Are AIs collectively shifting away from chat-only models? Needle 3 focuses solely on tool calling, and runs on as little as 9MB.

Beating AI News Flash: AI data center operator Nscale has filed an initial public offering (IPO) application with the U.S. Securities and Exchange Commission (SEC), planning to list in New York and raise up to $3 billion, with the proposed stock ticker NSCL. The company was spun off from its cryptocurrency mining business in early 2024 and has since transitioned into an AI computing infrastructure service provider. According to its prospectus, as of August 2026, Nscale owns and controls data centers with a total power capacity exceeding 10GW, holding total contracts worth approximately $103 billion, including a $45 billion lease commitment from Anthropic this August. Nscale’s clients and partners include NVIDIA, Microsoft, and other firms. The company has posted rapid revenue growth alongside expanding losses: in the six months ended June 30, it generated $140.6 million in revenue, a year-on-year increase of over 12 times, while its net loss reached $1.02 billion, compared with a net loss of $368.9 million in the same period last year. In March this year, Nscale closed its Series C financing round, led by Aker ASA and 8090 Industries, with participation from NVIDIA, Nokia, and others, valuing the company at roughly $14.6 billion. The prospectus shows NVIDIA holds more than 5% of Nscale’s shares, which are non-voting. Additionally, Nscale plans to issue $1 billion in unsecured convertible loan notes or non-voting shares to NVIDIA. Nscale is currently accelerating its AI data center expansion: its Monarch campus in West Virginia is expected to reach a 2GW power capacity when it launches operations in 2028, with plans to scale to around 8GW by 2031. The company also announced in July this year that it would acquire AI software firm Anyscale for $1.65 billion to further extend its AI computing power software services. Goldman Sachs, JPMorgan Chase, and Morgan Stanley are acting as lead underwriters for the IPO.

Beating AI Express: OpenAI researcher Noam Brown stated on the Dwarkesh Podcast that even completely disconnecting an AI-run computer from the internet and physically isolating it is not an absolute security measure. This is because computers can secretly transmit messages via CPU heat—a method called BitWhisper, developed in 2015 by Israel’s Ben-Gurion University and unrelated to AI at the time. Researchers had two adjacent computers run matching programs in advance, then controlled one to adjust CPU load to alter temperature, while the other used a temperature sensor to read the temperature shifts and convert them into 0s and 1s. The two machines could exchange data without any network connection. However, this communication is extremely slow: the experiment only covered a distance of 0 to 40 centimeters, with a speed of 1 to 8 bits per hour, and both computers needed to run the matching programs beforehand. Brown noted that the biggest lesson from the Hugging Face incident is that humans underestimate AI’s ability to find vulnerabilities and bypass restrictions. BitWhisper is just one example; more advanced AI in the future could very likely uncover additional communication paths humans haven’t considered.

Oracle protocol Switchboard has announced it will cease operations. According to an official statement, core development contributor Switchboard Technology Labs Inc. decided to wind down its remaining business after exhausting all alternatives, with all existing support services to be fully halted on September 25. Switchboard noted that since its founding in 2021, its goal has been to provide a permissionless and customizable data layer. Its permissionless Feed Creation Tool, launched in 2022, allowed users to build custom data sources independently, but the oracle industry has seen weakened pricing power and demand amid factors including AI lowering the barrier to oracle development, a bear market slowing the launch of new public chains, and sharp contractions in project budgets. Switchboard added that following recent security vulnerabilities and the exit of other contributors and node operators, all related implementations are deprecated effective immediately. Protocols relying on Switchboard must migrate immediately to other oracle services such as Pyth and RedStone to avoid service disruptions. BlockBeats previously reported that on May 28, 2024, oracle protocol Switchboard announced the completion of a $7.5 million Series A funding round led by Tribe Capital and RockawayX, with participation from Solana Foundation, Aptos, StarkWare, Mysten Labs, InfStones Global, OtterSec, Bixin Ventures, and various angel investors.

According to Arkham's disclosure, Morgan Stanley's Bitcoin ETF (MSBT) has cumulatively purchased Bitcoin worth approximately $51.5 million over the past 20 trading days, with no single-day capital outflows recorded during the period. MSBT is also one of the few Bitcoin funds that has not seen any single-day net outflows this month.

Yesterday, Grayscale Zcash ETF saw a net inflow of $270 million, marking its 16th consecutive trading day of net inflows. Meanwhile, the U.S. spot Ethereum ETF recorded a net inflow of $143.7 million, ending its three-day streak of net outflows.

Beating AI News Flash: Vercel Labs engineer Chris Tate has integrated Jev into json-render, launching experiments to enable AI to directly compose UI at runtime. Json-render pre-prepares buttons, cards, input fields, data, and actions. Unlike large language models (LLMs) that generate full JSON token-by-token, Jev only determines which components to use, their order, and placement before handing off to a standard UI renderer. The new interface defaults to selecting core components in one pass, with a second pass for layout when required. In a demo generating an identical train ticket interface, the default JSONL mode took 3.21 seconds from request to completion, while Jev clocked in at just 0.88 seconds. Future app interfaces may no longer need to be fully hardcoded. UI can become runtime-decisive, allowing the same app to dynamically assemble different pages for different users, tasks, and states. Jev’s latest demo has pushed full generation time to under 1 second, at least bringing this type of interaction closer to real-time. Runway’s Solaris, released at the end of last month, is pursuing a similar direction but with a more radical approach. Solaris uses a world model to generate entire interfaces frame-by-frame, directly bypassing traditional code and components; Jev, by contrast, still relies on pre-built components and actions written by developers, only dynamically deciding how to combine them. One aims to turn software into a real-time generated video world, while the other is more like giving software a face that can be rearranged at any time.