Good morning. Wall Street is the through-line today: OpenAI has filed confidentially for an IPO just a week after Anthropic, setting up what could be a trio of trillion-dollar offerings if you count SpaceX. Meanwhile at WWDC, Apple quietly admitted it couldn’t build a competitive assistant on its own and handed the keys to Google — a move that may have just commoditized large chunks of the frontier-lab business model.
OpenAI joins the IPO queue. OpenAI confidentially submitted a draft S-1 to the SEC, per The Verge and Wired, with timing left open and CFO Sarah Friar reportedly cooler on the idea than Sam Altman given missed revenue targets and the scale of compute commitments. The company is valued at $852 billion against Anthropic’s $965 billion, and its own announcement carries a notably hedged tone, conceding an IPO may be “a while” off. HN commenters were blunt: OpenAI, Anthropic, and xAI may end up bursting their own bubble once quarterly scrutiny and lockup periods arrive.
Apple bets on Gemini. At WWDC 2026, Apple unveiled “Siri AI” and a redesigned Apple Intelligence architecture built on foundation models co-developed with Google using Gemini, running both on-device and through Private Cloud Compute. The Verge’s writeup and Apple’s own page detail a more conversational Siri with cross-app actions, a dedicated chat app, expanded Visual Intelligence, and CarPlay integration. HN reaction was tepid — many of the demos are things Apple promised two years ago — and EU users appear to be excluded again.
The implications of Apple’s Google deal. One HN commenter caught what may be the real story: Apple reportedly pays Google only about a billion a year to power Apple Intelligence, essentially licensing the IP. If that’s roughly right, Google just handed a competitor the right to operate and distill its flagship models for a rounding error, suggesting consumer-model economics are collapsing faster than the IPO valuations imply. Another commenter put it directly: Apple may have just sherlocked a trillion-dollar private company.
Xiaomi pushes the speed frontier. Xiaomi released MiMo-V2.5-Pro-UltraSpeed, claiming the first trillion-parameter model to break 1,000 tokens/second decode, priced at 3x the standard MiMo-V2.5-Pro for roughly 10x the speed. HN discussion noted that even at 3x, the pricing is shockingly cheap by US standards, though Cerebras is already trialing Kimi K2.6 at 3,000 t/s. One commenter flagged voice as the obvious use case — reasoning models finally fit inside a conversational latency budget.
NotebookLM gets a cloud computer. Google upgraded NotebookLM with Gemini 3.5, adding Google Search-powered source discovery and a “secure cloud computer” per notebook via Google’s Antigravity agentic platform, with new output formats including PDFs, spreadsheets, and generated images. It’s gated to AI Ultra and Workspace customers for now. Combined with the Apple deal, it’s a productive day for Google’s distribution story.
Ed Zitron says the math doesn’t work. Zitron’s latest, AI is slowing down, pegs planned data center buildouts at $9.5–15 trillion and argues the industry needs $3 trillion in annual revenue by 2030 to justify it. HN was split: even commenters who find Zitron’s tone insufferable admit the financial math is directionally sound, while others note his blanket hostility to LLMs undermines the analysis. The piece lands awkwardly well against the IPO news.
Meta quietly pulls face recognition. Wired reports Meta stripped a face-recognition system called NameTag from its Meta AI app — installed on over 50 million phones — one day after Wired first reported its existence. Meta executives had initially called the reporting “incredibly misleading” and claimed the feature didn’t exist; the code had reportedly been embedded since January and was designed to create on-device faceprints of unrecognized individuals.
That’s the morning. Worth watching how Anthropic and OpenAI’s S-1s eventually thread the needle between Zitron-style cost concerns and the Apple-Google deal’s implications for model commoditization.