An AI infrastructure company, is seeking to raise about £2B in an IPO at a £15B to £20B valuation.
- OpenEvidence: An AI search engine for doctors, raised $250M at a $15B valuation.
- Palma.ai: An AI-agent permissions and governance platform, raised a $1.8M pre-seed round.
- Penelope Health: An AI-powered healthcare benefits platform, received a $500M commitment from Thoreau Capital Management.
- Primo: An AI platform for corporate IT management, raised $8M.
- Sela: A voice-agent platform for mortgage sales, raised $21M across seed and Series A funding.
- Snorkel AI: An agentic data-development platform, raised $350M at a $3.5B valuation.
- Sol: An AI chief-of-staff company, raised $4M.
- Spott: An AI recruitment-workflow platform, raised a $21M Series A.
- TypeSafe AI: An AI model developer, is in talks to raise $1B+ at a $10B+ valuation after announcing a $40M seed last week.
- Upwind Security: Acquired Aegis, a cybersecurity startup focused on stopping AI-generated attacks, in an all-equity deal valued at $25M-$50M.
- Verda: An AI neocloud, raised $189M at a $1B+ valuation.
- Wander: Acquired Polarity, an applied AI research lab focused on agents and model post-training; terms were not disclosed.
- Within: An AI workflow-context company, received an undisclosed strategic investment from Accenture Ventures.
- Zeliq: An AI sales-agent company, raised a €7M seed extension.