Support Us & Upgrade to the Operator Edition

Five weeks after giving away the most downloaded AI app in America for free, Meta just locked its best model behind a subscription. On the same Monday, Microsoft told Nvidia it's moving out, politely, with a 2027 deadline.

This week, we'll deep dive into who actually pays for AI, and who's tired of paying Nvidia.

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free

The 5.5 models think differently, cache differently, and charge differently. Habits that worked a few months ago now waste usage and can even give you worse answers. This breakdown from Jay went through Anthropic's own 5.5 guides and hundreds of power-user tips. Here are the five fixes worth doing today:

💡 Pro tip: Type /usage to see exactly what's eating your limits, broken down by skill, subagent, plugin, and MCP server. It even flags any habit taking 10% or more.

Watch the full 23-mistake breakdown here

Want the version of this that actually runs in your business?

The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.

This week: Learn how to use Claude inside of Google Workspace

👉 Upgrade to The Operator Brief

💰 The CEOs of Intel and Broadcom are warning that the next AI bottleneck might not be chips at all but money, as a capital crunch threatens the buildout. The industry spent three years worrying about running out of silicon. Now it turns out the real bottleneck is just money.

🧬 TikTok's CEO says ByteDance is aiming its AI at scientific breakthroughs. Consider this a signal that the company behind your For You page wants to be taken seriously as a research lab. The algorithm feeding you dance trends is about to start curing diseases.

💵 The FTC is sending refunds to Cleo AI users misled by the cash advance app, with PayPal payments starting October 27. Check to see if you qualify before the 30-day window closes. A cash-advance AI promising easy money just got a real-world reminder that financial regulators still write the checks.

🛒 Amazon is reportedly circling Decart for around $7 billion, a 3-year-old startup that already walked away from Anthropic and makes AI run on non-Nvidia chips. Amazon dropping $7 billion on a startup that bypasses Nvidia chips shows just how desperate Big Tech is to break Jensen Huang's monopoly.

Meta just launched Muse Wonder, and for the first time, the smartest version of Meta AI isn't free.

The model, codenamed "Watermelon" internally, will power the paid tiers of the Muse personal agent app, which shot to the top of the App Store after launching on September 8.

Here's what you need to know:

Wang is pitching Wonder as the consumer play. While everyone else is building coding agents, Meta wants the agent that books your flights, cancels your gym membership, and negotiates your phone bill.

Here's the uncomfortable part: the model is "state of the art" mostly on Meta's own account. Independent benchmarks will take weeks. Meanwhile, some users have complained on social media that Muse accessed personal messages without authorization, a claim Meta disputes. And Meta wants you to hand that same agent your bank login and a monthly card charge.

🔮 Prediction: Muse Wonder is less a model launch than a pricing experiment, and the answer matters for every AI company, not just Meta.

For two decades, Meta's entire business has run on one rule: the user never pays, the advertiser does. Muse broke that rule quietly. The free tier gets you in the door. The good brain costs $20. That's the ChatGPT playbook, and it tells you Meta has done the math on what agents cost to run. An always-on agent sitting in its own virtual machine, browsing, clicking, and filling forms around the clock, burns far more compute than a feed of Reels. Ads alone can't cover that bill.

But Meta has an advantage OpenAI and Anthropic don't: 3 billion people already inside its apps. If even 1% convert, that's 30 million subscribers, which would put Muse in the same league as ChatGPT's paid base overnight.

Here's where I think it breaks, though. Trust. People will pay $20 for something that saves them time. They will not pay $20 to an agent they suspect is reading their DMs. The first viral "Muse did something I didn't ask for" story with a real dollar figure attached will hit conversion harder than any benchmark.

So expect two things in the next 90 days: Meta leaning hard on privacy controls in its marketing, and OpenAI responding with a cheaper agent tier to undercut it. The personal agent war isn't going to be won on intelligence. It's going to be won on who people let into their inbox.

💬 Hit reply: Would you pay Meta $20 a month to run your life admin? Yes, no, or "absolutely not, it's Meta"? I'm reading every answer.

Check out Muse Wonder here

Microsoft just unveiled Maia 300 and said its own chips will become "the primary AI inference system" for Copilot and the rest of its products.

Azure chief Scott Guthrie laid out the plan in a blog post on Monday, with large-scale deployments starting in 2027. Satya Nadella called it "our biggest bet yet on our own silicon."

What stood out:

The current Maia 200 already powers Microsoft 365 Copilot and runs OpenAI models. GitHub Copilot and Microsoft Foundry move onto it by the end of the year.

Here's the asterisk: just last week, Nadella and Jensen Huang shared a stage to launch Nvidia-powered Windows PCs. Evercore ISI estimates Maia will account for only about 10% of Microsoft's AI chip spend by early 2027. This is a breakup on a very long timeline.

🔮 Prediction: Nvidia won't lose Microsoft as a customer. It will lose Microsoft as its most profitable customer, and that's a different, slower kind of damage.

Watch the split Microsoft is setting up. Its own products (Copilot, GitHub, Office) move to Maia, where it controls the cost. Nvidia's expensive Vera Rubin systems get reserved for Azure customers willing to pay premium prices for them. That means Microsoft keeps buying Nvidia, but increasingly as a reseller, not as the end user eating the margin.

Now look at the timing. In the same week, Amazon is reportedly in talks to buy Decart for around $7 billion, a startup whose entire job is making AI models run on non-Nvidia chips like Trainium. Google already sells TPUs to outside customers. OpenAI shipped a chip with Broadcom in June. Every hyperscaler is attacking the same weak spot: CUDA lock-in.

The real constraint isn't design anymore. It's power and fabs. A chip that does 40% more work per watt matters enormously when electricity, not demand, is what's capping data center growth. That's why "multi-gigawatt" is in the headline, not "faster."

My bet: by 2028, the question on earnings calls shifts from "how many Nvidia GPUs did you buy?" to "what percentage of your inference runs on your own silicon?" Microsoft just told you its answer will be "most of it."

💬 Hit reply: Is Nvidia's moat real, or are we watching the slow start of its Intel moment? Tell me which side you're on.

Read Microsoft's full announcement here

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI