AI in 15 — June 09, 2026

Eight hundred fifty-two billion dollars. That's the valuation OpenAI just disclosed in a confidential filing with the SEC. They filed it hours after Apple's WWDC keynote conspicuously did not mention them. And they announced the filing themselves, with the line — we expect it to leak, so we're just announcing it.

Welcome to AI in 15 for Tuesday, June ninth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big day, Marcus. The full picture of Apple's WWDC announcement is settling in — and the price tag on the Gemini deal is now public. OpenAI confirms a confidential S-1 filing at roughly eight hundred fifty-two billion. Ed Zitron drops a twelve-thousand-word essay arguing the AI industry needs three trillion dollars in revenue by 2030. Cognition releases a brutal new coding benchmark — Claude Opus 4.8 leads at thirteen percent. Apple quietly retires Core ML. Nvidia and LG break ground on a humanoid robot factory in Korea. And GitHub fell over for two hours yesterday.

Apple's price is a billion. OpenAI's price is the public market.

The bear case finally gets written down.

And a new benchmark scores AI code like a tech lead.

Lead story, Marcus. We covered the WWDC keynote yesterday — what's actually new today?

Two things landed overnight that change the read, Kate. First, the price on the Gemini license is now reported as roughly one billion dollars per year. One billion. That's it. For distribution to two billion Apple devices. The cheapest distribution deal in tech history if it holds. And second, OpenAI confirmed a confidential S-1 filing with the SEC at approximately eight hundred fifty-two billion dollars, with Goldman Sachs and Morgan Stanley as lead underwriters. ChatGPT is at roughly nine hundred million weekly users, fifty million paying subscribers, and about two billion dollars a month in revenue.

And the timing of the filing.

Almost theatrical, Kate. Apple does a two-hour keynote where it never says the words OpenAI or ChatGPT, deprecates the existing ChatGPT-Siri pilot, and routes the future of Siri through Google. Hours later, OpenAI puts out a statement saying — quote — we recently submitted a confidential S-1. We expect it to leak, so we're just announcing it. They also said the filing is, quote, not a signal that a listing is near. Which is the kind of sentence you write when you're trying to lower expectations on a date that's actually coming fast.

So what's the real story, Marcus?

The power structure inverted in twenty-four hours, Kate. For two years OpenAI was the consumer brand. ChatGPT was the verb. Apple just demonstrated that being late was a strategy — let the labs burn a hundred billion training models, then license one for a billion a year and embed it under your privacy wrapper. And the architectural detail matters. Apple runs smaller models distilled from Gemini on-device, heavier requests through Private Cloud Compute, with outside experts able to audit that Google never sees user data. That's a genuine privacy story, and it sets a new floor for what consumers will expect from any personal AI. Meanwhile OpenAI and Anthropic are both filing for public markets in the same week — Anthropic at nine hundred sixty-five billion last Monday, OpenAI at eight hundred fifty-two billion now. Two of the three frontier labs racing to public markets the same week Apple announced it doesn't need either of them. That's the subtext Wall Street will absorb.

Quick hits. Marcus, Ed Zitron's essay. He's been waiting for this moment.

Timed almost too perfectly for Apple-keynote day, Kate. Twelve thousand words on his Where's Your Ed At newsletter, headlined AI Is Slowing Down. The central claim — the AI industry needs roughly three trillion dollars in cumulative revenue by end of 2030 to justify the compute commitments already booked. He pegs OpenAI's 2026 compute spend at fifty billion, Anthropic's at thirty to fifty. Their combined projected 2026 revenue is around sixty billion. To hit break-even targets he calculates they need roughly four hundred ninety-six percent growth by 2029. And he flags a Q1 2026 OpenAI non-GAAP operating margin of negative one hundred twenty-two percent.

And the data on user growth.

He claims ChatGPT user growth has plateaued, Kate, and Oracle's seven-point-one gigawatt buildout for OpenAI carries an estimated cost between three hundred forty and seven hundred billion dollars. Hacker News reaction was unusually split — four hundred eighty points, five hundred comments. Defenders say the arithmetic is right. Critics say his impassioned-ranter tone obscures genuine analytical weaknesses. The most cited comment in the thread pointed out that Apple's one-billion-a-year Gemini license actually proves Zitron's broader thesis — foundation models are racing toward commoditization at the wholesale layer.

Why this matters.

Zitron is wrong about plenty, Kate, but the unit-economics critique is the one OpenAI's S-1 will have to answer in plain prose, on the record, to public-market investors. The bear case has now been written down clearly enough to be cited in equity research. Whether this gets dated later as a Buffett-warns-about-Cisco moment, or as a cautionary tale about angry bloggers, depends entirely on what unit economics OpenAI prints in its first public quarter.

Marcus, the Cognition coding benchmark. Talk me through it.

Genuinely interesting release, Kate. Cognition — the company behind Devin — published FrontierCode yesterday. The pitch is that existing coding benchmarks measure whether AI-written code passes tests. FrontierCode measures whether real open-source maintainers would actually merge it. Twenty-plus maintainers from thirty-six flagship repos each spent over forty hours building tasks on their own codebases. The benchmark scores along six dimensions — behavioral correctness, regression safety, mechanical cleanliness, test quality, scope appropriateness, and design quality. A hundred fifty tasks split across Extended, Main, and Diamond tiers. Cognition claims an eighty-one percent lower false-positive rate than SWE-Bench Pro.

And the scores.

On the hardest Diamond tier, Kate — Claude Opus 4.8 leads at thirteen-point-four percent. GPT-5.5 at six-point-three. Gemini 3.1 Pro at four-point-seven. Kimi K2.6, the best open-source entry, at three-point-eight. Cognition notes GPT-5.5 is significantly more token-efficient than Opus despite the lower score, which matters for cost.

So thirteen percent is the top score.

Which sounds embarrassing if you believe the developers-are-obsolete line, Kate, but the story is the trajectory. SWE-Bench Verified went from twelve percent to over eighty in a year. Frontier benchmarks keep saturating in months, and labs now need new evals that measure taste rather than just correctness. FrontierCode is the first major benchmark explicitly scored like a tech lead. Worth noting — the benchmark itself was largely built using Claude. And Anthropic's commanding lead here lands the same week as its IPO filing. Good marketing if you're trying to justify nine hundred sixty-five billion.

Marcus, Apple's Core AI framework. This is the developer story underneath the Siri story.

And it may matter more long-term, Kate. Tucked under the consumer-facing Siri AI announcement, Apple is sunsetting Core ML — nine years old, introduced in 2017 — and replacing it with Core AI. The new framework lets apps run on-device large language models across CPU, GPU, and Apple's Neural Engine through a unified PyTorch-to-Apple-Silicon conversion pipeline. The Foundation Models API now gives developers a three-line Swift call into a significantly bigger on-device base model, with expanded context window, structured output, streaming, tool calling, and — this is new — on-device fine-tuning that never leaves the device.

And the pricing twist.

This is the part that will reshape the iOS app economy, Kate. Apps with under two million downloads get free server-side Foundation Models access via Private Cloud Compute. Free. Larger apps pay a rate Apple hasn't published, with end-user quotas, where iCloud Plus subscribers get higher quotas. Apple also bundled curated open-source models — Qwen, Mistral, SAM3 — pre-optimized for Apple Silicon and discoverable through the framework. And the new platform supports MCP — Anthropic's Model Context Protocol — as a default.

Why this matters.

Three things, Kate. First, this is Apple recasting itself as the substrate for AI apps, not a model competitor. Free on-device intelligence for indie developers undercuts OpenAI's API economics and the entire third-party AI assistant market. The next generation of iOS apps stops calling OpenAI's API entirely. Second, combined with the Gemini deal, it's a clean two-layer strategy — Google for the heavy reasoning the user can see, Apple silicon for everything else. Third, the MCP adoption is a sleeper story. Anthropic's open standard just became platform-default on a billion devices. Anthropic loses the iPhone consumer pilot but gets the protocol layer of the entire Apple ecosystem. Not a bad consolation prize the week of your IPO filing.

Marcus, Nvidia and LG. Humanoid robots in Korea.

Announced June seventh, trending into yesterday, Kate. Nvidia and LG Group are building a full-stack AI infrastructure facility in South Korea spanning humanoid robotics, autonomous vehicles, data centers, and GPU cloud. The robotics piece uses Nvidia's Isaac Sim and Isaac Lab for simulation, Isaac GR00T as the foundation reasoning model for humanoids, and Cosmos world-foundation models to generate synthetic training data. The infrastructure layer uses Nvidia's DSX modular AI factory platform with liquid-cooled Blackwell GPUs. LG is also using NeMo and TensorRT-LLM to develop EXAONE — a Korean sovereign-AI model. Financial terms not disclosed.

Why this matters.

Second major non-China, non-US robotics build-out announcement in as many weeks, Kate. South Korea joining Japan, Germany, and the US as a physical-AI hub matters because it keeps the humanoid supply chain inside the Western and allied bloc rather than ceding it to Chinese manufacturers like Unitree and Xpeng. For Nvidia it's another sovereign-AI customer. For LG it's a way to leapfrog the manufacturing-robotics market dominated by Japanese and Chinese players. Hacker News top comment captured the consumer skepticism — quote — we're going to be conditioned to think twenty to fifty thousand dollars for a robot is okay to spend. But the business case here is industrial, not domestic. And the strategic case is keeping a critical category out of Beijing's supply chain before it becomes the next solar panels.

Marcus, GitHub fell over yesterday.

Two-hour outage Monday morning, Kate. Took down git operations, pull requests, and the Copilot API. GitHub posted incident notes but Hacker News noticed the status-page numbers may have been cleaned up. A third-party tracker showed pull-request availability at ninety-five-point-nine percent over the last ninety days, versus GitHub's claimed ninety-nine-point-six-one. The bigger story — with LLM coding agents now making thousands of API calls per workflow, GitHub uptime is rapidly becoming AI-pipeline-critical infrastructure. A two-hour GitHub outage in 2022 was an inconvenience. In 2026 it stalls every agent that needs to clone a repo, open a PR, or read CI status. The reliability bar GitHub needs to clear just went up by an order of magnitude, and the SLA disclosure debate is going to get louder.

Marcus, last quick hit. There's a font that mangles for AI.

Small viral curiosity, Kate. A developer released SoulsOnly.ttf — a custom TrueType font designed to be human-readable but break optical character extraction. Plus keyboard firmware so you can actually type in it. Mostly conceptual, but it captures the brewing AI-unfriendly backlash where some creators are deliberately making content harder for crawlers to consume. The bonus irony — Hacker News noticed the commit history shows the maintainer and Claude committed two weeks ago. So the anti-AI font was built with the very tools it tries to defeat.

Pair that with one more.

Worth flagging quickly, Kate — Microsoft's MAI-Code-1-Flash started rolling out across all Copilot tiers last week. Five billion parameters, two hundred fifty-six kilo context window, adaptive reasoning. Microsoft claims fifty-one-point-two percent on SWE-Bench Pro versus Claude Haiku 4.5's thirty-five-point-two. Microsoft's first wholly in-house coding model, and the strategic significance is Microsoft is finally weaning Copilot off OpenAI. Combined with Apple's Gemini deal and the new Core AI framework, the message from the platform owners is unmistakable — they are insourcing or commodity-shopping their model stacks, and the labs are wholesalers now.

Big picture, Marcus.

Today is the day the AI industry's power structure visibly inverted, Kate. Apple — which had been mocked for a year for being late to AI — used WWDC to demonstrate that being late was a strategy. Let the labs burn a hundred-plus billion training models, then license one for a billion a year and embed it under your privacy wrapper on two billion devices. The same day, the two leading labs are both quietly filing for public markets — not from a position of strength, but because private capital can no longer absorb their burn rate. Ed Zitron's essay isn't gospel, but it's now the bear-case template every IPO analyst will reach for. Real value keeps moving down the stack — to devices with Apple, to operating systems with Google, to industrial robotics with Nvidia and LG, to coding tools with Microsoft and Cognition.

And the libertarian read.

Apple's strategy is straight out of Joel Spolsky's playbook, Kate — commoditize your complement. Turn foundation models into a wholesale input the way IBM commoditized Intel and Microsoft for itself in the eighties, then ate the margins for thirty years. Markets are doing exactly what markets are supposed to do. The platform owners are negotiating with the model labs, not bowing to them. The OpenAI and Anthropic IPO pitches don't want investors to notice that, but Wall Street has now noticed, and the S-1s will have to explain it. The lab-as-product era peaked somewhere in the last six months. What comes next is the platform-distribution war. And Apple just fired the opening shot — for a billion dollars a year, paid to Google, while keeping the operating system, the silicon, the privacy story, and the right to switch providers anytime Gemini gets sloppy.

That's your AI in 15 for today. See you tomorrow.