Know Thyself

Knowing What You Are

Two winners looked at the hottest game in AI this week and, in completely different ways, decided to stop playing it. Both were right, and both were doing the same thing: acting on a clear-eyed read of what they actually are.

Jeff Dean left Google. Twenty-seven years, the man behind MapReduce, TensorFlow, and the TPU, gone with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals to start Discovery Loop. The pitch deck did 3.1 million views and every marquee fund tripped over itself to lead the seed. Dean doesn't need the money. Nobody walks away from a Google comp package chasing a raise. He's leaving because a founder who built the machine can't build inside the bureaucracy that grew up on top of it. This is the oldest move in the Valley: the people who make the thing walk out of the empire it became, to go build again, unshackled. Noyce and Moore did it to Fairchild in 1957 and got Intel, the same walkout that later seeded Sequoia and Kleiner Perkins, one of which is writing a check in Dean's round today. And Dean isn't alone. Google has spent years shedding the exact people who built it, a Nobel laureate and its earliest engineers among them. When your best builders keep choosing risk and a blank whiteboard over your stock and your cafeteria, that's not a retention problem. It's a verdict on the building.

Sequoia made the opposite move for the identical reason. It put $10 billion into "AI and reindustrialization," the biggest bet in its 54-year history, and pointedly not into another frontier model. Read it as Moneyball. When every fund on Sand Hill is bidding the same model rounds up to fifty-times oversubscribed, the last thing a disciplined investor does is pay top dollar to stand in that line. Billy Beane didn't outspend the Yankees. He found the value the market was mispricing. Sequoia's doing the same, stepping one ring out into robots, defense, power, and mines, the overlooked atoms nobody's fighting over, where a check still names its own terms and takes the board seat. Path Robotics just grabbed $600M of a $900M Navy deal. Unitree's going public off 5,500 robots. Let everyone else pay up to make the models better. Every gain lifts the value of the boring stuff underneath.

Two winners, one instinct: know what you are, and don't keep playing a game that's stopped rewarding it. Dean's edge is building, so he left to build. Sequoia's edge is finding value, so it left to find it. The people still crowded in the middle, paying any price for a sliver of the model layer, are playing everyone else's game at everyone else's prices.

At COAI today: the full Signal/Noise, with the Delphic frame, the 1957 Fairchild walkout that seeded both Intel and Sequoia, and why the "safe" floor Sequoia's buying is literally cracking, is live at getcoai.com.

$10 billion. The biggest check Sequoia has written in 54 years, and it skipped the models entirely, going into robots, defense, and power instead. That's not caution, it's Moneyball: when everyone's overpaying for the same frontier bet, the edge is the value they're ignoring. When the sharpest fund in venture stops bidding up the crowded game and goes hunting the overlooked one, that's the tell about where this cycle's returns actually live.

Figure out which game you're actually playing. Go bet by bet: are you paying up to crowd into the same frontier-model trade as everyone else, or backing the overlooked stuff underneath it? Sequoia just wrote its biggest check in 54 years choosing the second. If you're all in on the crowded game, you're a price-taker in the most oversubscribed market alive.

Price the mission before your best people do. Jeff Dean walked from Google because the money was long solved and only the work was left to move him. Once you've paid a key person market, your retention plan is entirely mission and autonomy. Write down today what each one is building and why it matters, before someone offering a blank whiteboard and real ownership asks them first.

Stress-test the floor you're standing on. The "safe" physical assets backstopping AI bets are cracking: data centers frying their own turbines, running near 80% uptime against a 99.999% promise. Pick the hard asset your AI plan quietly assumes is reliable forever, and pressure-test it this afternoon with someone who actually runs one.

Friday, August 7

— Harry and Anthony

Sources: