AI in 15 — June 06, 2026

Fourteen billion dollars in passive index buying. That's what S&P 500 inclusion would deliver to SpaceX alone. And on Thursday, the keepers of America's most important stock index just said no.

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

And I'm Marcus, your co-host.

Big day, Marcus. S&P Dow Jones blocks fast-track entry for SpaceX, OpenAI and Anthropic. Google DeepMind ships Gemma 4 QAT — frontier-class AI in one gigabyte of memory. Google agrees to pay SpaceX nine hundred twenty million dollars a month for compute. Claude Opus 4.8 finds a four-year-old counterfeit bug in Zcash and the token crashes thirty-eight percent. The rsync community erupts over Claude co-authorship. A leaked Microsoft memo says the quiet part out loud — make people addicted to Scout. And Apple's WWDC starts Monday with a Gemini-powered Siri.

Wall Street tells the AI darlings to wait.

Google rents GPUs from a rocket company.

And Microsoft writes "addictive" into the strategy doc.

Lead story, Marcus. Walk me through the S&P 500 decision.

This is a real institutional pushback, Kate. On Thursday, S&P Dow Jones Indices rejected a proposal that would have created a fast-track path into the S&P 500 for the three most-anticipated tech IPOs of the decade — SpaceX, OpenAI and Anthropic. They refused to waive the GAAP profitability rule — four consecutive quarters of positive net income — and kept the twelve-month seasoning wait for newly public companies fully intact. Nasdaq and FTSE Russell already softened their rules to accommodate mega-cap newcomers. S&P said no.

And the financial impact, Marcus.

Enormous, Kate. Bloomberg Intelligence estimates S&P 500 inclusion would trigger roughly fourteen billion dollars in passive index buying for SpaceX, more than eight billion for OpenAI, and four-point-six billion for Anthropic. SpaceX reported a four-point-nine-four billion loss in 2025. Both AI labs spend more than they earn — Anthropic is paying SpaceX one-point-two-five billion dollars per month for compute through May 2029. None of them qualify under current rules. And the timing is sharp. Anthropic confidentially filed its draft S-1 with the SEC on June first, less than a week after closing a sixty-five billion Series H at a nine hundred sixty-five billion valuation, on a forty-seven billion annualized run-rate.

So is this Wall Street finally saying enough?

In a measured way, Kate, yes. This is the financial system pushing back on AI exceptionalism. The magnificent seven rally has masked the fact that the most-hyped AI companies still lose money at staggering rates, and the keepers of the world's most important equity index just said size and hype don't substitute for cash flow. For retail investors holding S&P 500 index funds, it's a quiet consumer-protection win. For the AI labs, it means hundreds of billions in deferred passive demand and IPO pricing that has to lean on active institutional buyers instead of the index machine. The libertarian read — discipline is showing up exactly where it's supposed to, in the rulebook of a private index provider, not in a federal mandate. The honest read — public markets are about to test whether revenue-without-profit is a story investors still want to buy.

Quick hits. Marcus, Gemma 4 QAT.

Significant on-device release, Kate. Google DeepMind yesterday published Quantization-Aware Training checkpoints for the Gemma 4 family, designed for phones, laptops, and consumer GPUs. The headline — Gemma 4 E2B shrinks to roughly one gigabyte of memory in the new mobile format. E2B is three-point-two gigs, E4B is five gigs in the laptop Q4_0 format. Google says QAT preserves about ninety-five percent of full-precision quality while cutting memory footprints by forty percent, beating standard post-training quantization at the same four-bit size. They've introduced a mobile-specialized quantization schema with pre-calculated static activations and targeted two-bit quantization for token-generation layers.

And the developer reception.

Hot, Kate. Top story on Hacker News at three hundred eighteen points. Unsloth has already published competitive quants approaching full-precision accuracy. Simon Willison demonstrated running E2B locally on a Mac via litert-lm with a single uvx command. And the timing is no accident — three days before Apple's WWDC, where Siri is rumored to be powered partly by Google models under a reported one-billion-dollar-a-year licensing deal.

Why it matters.

On-device AI just crossed a real threshold, Kate. A one-gig model with useful quality fits in a phone's RAM budget without breaking everything else. If Apple does showcase Gemini-derived models on iPhones at WWDC Monday, this release is the technical scaffolding underneath. Strategically, open-weights Gemma is the Western spear point against Chinese open-source models like Qwen and DeepSeek, and against closed labs that refuse to release weights at all.

Marcus, the Google-SpaceX compute deal. The number is wild.

Nine hundred twenty million a month, Kate. CNBC confirmed Friday that Google will pay SpaceX roughly thirty billion dollars total for access to about a hundred and ten thousand NVIDIA GPUs, plus CPUs and supporting infrastructure housed at SpaceX-operated data centers. The contract runs October 2026 through June 2029. Google can terminate if SpaceX fails to deliver committed capacity by September thirtieth, and after year one either side can exit on ninety days notice. This deal sits alongside Anthropic's previously reported one-point-two-five billion per month arrangement with SpaceX.

The strangeness of this is staggering.

On multiple fronts, Kate. A hyperscaler renting GPU capacity from a rocket company. Anthropic and one of the largest cloud providers in the world both becoming SpaceX customers. And the suggestion that Elon Musk's xAI infrastructure is effectively being sub-let to Google — a direct competitor — right before SpaceX's IPO. Two takeaways. First, GPU supply is so constrained that Google, which builds its own TPUs, is willing to pay a competitor's affiliate nearly a billion a month to get NVIDIA silicon online by fall. Second, this turns SpaceX into a major AI infrastructure landlord just as it courts public investors, validating a one-point-seven-five trillion target valuation with a new high-margin recurring revenue line. The deeper signal — almost none of the AI labs' revenue stays in the AI labs. It flows straight to GPU owners and power utilities.

Marcus, this Zcash story is wild. Claude finds a four-year-old counterfeit bug?

One of the strongest concrete proof points yet, Kate. Security engineer Taylor Hornby disclosed yesterday that he used Anthropic's Claude Opus 4.8 to audit the Zcash Orchard circuit — the cryptographic component powering Zcash's most advanced shielded-transaction pool — and discovered a two-line flaw that has existed since Orchard activated in May 2022. The bug would have allowed an attacker to mint an unlimited number of counterfeit ZEC tokens inside the shielded pool, with no on-chain signature, completely undetectable. Hornby wrote a full working exploit in a local test environment that produced unlimited untraceable counterfeits. Shielded Labs shipped an emergency hard fork by June first.

And the market.

Brutal, Kate. ZEC fell roughly thirty-eight percent in twenty-four hours and as much as fifty percent over forty-eight. Shielded Labs admits there's no cryptographic way to know whether the flaw was exploited before the patch, so they're proposing a network upgrade with new supply-integrity accounting.

Why this matters.

Orchard had been reviewed multiple times by professional cryptographers across four years and nobody caught a two-line forgery flaw, Kate. Claude Opus 4.8 did, in a targeted review. This dovetails with the Project Glasswing expansion we covered Wednesday — Anthropic says Mythos Preview has already surfaced over ten thousand high- and critical-severity vulnerabilities across major operating systems and browsers. The defensive upside is huge. The offensive upside, in the hands of an adversary with the same class of model, is what keeps the safety teams up at night. Frontier models as production-grade cryptographic auditors is no longer hypothetical.

Marcus, the rsync drama. This one's a culture story.

And the data is interesting, Kate. A storm broke this week after rsync users discovered backups failing under version 3.4.3, and that dozens of recent commits were co-authored by Claude alongside maintainer Andrew Tridge Tridgell. Critics accused Tridge of vibe-coding production-critical backup software. But a statistical analysis published yesterday by Alexis Purslane examined thirty-six rsync releases using a severity-weighted bugs-per-ten-commits metric, ran exact permutation tests, and found Claude-era releases were not statistical outliers. The biggest single offender was actually version 3.4.1 — the release before Claude commits began — at thirty-nine-point-three-nine severity per ten commits. Nobody noticed because there was no AI to blame.

Tridge's response.

He pushed back in a Medium post titled "rsync and outrage," Kate. Said Claude was used for grunt work — test suite scaffolding, code review of security patches — and the regression spike in 3.4.3 came from absorbing a flood of security reports, not from AI writing the code. Hacker News commenters split predictably. Defenders said this is the future of OSS maintenance. Critics said disclosing AI involvement just invites mob harassment, and others will quietly stop attributing.

The bigger story.

Open-source projects now face a choice between transparent AI co-authorship — and getting harassed for it — or quietly hiding the assist, Kate. The data so far suggests the panic is overcalibrated. The social pressure is not. Expect more projects to drop the Co-Authored-By Claude trailer to avoid the drama. Which is exactly the opposite of what anyone genuinely concerned about AI code quality should want.

Marcus, the Microsoft memo. Did this actually say make people addicted?

It did, Kate. 404 Media obtained an internal Microsoft strategy document titled "ClawPilot — Overview and Plan with Project Lobster," laying out Phase 1 of the launch plan for Scout, Microsoft's new always-on Microsoft 365 personal agent we covered on Wednesday. The document states the explicit goal — quote — make people addicted. Continue shipping the standalone ClawPilot experience. Pilot the UX, grow the user base, and build the skill and tool ecosystem that makes people depend on it daily. The document was authored by Microsoft execs Omar Shahine — who leads the Scout project — and Jakob Werner. Apparently with help from an AI writing tool.

And Nadella?

Told The Information he was, quote, not sure what this document is or who is writing and leaking this nonsense, Kate. 404 noted, dryly, that Nadella absolutely knows who wrote it, because one of them is his project lead. One anonymous Microsoft employee called the explicit references to addiction very troubling.

Why it matters.

Every consumer product company chases engagement, Kate. But writing addictive as a top-line strategic objective — and getting it leaked — is a regulatory and PR own-goal at exactly the moment EU and state lawmakers are deciding what AI products can and can't do. Expect this document to surface in Senate hearings and the FTC's AI consumer-protection inquiries. It also hands competitors — notably Apple, with its we-don't-sell-engagement pitch — a free marketing weapon a week before WWDC. And it's a useful contrast with Friday's OpenAI Dreaming launch. OpenAI is building memory that makes ChatGPT stickier. Microsoft is writing the stickiness goal into the strategy doc. One of those companies is going to look much better in front of a Senate panel than the other.

Last one, Marcus. Apple WWDC starts Monday.

Worth flagging on a Saturday, Kate. WWDC 2026 opens Monday June eighth — Tim Cook's final developer conference as CEO. The headline expectation — a fully rebuilt Siri powered by a custom one-point-two-trillion-parameter Google Gemini model under that reported one-billion-a-year licensing deal. Plus a default-AI-provider picker letting users route queries to Claude, ChatGPT, Grok, Copilot or Perplexity. Apple plans to run the new Siri partly on Google Cloud using NVIDIA chips. The strategy — Apple wants to be the AI distribution platform, not the AI model.

So Gemma 4 QAT this week was setup.

Looks coordinated, Kate. On-device Gemma for offline tasks. Custom Gemini in the cloud for heavy lifting. A provider picker for users who prefer Claude or ChatGPT. Apple gets a credible AI story without spending five years closing the model gap. Google gets distribution on a billion-plus iPhones. NVIDIA wins either way. We'll cover the actual announcements Monday.

Big picture, Marcus.

Three threads pull through today, Kate. First — money meets math. S&P 500 won't waive its profitability rule for AI darlings. Google is paying SpaceX nearly a billion a month for GPUs. Anthropic pays one-point-two-five billion. The AI economy is colossal, but most of the revenue flows past the labs into infrastructure. The IPOs will test whether public-market investors care. Second — models are getting useful in two opposite directions at once. Gemma 4 QAT pushes capable AI down to one gigabyte on a phone. Claude Opus 4.8 finds cryptographic bugs that four years of human auditors missed. The high end gets sharper. The low end gets cheaper. There's no comfortable middle ground left for human cognitive labor. Third — the trust war is just starting. Microsoft writes addictive into a strategy doc. rsync maintainers get harassed for disclosing AI co-authorship. The technical race is being won. The legitimacy race is wide open. The libertarian read — S&P discipline, market pricing of compute, open-source Gemma, all moving in the right direction without a federal mandate. The honest read — when a leaked memo at the world's most valuable software company names addiction as a strategic objective, the case for self-regulation gets harder to make every week.

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