Google unveiled Gemini 4 Carbon on Wednesday, fifteen days after widening access to Gemini 4 Argon. The new frontier model is restricted to a small group of cybersecurity defence firms through Google's Fairwind programme rather than a public API. Google says Carbon finds vulnerabilities, builds exploits and writes patches better than Argon, and produces secure code at more than twice Argon's success rate on internal tests. It holds first place on SEC-bench Pro and ExploitGym. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens.
The United States and China are building rival technology clubs, and middle powers are discovering that membership may soon be exclusive. Washington has told partners they cannot sit inside both the US-led Pax Silica grouping and China's World AI Cooperation Organization. Kazakhstan tried: its AI minister joined US supply chain talks in Washington in June, then signed in Shanghai three weeks later as a founding member of the Chinese initiative. China has funded AI projects in at least 77 low- and middle-income countries, according to AidData, while US export controls keep the most advanced chips on one side of the line.
SpaceX has closed $40bn of debt to buy Nvidia chips for artificial intelligence, Elon Musk said on Sunday. Musk described the financing as callable debt for GPU purchases arriving in the first half of 2027, and said it would not be a major factor this quarter. Bloomberg reported the deal is off-balance-sheet and was possible because SpaceX holds investment-grade ratings and earns revenue from Starlink and rocket launches. Musk used the announcement to draw a contrast with 'neocloud' firms that borrow to resell computing time. The neocloud sector has come under scrutiny over debt loads and doubts about demand, and the deal makes SpaceX one of the largest private borrowers in the AI build-out.
Anthropic has opened a research preview of Claude Duet, an AI agent that runs on its own cloud computer and contacts the user when it finds something worth acting on. The preview is limited to Claude Pro and Max subscribers in the US, who join a waitlist, and costs nothing extra for now. The agent sends a morning brief, works inside connected apps and asks for approval before sending messages, making payments, signing documents or deleting files. It shares usage limits with the rest of the user's plan. Meta, OpenAI and xAI have released similar products. Anthropic says it will use the preview to learn which notifications users keep.
The Nasdaq Composite fell 2.6 percent on Thursday, its steepest one-day drop since June, as investors questioned whether spending on artificial intelligence can keep paying off. The S&P 500 lost 1.21 percent and six of the seven biggest technology companies closed lower. Meta fell nearly 7 percent after lifting its capital spending forecast to between $135bn and $145bn while missing profit expectations. Nvidia dropped 4.4 percent, Intel and AMD each fell about 8 percent, and Broadcom lost 4.3 percent. Brent crude above $114 a barrel added to the pressure, and Asian markets opened lower on Friday.
Two critical flaws in PHP, the language behind an estimated 70% of websites, triggered a global patching scramble this week. Researchers using OpenAI models, working with security firms Wiz and Calif, showed the bugs could let attackers take over web servers without authentication and read memory that may hold passwords. The companies nicknamed the bug set PHP2Shell. One flaw had been in the language for 14 years. The US National Security Agency and the EU cybersecurity agency urged immediate patching and said unhardenable systems should go offline. Some European government sites did so. Cloudflare reported early attack attempts before the patch was public.
Intel has reached a construction milestone at its $28 billion Ohio chip campus, with the first steel truss lifted into place at the Ohio One site built by Bechtel. The project is at least five years behind its original timeline, and the company now points to operations in the 2030-2031 window. Intel says its next Arizona fab should start running by the end of 2028, while its most advanced 18A chips are already in high-volume production in Arizona. Analysts say the foundry business still needs a large outside customer, and Intel reports third-quarter results on October 29. The campus was announced in 2022 as the centrepiece of Intel's push to build leading-edge capacity in the United States.
At GTC Berlin, Nvidia showed the DSX POD, a reference design for building a complete Vera Rubin AI system inside a shipping container, and pushed co-packaged optics for denser networking between racks. On software it introduced the OpenShell reference architecture for running AI agents with guardrails. Google Cloud said A5X instances on Vera Rubin NVL72 systems are now generally available for European customers, and WPP demonstrated a creative agent built on Nvidia Nemotron models and OpenShell. The company also announced a compute donation worth $250 million over five years for UK research on agentic AI safety. Power efficiency was a central theme.
The Department of Homeland Security has ordered federal civilian agencies to remove all products and services linked to Chinese AI developer DeepSeek from their IT systems and to stop buying them. A notice filed on Monday and published in the Federal Register covers the Treasury, State and Justice departments but exempts the Defense Department. DHS said the models present multiple threat vectors, including the risk that sensitive data could be transferred to China. The order relies on a 2018 federal supply-chain law.
Samsung will build the base dies for Nvidia's custom high-bandwidth memory, NVHBM, on TSMC's 3-nanometer process rather than its own foundry, according to the Seoul Economic Daily. Industry estimates put TSMC's 3nm yield above 90 percent against roughly 60 percent for Samsung's 2nm line. The design moves the memory controller onto the base die, freeing GPU space and delivering up to 30 percent more bandwidth and 15 percent less power than HBM4E. It would be the first time the two foundry rivals have worked together on HBM, and it comes as UBS projects a 10 percent global HBM shortfall next year.