Top AI stories · Wednesday, Sep 9, 2026
About 10,000 concurrent agents running a next-generation internal model that OpenAI described as significantly more capable than GPT-6 Astra worked on the problem for about 88 hours. OpenAI said the system produced a proof that smooth three-dimensional fluid motion can develop a singularity in finite time under a smooth external force, and that GPT-6 Astra spent another 17 hours formalizing and verifying the work in Lean. OpenAI said the work resolves the Navier-Stokes problem, one of mathematics' six remaining Millennium Prize problems, though outside mathematicians still need to independently validate the result. The announcement also intensified a dispute with Tristan Buckmaster and Levent Alpöge over related work and authorship. OpenAI said no specific user data was accessed to solve the problem, but it cannot rule out that de-identified data from product usage helped improve its models.
Meta's new Muse personal assistant utilizes a confidential cloud virtual machine and a Sentinel agent to send emails, manage calendars, and process payments across third party apps. The agent is powered by the Muse Spark 1.3 model and provides a free tier of 100M tokens per week, alongside monthly paid subscriptions at $20 and $100. Access is currently available for US users on web, iOS, Android, and WhatsApp. Early adoption has reached 10 times the volume seen in testing cohorts. To build the secure VM architecture that hides user data from Meta, Mark Zuckerberg recruited Signal founder Moxie Marlinspike and Nat Friedman. Zuckerberg's personal use cases include tracking mountain climbing permits and reviewing MMA gym footage to receive performance feedback.
Jacob Coxon left his position at Anthropic and the broader artificial intelligence field after three years of pretraining work at the company and OpenAI. He warned that competing labs are making safety trade-offs to win a race toward self-improving superintelligence that could spiral out of control by the end of next year, rendering any individual lab's efforts to build safely ineffective. Evan Hubinger, Anthropic's alignment science lead, estimates a greater than 10% chance that AI kills all humans within the next decade. Hubinger said the company does not yet have a plan to solve the alignment problem for superintelligence, even as the firm prepares for a potential IPO and researchers inside the lab use terms like "crunchtime" to describe the pace of capabilities.
The tech giant will deploy $15.1 billion toward energy and data infrastructure between 2027 and 2028 to increase computing capacity. This initiative represents Google's largest single investment in Europe and includes 3 new data centers in Kajaani, Muhos and Vaala. The sites will utilize a 94 MW battery system and 629 MW of wind energy to maintain power stability. A 22 year power agreement with Fortum provides the company up to 50% of the output from the 2 reactor Loviisa nuclear plant, supplementing an expansion of the existing Hamina facility. These projects will support more than 37,000 construction jobs and 7,000 permanent positions in Finland, which CNBC describes as the Texas of Europe.
China's Commerce Ministry rejected allegations from US agencies that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI systematically mined American models. A joint advisory from the NSA, FBI, and CISA stated these six firms used industrial-scale knowledge distillation campaigns to extract proprietary functionalities through billions of tokens and millions of requests with ChatGPT, Gemini, Claude, and Grok since late 2024. Beijing described the claims as "unfounded" and "baseless," noting that distillation is a standard practice among AI developers globally, including those in the US. Chinese officials warned of "resolute countermeasures" if Washington implements actions to suppress Chinese AI firms under the guise of cracking down on distillation. The diplomatic clash occurs as the heads of state of both nations agreed to launch a separate intergovernmental dialogue on artificial intelligence to pursue greater cooperation.
Chief AI Officer Alexandr Wang reported a surge in early engagement for the new Muse assistant, noting that activity has far exceeded internal forecasts. Current users are interacting with the tool 10 times more than previous test cohorts, and the app reached the No. 4 position on the Apple App Store in less than 24 hours after its rollout to users in the United States. Meta shares rose more than 5% in premarket trading on Sept. 9 and broke above $650. The price action follows the debut of the personal AI agent, which uses a secure virtual machine to let users manage tasks and make purchases across the web and within other apps.
Company executives may restrict new sign-ups for the ChatGPT Pro tier to ensure stability for current customers. OpenAI's Thomas Sottiaux reported that demand for the GPT-6 Astra model has reached levels unseen during previous growth periods, requiring the firm to utilize all available operational levers to sustain the service. The potential halt on new subscriptions follows a surge in usage and capacity strain accompanying the launch of the latest model. Market observers note that the resulting bottleneck could serve as a strategic brand signal to enhance investor interest prior to a projected initial public offering.
Founder Liang Wenfeng plans to use a public listing to offer employee economic incentives and reduce talent attrition at the Chinese AI startup. DeepSeek engaged CITIC Securities to lead an initial public offering on Shanghai's STAR Market, with the firm planning to initiate the process this year. The listing push follows a funding round at a reported $75 billion valuation, an increase from the $7.4 billion the company secured a few months ago. DeepSeek is also recruiting approximately 150 senior engineers to overhaul backend systems that have been strained by high user demand and the rollout of AI agents.
The partnership between Qualcomm and Amazon focuses on the development of customized silicon for AWS AI infrastructure and large-scale AI inference. Amazon holds warrants for as many as 25 million Qualcomm shares tied to purchase milestones totaling $60 billion, with an exercise price of $161.26 per share expiring in 2036. The collaboration also includes advanced optical connectivity using Qualcomm's SerDes and DSP technology for links up to 1.6T. Qualcomm is providing ASIC services to ByteDance alongside the AWS agreement and will use Amazon Bedrock for its chip-design and EDA workloads. The warrant agreement saw 3.75 million shares vest immediately based on initial purchase commitments. Qualcomm shares rose nearly 10% following the official announcement of the multi-generational deal.
A four part debt offering with maturities from 3 to 19 years is scheduled for delivery Wednesday to expand the company's funding sources. The sale targets at least £3 billion in its first foray into the UK sterling market to finance artificial intelligence infrastructure. This issuance follows previous non dollar moves this year in euros, Swiss francs, and Canadian dollars. Amazon has sold more than $92 billion in bonds in 2026, making it the largest hyperscaler debt issuer of the year. The 19 year tranche specifically tests the market's willingness to absorb long duration AI debt after recent similar deals faced higher borrowing costs and weaker demand.
The ChatGPT maker is expanding its ties with Samsung Electronics to include the joint research and manufacture of specialized semiconductors. Harrison Kim, General Manager of OpenAI Korea, detailed the expanded cooperation across semiconductors and enterprise AI services during a press conference in Seoul on Sept. 9. OpenAI now provides AI specifically for chip design, which Chief Financial Officer Sarah Friar says offers a cost advantage over open-source alternatives to help developers build hardware more efficiently. Samsung is one of the largest global deployments of ChatGPT, and enterprise usage across South Korea has grown approximately 28x year over year. The hardware push is part of a strategy to penetrate specialized sectors such as financial services and life sciences, alongside efforts to drive down compute costs for training and serving frontier intelligence models to reduce dependence on third party providers.
The developer of the Claude AI assistant has lined up data-center and chip agreements that drastically increase its capital expenditure forecast through 2029. These commitments total $517 billion, compared to the $180 billion the company previously disclosed to its investor base, signaling an aggressive push for growth ahead of a potential public listing. Bankers for the AI firm are currently lobbying rating agencies for investment-grade credit ratings to lower financing costs for its buildout post-IPO. A potential listing before the November midterm elections could value Anthropic at $2 trillion, with the official prospectus release slated for late September and marketing beginning in mid-October.