AI in 15 — June 16, 2026

Thirty-eight and a half billion dollars. That's not OpenAI's revenue. That's what OpenAI lost last year — nearly eight times the year before. And the company is about to ask the public markets to buy in at close to a trillion.

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

And I'm Marcus, your co-host.

We've lived inside the Fable shutdown all week, Marcus, and today there's a small but telling new wrinkle on it — that's our lead. Then the leaked OpenAI numbers that are scaring everyone. GitHub literally buckling under AI-written code, so badly Microsoft is renting from Amazon. A Chinese lab giving away a trillion-parameter coder the same day Fable went dark. Developers quietly retreating to their own hardware. And the fake job interview that shipped a backdoor.

Why GitHub is renting servers from its owner's biggest rival.

How a free Chinese model landed at the worst possible moment for Washington.

And the AI that caught a state-sponsored backdoor in seconds.

Lead story, Marcus. We know the government pulled Fable 5 and Mythos 5. What's genuinely new today?

The shape of the demand, Kate. We've reported the shutdown and the Amazon angle all week. The new detail clarifies why Anthropic pulled both models rather than comply. The directive wasn't "fix the jailbreak." It applied ITAR-style export rules — the kind reserved for munitions — and required that only U.S. citizens get access. Not U.S. residents. Citizens. That would have barred Anthropic's own foreign-national engineers from touching their own models, inside or outside the country.

So a citizens-only Fable was technically on the table, and they said no.

They said no, Kate. Rather than ship a passport-gated version, they pulled both products entirely and they're still offline as of yesterday. And notice the framing underneath. The fear driving Mythos is that a China-linked group already got access and could distill it — train a cheaper clone on its outputs and release it open-source. Which is the now-familiar pattern. But here's the contradiction this show keeps circling. If a model can be cloned into a free knockoff within months, how durable is any wall built around the original?

And Anthropic says there was no specific rationale given.

That's their claim — that Secretary Lutnick's letter offered no specific national-security justification, and they flew senior technical staff to Washington to defuse it. Whatever the truth, the precedent stands. Treating frontier models as munitions, demanding citizenship checks on who can run software — that's a power that's easy to grant and brutal to claw back. We've said it all week, and a fourth day in, nothing's changed our mind.

Quick hits. And Marcus, this OpenAI leak is the one everyone's chewing on.

It's a stunner, Kate. Leaked financials, first reported by Ed Zitron and lining up with the Financial Times' independent spending figure. In 2025, OpenAI booked about thirteen billion in revenue against thirty-four billion in total costs. Net loss: thirty-eight and a half billion dollars — up nearly eight times from a five billion loss the year before. They spent roughly nineteen billion on R&D, almost six on sales and marketing, and paid Microsoft alone seventeen-point-two billion.

Okay, but revenue also tripled, right? Roughly four billion to thirteen.

It did, and that's the honest bull case, Kate — I won't pretend otherwise. Revenue grew three-and-a-half times while the operating loss only grew about two-and-a-half. On that math, the unit economics are arguably improving, not collapsing. The bear case is just as defensible: the burn is structural, not a one-off, and they're heading for a confidential IPO at a valuation near a trillion dollars.

So which is it?

Genuinely both, depending on your horizon, Kate. If compute costs keep falling and revenue keeps compounding, thirteen billion in losses today looks like Amazon in 1999. If the frontier stalls or the price war from open models bites, it looks like WeWork with better PR. My instinct? Show me the cash, not the keynote. A company losing thirty-eight billion a year is betting the public markets will fund the gap. And it reframes the Anthropic fight — Washington is squeezing the labs at the exact moment their economics depend on selling to absolutely everyone, everywhere.

Marcus, this next one made me laugh. Microsoft is renting Amazon's servers to keep GitHub alive?

It's almost absurd, Kate. Microsoft owns GitHub. Microsoft owns Azure. And yet they're renting AWS capacity to keep GitHub online, because AI-driven usage has overwhelmed it. The driver is autonomous coding agents. GitHub commits are on pace for fourteen billion this year — up from one billion in all of 2025. And pull requests opened by AI agents jumped from about four million last September to over seventeen million by March.

Those numbers are insane. Fourteen times the commits in a single year.

And the infrastructure simply can't keep up, Kate. GitHub logged nine service incidents in May alone, and they've paused new Copilot Pro and Student sign-ups just to protect existing customers. This is the physical bill for the agentic-coding boom coming due. The bottleneck isn't model quality anymore — it's raw compute and storage throughput. And the tell is this: even Microsoft, one of the three biggest cloud operators on Earth, can't provision its own data centers fast enough. They had to go to a rival.

So the bottleneck moved from the brains to the plumbing.

Exactly, Kate. We spent two years asking whether the models were smart enough. Now the question is whether the planet can pour enough concrete and run enough power to let the agents do what they're already capable of doing. The agents are ready. The buildings aren't.

Marcus, and the timing on this next one — almost too perfect. China's Moonshot.

You can't script it better, Kate. On June twelfth — the same day Fable went dark — Moonshot AI shipped Kimi K2.7 Code. A one-trillion-parameter mixture-of-experts model, only thirty-two billion active at a time, 256K context, under a modified MIT license, free to download on Hugging Face. They report a near-twenty-two-percent jump over the previous version on coding tasks, using about thirty percent fewer reasoning tokens.

So Washington locks up America's best, and Beijing gives a near-frontier one away the same afternoon.

That's the strategic message, Kate, stated almost too neatly: you lock yours up, we'll give ours away. But I want to flag a big caveat on air. Those benchmarks are Moonshot's own internal numbers. They skipped independent benchmark submission entirely. So treat the precise figures with real skepticism — self-reported scores are marketing until a neutral party confirms them.

You're always quick to put an asterisk on the Chinese releases.

Because somebody's paying for that trillion-parameter training run, Kate, and it isn't charity. Free is a strategy. The capability is real and genuinely useful — I won't deny developers a good tool. But the move is to erode the West's commercial moat by flooding the zone with free, permissive weights. It's the exact distillation dynamic the whole Anthropic fight is supposedly about, playing out in real time, the very same day.

Which connects perfectly to the next one. Marcus, developers are quietly going local.

And it's the more durable trend, Kate. A Hacker News thread hit over eight hundred points — "Has anyone replaced Claude or GPT with a local model for daily coding?" And the answer from a real, growing cohort is yes. They're running models like Qwen 3.6 — thirty-five billion parameters, only three billion active — on a Mac Studio or a single graphics card, paired with open coding harnesses.

And the quality's actually good enough?

The consensus is sharp, Kate: it's roughly the frontier from eight to twelve months ago. But it's fast, it's private, it's free, and — quote — "enough to get most work done." The dissenters have a fair point too: for professional work, the opportunity cost of not using the absolute best cloud model is still real. You're leaving capability on the table.

But pair it with everything else today.

That's the whole thread, Kate. When access to the top models gets politically fragile — when a government can switch one off on a Friday — "good enough and on my own hardware" stops being a hobby and becomes a strategy. The frontier keeps racing ahead, sure. But the floor of free, capable, ownable models is rising fast. And nobody can revoke what's already sitting on your desk.

Last hit, Marcus, and it's a thriller. A fake job interview that ships a backdoor.

A clean little story, Kate. A developer gets approached by a recruiter at a crypto startup — a GitHub coding challenge as part of the interview. Buried in one file is code that executes whatever a remote server sends it, triggered automatically the moment you run the install command. The recruiter's profile belonged to a real arts journalist. The commit history spoofed a real developer who'd never worked there. Classic hallmarks of a state-aligned fake-recruiting operation — the kind North Korea runs.

And how did he catch it?

This is the satisfying part, Kate. He was suspicious, so he cloned it onto a throwaway server, pointed an AI agent at the code in read-only mode, and it flagged the malicious payload within seconds. So the dual-use story cuts both ways in a single anecdote. AI lowers the bar for attackers — commenters bet the polished lure itself was AI-written. But here AI was the defender, catching in seconds what a tired job-seeker might have run without a second thought.

And the timeless takeaway?

Never run a stranger's code on your real machine, Kate. That advice predates AI by thirty years and it'll outlive all of us. The tools changed. The discipline didn't.

Big picture, Marcus. Pull it together.

One word again, Kate: fragility. Fragility of access. A government forcing America's best model off the shelf. A Chinese lab giving a near-frontier one away. Developers retreating to their own hardware. And OpenAI burning thirty-eight billion a year just to stay in the race while GitHub can't even keep the lights on. Every story orbits the same anxiety — control over frontier AI has become political and contested, and the market's answer is to diversify away from any single gatekeeper.

Whether that gatekeeper is Washington, Anthropic, or OpenAI.

Exactly, Kate. And here's the pro-competition read I'd leave people with. Walls invite cloning — the Kimi release the same day as the shutdown proves it. The durable edge was never the lock on the door. It's execution, and it's the steadily rising floor of cheap, capable, ownable models that nobody can switch off. The West wins this by out-building and out-shipping, not by passport-gating its own software and hoping the rest of the world waits politely.

And the money under all of it?

Is the real test, Kate. Thirty-eight billion in losses, GitHub renting from a rival, seven-figure training runs given away for free. Somebody, somewhere, eventually has to make this math work. Until then, keep one eye on the capability — which is genuinely dazzling — and one eye firmly on the cash. The hype and the burn are both running hot. Both things are true.

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