Today's stories share one thread: AI's center of gravity is moving from models to capital, power and speed. OpenAI decides it can afford to wait on Wall Street, Apple finally outsources Siri's brain to Google, Nvidia's next rack doubles the current king before it even ships, a swarm of AI agents runs the first documented machine-speed global hacking campaign, and an Nvidia-backed startup launches Australia's biggest IPO in decades to build 16 gigawatts of compute.
OpenAI says going public in 2026 would be “ill-advised”
- Sam Altman confirmed to Fortune that OpenAI will not go public this year, calling it an “ill-advised moment” and pushing the IPO window to 2027 — with safety concerns, not market conditions, as the stated reason.
- The contrast is striking: rival Anthropic is racing toward a mega-IPO at a valuation near $2.3 trillion with Nvidia negotiating a $10B stake, while the world's most famous AI company chooses to wait.
Why it mattersAn OpenAI listing would be one of the largest in history; delaying it “for safety” tells you even insiders see turbulence ahead — and leaves Anthropic to test the public markets first.
Apple's Gemini-powered Siri finally ships (but not everywhere)
- The new Siri began rolling out September 14 with iOS 27, built with Google's Gemini models running on-device and on Apple's Private Cloud Compute, with WhatsApp and Audible app integrations at launch and an iPhone 15 Pro or newer required.
- Coverage has holes: no iPhone availability in the EU, nothing at all in China, and server-side features carry still-unspecified daily usage limits.
Why it mattersAfter years of delays Apple outsourced Siri's brain to its fiercest rival in search — and Latin America is included at launch while the EU and China wait, a rare case of the region getting frontier AI first.
Nvidia's Rubin doubles Blackwell before it even ships
- SemiAnalysis published the first deep benchmarks of the Vera Rubin NVL72 rack: about 59.4 million tokens per second per megawatt on DeepSeek's 1.6-trillion-parameter model — roughly 2.1x the best Blackwell GB300 configuration.
- Under realistic cost assumptions Rubin delivers up to 10x more tokens per dollar than Blackwell, and these numbers come from pre-release software that will only improve.
Why it mattersPower, not chips, is now the bottleneck of AI: doubling tokens-per-megawatt resets datacenter economics and helps justify the hundreds of billions in compute commitments the labs keep signing.
Hundreds of AI agents ran a global hacking campaign
- Security firm GreyNoise revealed that an attacker used hundreds of AI agents — built on OpenAI's Codex and DeepSeek models — to exploit two PaperCut print-server flaws, compromising 440 servers across 395 organizations in 48 countries.
- Once launched, the swarm breached 11 organizations in 26 seconds and went from first access to full domain administrator at a US high school in seven minutes; education made up roughly half the victims.
Why it mattersThis is the first well-documented AI-orchestrated mass intrusion: attack timelines just collapsed from weeks to seconds, and patching cadences everywhere — very much including Latin America's under-resourced IT teams — are not built for machine-speed adversaries.
An Nvidia-backed startup launches Australia's biggest float in decades
- AI data-center company Firmus kicked off its roadshow to raise about US$5 billion (A$7 billion) on the Australian stock exchange in October, funding a 16-gigawatt compute buildout.
- Firmus already raised roughly $2 billion in August, counts Nvidia among its backers, and reports link OpenAI as an anchor capacity customer.
Why it mattersWhile OpenAI delays its own listing, the AI infrastructure gold rush keeps going public: power-hungry compute is becoming its own asset class, and stock exchanges far from Silicon Valley want a piece.
The big picture
One day, one lesson: the AI race is no longer only about who has the best model. It is about who controls capital (OpenAI waiting, Firmus and Anthropic listing), distribution (Apple renting Google's brain for a billion phones) and energy (Rubin's tokens-per-megawatt). And while productivity doubles, so does risk: the same agentic AI that runs your workflows just proved it can run a global cyberattack in 26 seconds. The tools and the threats now scale together.
Sources & further reading
- Fortune — Sam Altman on OpenAI's IPO window
- TechCrunch — Altman: going public in 2026 'ill-advised'
- The Rundown — Apple ships Siri AI with iOS 27
- SemiAnalysis — Vera Rubin NVL72 agentic inference benchmarks
- BleepingComputer — AI-powered PaperCut attack hit 395 orgs
- GreyNoise — AI-orchestrated campaign against PaperCut NG/MF
- Bloomberg — Firmus seeks $5B in Australia IPO