TLDR: This week on my Moonshots Podcast we did a deep dive into Google I/O 2026 (the most ambitious developer conference in the company's history) plus Andrej Karpathy's move to Anthropic, Andrew Feldman's record-breaking Cerebras IPO, and the new Build with Gemini XPRIZE. If you haven't had a chance to listen to this week's Moonshots episode or would like a reminder of the most important points, let's dive in…
GOOGLE I/O 2026: THE FULL-STACK AI PLAY
Google’s Staggering Numbers and Why 6X CapEx Is Bullish
Sundar Pichai opened Google I/O with numbers that should make your head spin: quadrillions of operations, hundreds of billions in infrastructure spend, and an array of products with over a billion users each. Google has 6X’d its CapEx… and the stock went up. Five years ago, nobody would have believed that was even possible.
Gemini at Scale: The Gemini app now has 900 million monthly users, approaching ChatGPT territory. We now have a three-horse race between OpenAI, Anthropic, and Google, and that’s healthy for the ecosystem.
The Singularity in Real Time: As AWG puts it: “Every sci-fi trope, everywhere, all at once over the next 10 years… that’s the singularity.”
Gemini Omni: Google’s Multimodality Bet
Google launched Gemini Omni, a new family of models generating video from any combination of text, photos, video, and audio. Google DeepMind is now the only remaining American frontier lab seriously pursuing multimodality. OpenAI cut Sora. Anthropic focuses squarely on cogen.
The Deeper Bet: AWG suspects Google is treating biological sequences (DNA, protein) as just another modality, betting on “modality scaling” as the path to superintelligence that other labs aren’t chasing.
Google’s AI Strategy: Fast, Cheap, and Everywhere
Google launched Gemini 3.5 Flash as the new default for the Gemini app. Alex Wissner-Gross’ honest assessment: “Solidly mid.” Not frontier, but optimized for throughput and tool use. That maps directly onto Google’s needs powering search at scale.
The Two-Horse Race: Dave’s Silicon Valley scuttlebutt: “It’s a two-horse race between OpenAI and Anthropic for the best AI in the world.” The other labs have pivoted to enterprise.
The Bifurcation: Salim identified the strategic split: premium cognition vs. ultra-cheap but fast cognition. Flash serves the latter. And that continuous march of marginal intelligence, cost trending toward zero, has massive implications.
The Distribution Play: Gemini Spark is Google’s 24/7 AI agent writing emails, creating study guides, monitoring financial fees. Dave raised the point that matters: “Over half the world uses Google. One click, fully integrated. They don’t need to push boundaries… they just need to be good enough.”
“Code is clearly going away as a human endeavor. It’s all being abstracted away by code agents.” — AWG
SynthID: The Proof-of-Reality Infrastructure
Google’s SynthID content credentialing standard was adopted by OpenAI the same day it was announced. This is the beginning of end-to-end authentication of realness.
The Irony: We’re going to get proof of reality not from the camera end, but from the synthetic end. All vendors want to claim credit that they generated the reality. Camera makers will be downstream adopters.
Self-Regulation: Dave: “AI can’t be kept up with by Congress. The industry is starting to self-regulate. We may look back on this as one of the first moves by the self-regulation community.”
The Formula: Salim’s elegant equation from Jerry Mikulski: Scarcity = Abundance minus Trust. Solve for trust, you solve for abundance. “We’re moving from the information age to the verification age.”
Universal Cart: The End of Shopping
Google launched a universal shopping cart across YouTube, Search, Gemini, and Gmail, working with Nike, Target, Walmart, and Shopify.
Agent Commerce: Salim saw the real disruption: “Today: human → website → cart → checkout. Tomorrow: intent → agent → transaction.” Every CMO will ask: how do I convince a hundred million agents to choose my product? Shopping becomes continuous, invisible, and returnable.
THE BUILD WITH GEMINI XPRIZE
$2 Million to Build, Market, and Scale: In 90 Days
My team and I at XPRIZE just announced the “$2M Build with Gemini XPRIZE,” funded by Google at $3 million for the prize purse plus $1 million for operations. Here’s the challenge: find a problem impacting 100,000+ people, build your product using AI, and scale revenue as much as possible in 90 days. Describe what you want in English, and the agent builds it. Dave Blundin called it “the highest calling for XPRIZE yet.”
Go to www.GeminiXPRIZE.com to learn more and register.
The five finalists for both the Gemini XPRIZE and the Future Vision XPRIZE will pitch at the Moonshot Gathering on September 25th in Los Angeles. Go to www.moonshots.com to register for the Gathering.
AI TALENT WARS
Andrej Karpathy Joins Anthropic
Andrej Karpathy, co-founder of OpenAI and former head of Tesla’s full self-driving, has joined Anthropic’s pre-training team. He’ll lead a new initiative using Claude to accelerate Claude’s own pre-training research.
Dave’s Analysis: “Every other OpenAI co-founder has either raised billions or is at a foundation lab. He’s the only guy who didn’t have access to the big machine. You can’t miss the singularity by being outside of the big game.“
Andrew Feldman’s Take: Karpathy’s point applies to hardware too: “If you’re not building hardware for one of the three most important labs (Google, Anthropic, OpenAI) your hardware will drift from what they need.“
CHIPS & DATA CENTERS
Cerebras: Record IPO and the Wafer-Scale Revolution
Andrew Feldman joined us on the Pod from Cerebras HQ, fresh off the biggest US tech IPO since Uber: up 68%, market cap of $95 billion. Dave visited the offices: “It’s like a Nobel Prize or Olympic gold medal. You carry it for the rest of your life.”
The Contrarian Bets: In 2015, Cerebras made two bets: AI would need dedicated silicon (not GPU derivatives), and you needed a clean sheet of paper. Both proved dead right. The result: a chip the size of a dinner plate, 58x larger than any chip ever built.
The Dark Days: “We spent 18 months burning $8 million a month and couldn’t solve the problem. Board meetings every six weeks, $100 million more in the hole, then $120 million, then $140 million.” They solved wafer-scale in August 2019. The world didn’t care until late 2024, when models got smart enough that everyone wanted fast inference.
NVIDIA’s Sleight of Hand: “NVIDIA doesn’t always tell you whether they mean tokens per second per user or aggregate throughput. A GPU generates 35 tokens/sec… painfully slow. Ask it to do 200 tokens/sec per user and it supports one or two users. That’s a $4 million solution working on one user.”
On TeraFab: Andrew puts Elon’s TeraFab at a 15-20 year project, not 5-10. “Fabs are pyramids. Even with the exact same ASML equipment, Samsung and TSMC aren’t at the same node. The received wisdom from generations of fab building cannot be underestimated. But if anybody can do it, Elon can.”
“This shit will kill you if you can’t modulate the highs and the lows. The number of times you get kicked in the gut before lunchtime and have it still be a good day as a startup CEO is amazing.” — Andrew Feldman
HERE’S THE BOTTOM LINE...
Google just showed the world what a full-stack AI company looks like when it fires on all cylinders: from custom TPUs to consumer apps with a billion users, from frontier science to universal shopping carts. The model race still belongs to OpenAI and Anthropic, but don’t underestimate distribution when half the world already lives in the Google ecosystem.
Meanwhile, the hardware story is quietly becoming the main story. Cerebras proved that a clean-sheet approach to silicon can produce a $95 billion company, and Andrew Feldman’s warning (fabs are our pyramids) should sober anyone who thinks we can reshore semiconductors on a political timeline. The punchline from Karpathy’s move: if you’re not at the frontier, you’re not seeing what’s coming next.
Catch the full episode wherever you get your podcasts, and join us at the Moonshots Gathering in Los Angeles on September 25th. Go to www.moonshots.com to register. And don’t forget to check out www.GeminiXPRIZE.com for our $2M “Build with Gemini” hackathon!
See you next week,
Peter
More From Peter
If you’ve enjoyed Metatrends, here are more ways to stay connected:
I appreciate the recap Peter👍
Where do I go to cancel?