Today in AI — 27 March 2026

The demo era is over; the IPO era is here. Anthropic and OpenAI are both racing toward public markets while the platforms they sell into — Apple, GitHub, enterprise IT — are redrawing who controls the AI layer. Meanwhile, the human costs of that transition are getting harder to wave away.

The race to go public

Anthropic is in early talks with Goldman Sachs, JPMorgan, and Morgan Stanley about an October listing, which would put it on a collision course with OpenAI's own IPO timeline. OpenAI, for its part, just finished pretraining its next-gen model codenamed "Spud" and is killing products (first Sora, now an adult chatbot) to focus on the bets that justify a public valuation. Two companies, same quarter, both arguing they deserve to be the default infrastructure of the AI economy.

• Anthropic weighs IPO as soon as October in race with OpenAI to go public — Bloomberg
• OpenAI finishes pretraining next-gen 'Spud' model as Altman promises it will 'accelerate the economy' — The Decoder
• OpenAI shelves erotic chatbot indefinitely after employee and investor pushback — Engadget

Distribution fights

Apple turning Siri into a choosable AI frontend is the biggest distribution story of the week. If iOS 27 ships with Claude, Gemini, and ChatGPT as selectable backends, the model layer becomes a commodity and the integration layer becomes the product. Separately, GitHub defaulting all Copilot users into training data collection tells you where the real value is: not in the code you write, but in the signal your usage generates.

• Apple will open Siri to ChatGPT, Gemini, and Claude rivals in iOS 27 — MacRumors
• GitHub will train AI models on Copilot interaction data by default starting April 24 — The Register

AI gets a voice

Three separate moves toward production-grade audio AI in a single day. Mistral's open-weight Voxtral TTS clones voices from three seconds of audio and runs locally. Google's Lyria 3 Pro generates structured music tracks. ElevenLabs plugged its voice stack into IBM's watsonx for enterprise agent deployments. Voice is moving from novelty to infrastructure fast.

• Mistral releases Voxtral TTS, an open-source voice model it says beats ElevenLabs — TechCrunch
• Google launches Lyria 3 Pro, its most advanced AI music generation model — Google Blog
• ElevenLabs and IBM bring premium voice AI to enterprise agents via watsonx Orchestrate — IBM Newsroom

The human cost compounds

Meta cut another 700 roles while redirecting $135 billion toward AI infrastructure. Across the sector, 59,000 jobs have gone in Q1 alone, with one in five explicitly attributed to AI. I think we'll look back on this quarter as the moment "AI will create new jobs" stopped being a sufficient answer.

• Meta cuts 700 more jobs across teams as AI infrastructure spending hits $135 billion — Salesforce Ben
• Tech sector cuts 59,000 jobs in Q1 2026 as one in five layoffs are directly attributed to AI — IBTimes

Safety and accountability

OpenAI expanded its bug bounty to cover agentic risks and prompt injection — a tacit admission that traditional security models don't map cleanly to AI agents. Baltimore's lawsuit against xAI over Grok-generated deepfakes is the first municipal legal action of its kind. Accenture's Claude-powered Cyber.AI cut security scan turnaround from days to under an hour. And MIT's warehouse robot work shows AI optimisation quietly delivering measurable gains in physical systems too.

• OpenAI launches Safety Bug Bounty targeting AI agent abuse and prompt injection — OpenAI
• Baltimore becomes first US city to sue xAI over Grok-generated deepfake images — CNBC
• Accenture launches Cyber.AI powered by Anthropic Claude to transform security operations — Accenture Newsroom
• MIT and Symbotic build AI that boosts warehouse robot throughput by 25% — MIT News

If you're building on top of these models, the question is shifting from "which model is best" to "which distribution channel owns the customer relationship" — and whether you're on the right side of that answer before October.

---