Two stocks now own 15% of the S&P 500, Meta hangs an AI agent off your bag, MoonPay buys its SEC licenses, and China puts a data center in orbit
Hello, dear TEA-mates! Here is what you need to know today.
1. 📈 Two Stocks Now Own 15% Of The S&P 500
Apple and Nvidia together make up more than 15% of the S&P 500, the highest two-stock concentration the index has ever seen, a figure flagged by Creative Planning president Peter Mallouk. For comparison, Microsoft and General Electric peaked at a combined 9.1% in the run-up to the dot-com bust. Apple hit an all-time high of $345 on Tuesday and is up 25% year to date, lifted by the iPhone 18 launch and its first foldable, the Duo. Nvidia is up 21% year to date after telling investors in late August that its business should grow at least 70% next fiscal year, with CEO Jensen Huang saying growth would top 100% if not for supply constraints. Both names came off their highs this week as rising 10-year Treasury yields pulled the broader market down. (Read More)
🫖 TEA For Thought: "Right now the bigger of these two is heavily hardware-focused, and it might shift into agent-focused, application-focused. I would say this is just the beginning, but AI moves too fast, so the beginning might just be the end also, or before you realize it, it's already the end."
2. 🚗 An AI Model Drove A Real Car, At 0.94 MPH
Researchers Tobias Gessler, Aditya Ramabadran, and Simon Mahns ran commercial AI models through a parking-lot cone course as part of DrivingBench, wiring a Toyota Corolla to a $999 comma four device running openpilot. GPT-5.6 Sol, Grok 4.6, and Claude Fable 5.1 all failed to finish. OpenAI's GPT-6 Astra completed the 134.7-meter course on its second attempt in 5 minutes 22 seconds, averaging 0.94 miles per hour and burning 6.6 million tokens for $7.74, which works out to about $92.47 per mile against roughly $0.18 per mile for gasoline. Models kept refusing to drive on safety grounds until the team renamed its tool server "DrivingBench Sandbox." Ramabadran said frontier models are not practical for real driving today, and expects specialized in-car models to keep winning on speed and cost. (Read More)
🫖 TEA For Thought: "This should be one of the samples for AGI being achieved, because they could just learn by taking courses and start driving. Full FSD is probably there."
3. 🧸 Meta's Muse Charm Is Betting On The Bag-Charm Economy
Meta announced the Muse Charm on September 23, a Tamagotchi-style dangling wearable for its Muse AI agent, and TechCrunch argues the form factor is aimed at an existing Gen Z habit rather than at AI hardware buyers. The bag-charm market is real money: Pop Mart's Labubu helped push revenue to $1.8 billion in 2024, a sevenfold jump from 2019, Hailey Bieber's Rhode hit $212 million in annual net sales from roughly 10 products before its billion-dollar sale to Elf Beauty, and one estimate puts global bag charms above $1 billion by 2030. TikTok has already turned old Apple Watches into keychains and charms. The open question is trust, given Meta's $5 billion FTC settlement, its $1.8 billion August settlement with 29 states, and a $942 million New Mexico ruling. Meta says Muse will make money by taking a small fee from transactions. (Read More)
🫖 TEA For Thought: "A great read for understanding the market position of the Muse Charm. While Meta may have lost customers' trust and may never fully recover, their product taste is still something worth learning from."
4. 🏦 MoonPay Buys Its Way Into SEC-Registered Rails
MoonPay has agreed to acquire North Capital in an all-stock deal that sources value at more than $60 million, announced September 23. North Capital, based in Midvale, Utah, has processed over $8.7 billion in primary and secondary transaction volume and brings a set of SEC-registered affiliates including a broker-dealer, an alternative trading system, a transfer agent, and an investment adviser. Both boards approved the deal, which still needs regulatory clearance and standard closing conditions. CEO Ivan Soto-Wright framed it as "building the regulatory foundation to support mass adoption of tokenized real-world assets." It follows MoonPay's 2026 purchases of Sodot for institutional key management, DFlow for on-chain trading infrastructure, and Entendre for AI finance operations. MoonPay is currently valued at $3.4 billion and last raised outside capital in October 2021. (Read More)
🫖 TEA For Thought: "This is just a sign that MoonPay is moving into RWAs, like everyone else is."
5. 🛰️ China Put An AI Data Center Into Orbit
CAS Space launched nine satellites on September 20 aboard its Kinetica 1 Y18 rocket, including Supercomputing-1, China's first integrated rocket-and-satellite AI computing project. Built by S-AIDC (Chaozhisuan Technology), the satellite carries a high-resolution optical payload plus an onboard image-processing AI computer, so Earth-observation data gets analyzed in orbit instead of being shipped down to ground data centers. The satellite decides which images are worth downlinking and uses laser intersatellite links to move data between craft. CAS Space and Chinese state media put the improvement at hours down to minutes for cross-regional processing. The launch was one of six Chinese orbital launches in six days, and follows the Space Computing Industry Innovation Center that Beijing set up in June to link rocket makers, satellite builders, chip fabs, and AI companies. (Read More)
🫖 TEA For Thought: "Space is a territory everyone is fighting for."
🛠️ Skill of the Day
The Ambiguity Sweep: catches the sentences your reader will misread, before you hit send.
You are a precision editor. I will paste something I am about to send: an email,
a project brief, a task ticket, or a prompt for an AI tool. Do not rewrite it yet.
Step 1. List every sentence or phrase that could reasonably be read more than one
way. For each one, quote the line, then give Reading A and Reading B in plain
language.
Step 2. For each ambiguity, say which reading the recipient is most likely to
pick, and why. Flag the ones where the wrong reading would cost real time or money.
Step 3. Find what is missing. List anything a reader would need in order to act
but cannot find in the text: dates, owners, budgets, file formats, what "done"
means.
Step 4. Only now, give me one revised version. Keep my voice and my length. Change
only what Steps 1 to 3 flagged, and show the edits as a short before-and-after list.
Here is the text: [PASTE YOUR TEXT HERE]Paste into ChatGPT, Claude, or your tool of choice. It works on anything you are about to send, from a one-line Slack message to a full spec.
TEAHEE Moment
Stay sharp, stay informed. See you Monday.
If you enjoyed this TEA, follow along on social for more: