OpenAI Shelves Its 2026 IPO as Apple and Google Ship Rival AI Platforms

AI & Machine Learning

Sam Altman told Fortune that OpenAI will not go public in 2026, calling the current moment “ill-advised” given the state of AI safety even though the company has filed confidentially for a listing. Altman said OpenAI is “not rushing into an IPO” and will wait until “the business is ready” and society has caught up with the technology, pointing to unresolved safety and alignment questions rather than financial pressure as the holdup. The company had previously been expected to target a Q3 or Q4 2026 offering; that timeline has now slipped toward 2027 or later. The remarks come as OpenAI works through fallout from a security incident at newly acquired Hugging Face and broader industry debate over how fast frontier labs should move. Source: TechCrunch Verified: True

Perplexity brought its local-first “Portable Computer” agent to Windows, letting the assistant plan and execute multistep tasks entirely on-device rather than in the cloud. The feature requires an Nvidia GeForce RTX or RTX PRO GPU with at least 24GB of VRAM, downloads through the Microsoft Store, and is limited to Pro and Max subscribers. It ships with a locally hosted Qwen 3.8 27B model that Perplexity post-trained specifically for the release, alongside its in-house PPLX 27B model and an upcoming Nvidia Nemotron 3.5 model for voice dictation. Because the work runs on the user’s own hardware, tasks consume no cloud credits and files never leave the machine, a pitch aimed at privacy-conscious power users and developers wary of sending sensitive data to third-party servers. Source: Tom’s Hardware Verified: True

Consumer Hardware

Apple shipped iOS 27 worldwide on September 14, delivering the biggest Siri overhaul in the assistant’s 15-year history, built on foundation models Apple developed jointly with Google using Gemini technology. The new Siri can hold extended back-and-forth conversations, draw on a user’s personal data across messages, email, and calendar to complete complex tasks, and search the web directly, though the AI features remain in an opt-in beta that may require a waitlist. Full Siri AI capability is reserved for iPhone 15 Pro and later, while the base iOS 27 update installs on iPhone 11 and newer, creating a split experience across Apple’s installed base. Apple also confirmed that some cloud-dependent Siri and Apple Intelligence features will carry daily usage limits, with expanded access planned as a future paid tier. Source: Apple Verified: True

Google used a New York media event to formally unveil Googlebook, a new laptop platform that fuses Android and ChromeOS under a Gemini-centered operating system previously known internally as Aluminium OS. Acer, Asus, Dell, HP, and Lenovo are the first hardware partners, with roughly eight devices — mostly clamshells plus one tablet — arriving this fall starting at $999, and pre-orders opening September 21. The platform runs Android’s app ecosystem natively while keeping ChromeOS’s browser-first design, and leans on Gemini for features like AI-generated widgets, a “Magic Pointer” input mode, and deeper handoff with Android phones. It’s Google’s most direct attempt yet to field a single hardware category to rival Apple’s tightly integrated Mac and iPhone lineup. Source: Android Central Verified: True

Cybersecurity

Microsoft disclosed an active cloud-intrusion campaign, running since May 2026, in which threat actors impersonate corporate IT help desks to trick employees into resetting their passkey, MFA, or SSO settings through fake sign-in pages. Microsoft ties the initial access activity to Storm-3121, which feeds the ShinyHunters and Falcon extortion brands, and Storm-3032, a splinter group now operating under the Helix banner. Once inside, the attackers enroll their own authentication methods to keep persistent access, then use the Microsoft Graph API to enumerate users and permissions before exfiltrating data from SharePoint, OneDrive, and Exchange Online over hours or days in patterns designed to blend in with normal activity. Microsoft is urging organizations to lock down help-desk identity-verification procedures and restrict device-code and cross-tenant authentication flows that the campaign abuses. Source: Microsoft Verified: True

CISA added five actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalog, spanning JFrog Artifactory, ConnectWise ScreenConnect, and MikroTik RouterOS. The most severe, CVE-2026-84869 (CVSS 9.9), is a missing-authorization flaw in ScreenConnect that lets an attacker transfer and execute files through an active remote session without host confirmation. The RouterOS issues, CVE-2026-67277 and CVE-2026-86060, allow kernel memory disclosure and privilege escalation, while the two Artifactory bugs can leak internal tokens or bypass authorization checks entirely. Federal civilian agencies face staggered deadlines running from September 13 through September 25 to patch the affected systems, underscoring how widely used developer and remote-support tools have become priority targets. Source: The Hacker News Verified: True

Enterprise Infrastructure

Cornelis Networks raised $205 million led by IAG Capital Partners and unveiled Active Compute Fabric, an open networking architecture that embeds programmable compute directly into the data-center fabric rather than treating the network as a passive transport layer. The Intel spinoff says the design lets accelerators process data as it moves between them, cutting the GPU idle time that comes from waiting on data transfers, and it works across different vendors’ GPUs rather than locking customers into one accelerator family. Cornelis is already shipping its 400 Gbps CN5000 switch and plans an 800 Gbps CN6000 generation, funded in part by the new capital. The raise positions Cornelis as one of a growing cluster of infrastructure startups betting that AI data centers will diversify away from Nvidia’s proprietary networking stack. Source: TechCrunch Verified: True

Independent benchmarking firm SemiAnalysis published the first verified agentic-inference results for Nvidia’s next-generation Vera Rubin NVL72 rack, measuring roughly 67 times the throughput per total cost of ownership of the current GB300 NVL72 generation on its AgentX benchmark, which replays real-world agentic coding sessions rather than single-shot prompts. At a lower interactivity target the advantage narrows to about 18 to 39 times depending on the workload, while separate CoreWeave testing on the simpler DeepSeek-R1 benchmark showed a 10x gain in tokens per second per megawatt. The figures matter because agentic workloads — where a model plans, calls tools, and spawns sub-agents — are far more sensitive to memory bandwidth and interconnect latency than traditional chatbot inference, making Rubin’s architecture changes a bigger deal for enterprises running AI agents at scale than raw compute specs alone would suggest. Source: SemiAnalysis Verified: True

Policy & Regulation

The UK government and India’s Cellular Operators Association (COAI) signed a memorandum of understanding at the CO.AI 2026 conference to cooperate on AI-enabled approaches to telecoms fraud, scams, and other digital harms. Signed between the British High Commission in New Delhi and COAI, the framework commits both countries’ telecom industries to share research and coordinate on applying AI to network security, fraud detection, and consumer protection as networks increasingly rely on AI for their own operation rather than just supporting AI workloads elsewhere. COAI Director General Lt. Gen. Dr. S.P. Kochhar framed the deal as recognition that telecom networks are becoming AI-driven systems in their own right, capable of autonomously spotting and blocking fraud patterns in real time. The agreement is non-binding but signals a template other governments may follow as AI-powered scam campaigns scale globally. Source: Unite.AI Verified: True