Starting soon, a company that already has a big Google Cloud commitment can spend it on OpenAI models. Same invoice that pays for Gemini. And that blurs the line between the labs and the clouds that host them. You're listening to AI Daily Briefing, and it's Thursday, October 22nd. OpenAI moves into Google's Model Garden, Meta creates a new kind of stock to pay for AI, Anthropic ships an agent that messages you first, and Nvidia pays its suppliers to build faster.
OpenAI, announcing the partnership on its own site Wednesday:
Google Cloud customers will be able to purchase OpenAI frontier models via model garden, transacted through the Google Cloud Marketplace. Enterprises with Google Cloud spend commitments will be able to draw down their GCP commitments for these purchases.
That second sentence is the real deal. Big companies sign multi-year spend commitments with their cloud, and that money has to land somewhere. Now it can land on OpenAI without a new contract or a new procurement fight. Per AI Compass, the lineup is GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, plus Codex workflows. OpenAI says its models will also show up soon in Google Antigravity and the Gemini Agent, with general availability planned in the coming weeks. And the last line of the post: Microsoft remains OpenAI's primary cloud partner. That's the line you add when you know Redmond is reading. Look at what Google gets. Per the Stillflood digest, Google Cloud says it's now the only major hyperscaler offering Gemini, Claude and OpenAI models at the same time. On October 14th we flagged Antigravity as Ultra-only and built around Argon. It's about to route to a rival's models. Thomas Kurian's pitch is the right model for any given job, at lower cost. Which is how I'd build anyway. Cheap calls go to the cheapest model, hard calls to whatever wins your evals. But nothing is generally available yet, and OpenAI's post has no pricing, so I can't tell you whether GPT through Google costs more than going direct.
The Verge, on the third-quarter report Meta released Wednesday:
"The scale of investment required to compete, sustain growth, expand infrastructure, and pursue new strategic initiatives amidst the AI transformation taking place in the industry is significant," Meta says in the press release.
Translation: the AI bill is big enough that Meta wants a new way to pay it. It's a new class of non-voting shares. Per The Verge, the stock will fund AI work, potential acquisitions and employee compensation, with trading starting in 2027. CFO Susan Li said on the call it will expand Meta's funding tools. For scale, the Stillflood digest, citing CNBC, Fortune and Business Insider, puts quarterly revenue at sixty-four point six billion, up twenty-six percent, with full-year capex narrowed to one hundred thirty-five to one hundred forty-five billion dollars. Non-voting is the tell. Meta gets fresh currency for hiring and deals, and the new holders get no say in how it's spent. This is also the next beat in our Muse thread. We've followed the models since Muse Wonder on October 12th, through Spark Open and, yesterday, Glimmer. On the call Meta gave the consumer number: TechCrunch reports the Muse app has over ten million users, with shopping a top activity. Per PoYo's summary, Zuckerberg said about a quarter of people who set up Muse use it daily, and the average purchase through it tops two hundred dollars. Meta plans to keep Muse free for most users and take a cut of commerce. It also expects Muse's compute cost efficiency to improve more than tenfold over the next year. That's the number I'd watch. If the tenfold slips, free plus a commerce cut doesn't pencil. Ten million is Meta's own count from an earnings call. And we still haven't seen independent evals of any Muse model.
Maximilian Schreiner at The Decoder, on Anthropic's new agent app:
Anthropic says Duet can only read, draft, and prepare on its own. Before it sends a message, makes a payment, signs a document, or deletes anything, it shows exactly what it plans to do and waits for the user's approval. Trading securities and moving money between accounts are off the table, even with approval.
Claude Duet opened as a research preview Wednesday. It's a standalone app on iOS, Android and the web, built on Claude Opus 5.5, with a waitlist for US Pro and Max subscribers. Team and Enterprise aren't included yet. It runs on its own cloud machine with its own browser, so it keeps working after you close the laptop. The pitch is that it comes to you. A morning brief across your connected apps, then a ping when something needs a decision. It can reach more than six thousand connectors, including Gmail, Slack, Stripe, Docusign and Shopify, and purchases go through Stripe Link with a virtual card. Schreiner names the field: Grok Bot since August, then Meta's Muse and OpenAI's Dots in September. Credit where due, Anthropic printed its own failure modes. It concedes Duet can misread a thread, misjudge urgency, and miss things that arrive between checks. Those checks run every few hours because, per Anthropic, constant checking would burn through subscription limits. Background work counts against the same plan limits as the rest of Claude. And per AI Compass, there's no model selector, and you can't watch or take over its cloud browser. So always-on, but metered. Approval on every outbound action is the right default. The risk is the stuff it decides isn't worth your attention. That's where I'd want an audit trail.
Dealroom News, drawing on reporting from Korea's Chosun, on what Nvidia announced at GTC Berlin:
NVIDIA signed a $1.3 billion deal with Amkor to raise advanced packaging capacity in South Korea more than 300% by the end of 2028. It paid KYEC a $1.1 billion advance to build testing facilities in the US. SPIL agreed to grow optical-engine capacity 15-fold by mid-2027.
So Nvidia isn't buying GPUs here. It's buying the factory time to package, test and connect them. Six suppliers in all. Ibiden gets a multibillion-dollar advance to triple its Ono plant in Japan, with new output from early 2028, and the substrate deals together aim to lift capacity two and a half times over three years. Jensen Huang's line: "Demand is extraordinary. Capacity is the constraint." Dealroom also flags the exposure. In August the Wall Street Journal reported Nvidia's supplier commitments hit two hundred seventy-nine billion dollars in its latest quarter, up from one hundred nineteen billion the quarter before. If demand or timelines slip, those advances sit on Nvidia's books. Yesterday it was Modal buying Vera CPUs. Today it's packaging lines in Korea. Same strategy, sell the whole stack. And the demand side showed up on the same stage. Google Cloud's A5X instances on Vera Rubin NVL72 went live Wednesday, with general availability planned for November. Google says its Virgo network can scale to eighty thousand Rubin GPUs at one site and nine hundred sixty thousand across several. Diego Almada Lopez at Crypto Briefing makes the key point: that's a capacity ceiling, not chips running. The claimed ten-times-lower cost per token is against the previous generation, and it's the vendors' number. One more from Berlin. SemiAnalysis says, per TechSpot, that some neoclouds that added AMD or TPU chips were turned away from the conference. Unnamed neoclouds, so treat it as a rumor. But it fits a company that wants every layer, from substrates to the guest list.
If this show has earned a slot in your morning, take ten seconds to rate it and leave a short review in your podcast app. That's how other builders and investors find us.
Today was OpenAI on Google's shelf, Meta's new stock, Claude Duet, and Nvidia's supplier deals, and every source sits in the show notes. AI Daily Briefing comes to you from Lantern Podcasts. We're back Friday.