1. Google releases Gemini 4 Argon, gated to cyber-defense partners first
Google DeepMind announced Gemini 4 Argon on September 30, calling it its most capable model for long-horizon software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense. The headline change is an output limit of 1M tokens, up from 64K on the previous generation. Citing safety, Google is initially limiting access to selected companies and organisations focused on cyber defense before a wider rollout.
2. Trump and AI CEOs sign a voluntary 'self-police' accord at the White House
At a September 29 White House luncheon, President Trump and the heads of the leading labs signed a voluntary accord on frontier AI development, described by CBS as 'morally binding' controls. Signatories included Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang and Elon Musk. Trump said the industry needs 'tremendous self-regulation' and restated his opposition to government intervention, drawing criticism from both Bernie Sanders and Steve Bannon at a rival AI-risk summit the same day.
3. FTC opens industry-wide safety probe into OpenAI, Anthropic and other frontier labs
One day after the White House accord, a senior FTC official confirmed the agency is conducting a sweeping investigation into Anthropic, OpenAI and other frontier labs over the dangers their systems pose to consumers. The probe follows a string of reported 'rogue agent' incidents, including OpenAI's own newly published misalignment-reports site. It lands awkwardly against the administration's stated preference for industry self-regulation.
4. OpenAI shelves 2026 IPO, seeks about $30B privately at roughly $1.4T
OpenAI is reportedly putting its IPO on hold and instead targeting a private round of nearly $30 billion at a valuation of about $1.4 trillion. Sam Altman has cited unfinished AI-safety work as the reason for not listing this year, with a public offering now pointed at 2027. The round is framed as a bridge to the delayed listing; secondary reports put 2025 revenue near $40 billion, up roughly 70 percent.
5. Anthropic's IPO prospectus leaks: $4.6B revenue, $42B net loss, $518B in compute commitments
Reuters reviewed Anthropic's draft prospectus ahead of a listing reportedly aimed at November. It shows 2025 revenue of about $4.6 billion against a net loss of $42 billion, with roughly $7.3 billion spent on compute and $518 billion in future cloud and infrastructure obligations. The filing carries long risk disclosures on AI safety. No public S-1 has appeared on EDGAR yet.
6. AMD to acquire Fei-Fei Li's World Labs for $8.2 billion
AMD announced on September 28 that it will buy World Labs, the world-model startup founded by Fei-Fei Li, in an $8.2 billion deal expected to close by year end. Li joins AMD as executive vice president and chief scientist reporting to Lisa Su, and the team will keep working on model research. Andreessen Horowitz is the largest outside holder at nearly 14 percent. The move signals AMD buying model expertise to shape its AI compute roadmap against Nvidia.
7. AI agent startup Instinct raises $1B Series C at a $10B valuation
Instinct, the San Francisco company building a personal agent for everyday life, raised $1 billion from Sequoia, Benchmark and Coatue on September 28. The round quadruples its valuation to $10 billion only a month after its previous mark. The company also unveiled new concierge and peer-to-peer features as consumer demand for agentic assistants surges.
8. ElevenLabs doubles to a $22B valuation on a $300M employee tender
The London-based AI voice company completed a $300 million employee tender offer on September 30 that values it at $22 billion, twice the $11 billion set in its February Series D. The tender was co-led by Wellington and T. Rowe Price. The company credits surging demand for AI voice agents.
9. California CEO arrested for smuggling $300M of Nvidia AI servers to China
Federal agents arrested Yiu Kong 'Greg' Lui, 38, owner of Earthmade Computer Inc., on October 1 on charges of conspiring to smuggle more than $300 million of export-controlled Nvidia hardware into China. Prosecutors say the shipments were routed through Malaysia and Singapore using false paperwork. He faces up to 20 years in prison, and the case adds to a run of arrests as Nvidia plays down the role of smuggled chips in China's AI build-out.
10. GPT-6 Astra Ultrafast goes live on Nvidia Blackwell
Nvidia said on October 1 that GPT-6 Astra Ultrafast, OpenAI's highest-speed service tier, is now available in the OpenAI API and to eligible ChatGPT Work and Codex users, running on Blackwell GPUs. OpenAI's docs describe Ultrafast as its fastest tier, priced above standard, with WebSockets recommended for agentic workloads that chain many tool calls. Nvidia separately launched a 64GB DGX Spark for local AI development.
// KEY TAKEAWAYS
The week was defined by a collision between money and oversight. On the capital side, OpenAI shelved its IPO for a roughly $30B private round, Anthropic's leaked prospectus showed a $42B net loss alongside half a trillion dollars of compute commitments, and AMD, Instinct and ElevenLabs posted billion-dollar deals within 72 hours of each other. On the oversight side, the White House chose a voluntary self-policing accord while the FTC opened a formal safety probe into the same labs a day later. Google's decision to gate Gemini 4 Argon to cyber-defense partners first shows frontier labs now treating release restrictions as a default rather than an exception.