🐶 Dogfooding CEOs 🥣

With Tesla’s Terafab project, the company is stepping into uncharted territory. Not only is it planning what could be the largest semiconductor fabrication plant in the world, it’s doing so with virtually no experience in chip manufacturing. Its ambitions are literally out of this world: Tesla has described it as a step toward “becoming a galactic civilization.”

• Terafab aims to bring all aspects of chip production — from design to fabrication to packaging — under one roof. CEO Elon Musk said the facility is intended to produce up to 1 terawatt of compute annually. 

• The goal is to supply Tesla, SpaceX, and xAI with the ballooning number of AI chips they expect to need, since Musk has said existing suppliers, including TSMC, Samsung, and Micron, can’t handle their future demand. 

• On Tesla’s last earnings call, Musk said chip supply would be a “limiting factor” for Tesla’s growth in about three or four years.

• The plant would manufacture inference chips for Tesla’s Robotaxis and Optimus robots, as well as custom AI chips for space-based applications, including solar-powered AI satellites. Morgan Stanley estimates the project could cost $35 billion to $45 billion in capital expenditure, likely shared between Tesla and SpaceX.

Early reports suggested Terafab could be located on the north side of Tesla’s Giga Texas facility, based on a slide Musk presented showing an “Advanced Technology Fab” there. But Musk later clarified that this referred only to a smaller facility for iterating on chip designs. The Terafab itself, he said, would be “far bigger than everything else combined” on the Giga Texas campus and require “thousands of acres.”

The Takeaway

Musk has not provided a timeline for the project, but even at his typical breakneck pace, it is likely years away. Building a semiconductor fab is one of the most complex industrial undertakings in the world, typically carried out by established players like TSMC or Intel, and often takes years to complete and ramp.

“Even under an aggressive scenario involving a retrofit of an existing facility, initial chip output would likely not occur until mid-2028 at the earliest,” Morgan Stanley analysts wrote last week. “A greenfield build would extend timelines further, with meaningful output more likely 4-5 years after project initiation.”

Go deeper

Mark Zuckerberg wants you to know he’s building an AI agent to help him be CEO — and that eventually everyone should have one. Jensen Huang is broadcasting that he’d be “deeply alarmed” if Nvidia’s $500,000 engineers weren’t burning through $250,000 in AI tokens a year. Salesforce keeps talking about “digital labor” like it’s already a line item in your budget.

• You can take all of this at face value. Or you can recognize a familiar move: the people selling the future are making a point of telling you they’re living in it first. It’s a little like Hair Club for Men. They’re not just pitching the product — they’re the testimonial.

• Tech has a long history of “dogfooding,” or using its own products internally to make them better. But this feels different: AI is still poorly understood by most of the people being asked to buy it, and at the same time they’re being told it’s inevitable.

• Across Big Tech, CEOs are starting to define what “good” looks like in an AI world. At Meta, that means flattening teams and pushing employees to use internal AI tools so aggressively that it shows up in performance reviews. At Nvidia, it means tying productivity to token consumption; if you’re not spending enough on AI, something’s wrong. 

This works because companies don’t just buy software — they copy norms. If the CEO of Nvidia says serious engineers should be using massive amounts of compute, that doesn’t stay contained to Nvidia. It seeps into how other companies evaluate their own teams.

The Takeaway

Overall business spending on AI has been growing, and the size of those contracts has been growing as well, according to data from Ramp, a corporate card and expense management platform, suggesting that companies are finding them useful.

So far, the external data showing AI productivity is limited. While about 40% of adults use AI at work, the time saved amounts to only about 2% of total work hours. A survey of 1,000 hiring managers by Resume.org found that AI’s impact on jobs has been minimal so far, with 9% saying it had fully replaced certain roles and 45% saying it had little to no impact on staffing. Researchers found that AI boosted productivity among customer service workers by about 15% on average — though gains were uneven and concentrated among less experienced employees.

In the absence of robust proof, marketing fills the gap.

Read more

Very big things are happening in the world of nutritionally complete products that taste like chalk: Danone has agreed to buy the celebrity-backed protein bar, powder, meal, and meal replacement shake maker Huel, a portmanteau of “human” and “fuel,” which counts Idris Elba as a fan. 

Why Danone wanted to bulk up with Huel

• 🏀 NCAAW: As the women wrap up the second round of March Madness, the UConn Huskies remain the overwhelmingly large favorite after dispatching Syracuse last night. While there are indeed 16 teams left, UConn has a 66% chance of winning the tournament. Indeed, the top 3 teams (UConn, UCLA, and the University of Texas) have a combined 93% chance of winning, leaving the remaining 13 teams to divvy up the remaining 7%. 

• 🏀 NCAAM: After a spellbinding weekend of basketball, the three big favorites — Arizona (22% chance of winning it all), Michigan (21%), and Duke (19%) — remain on top, though you may have noticed a change in those odds. Duke was once the top contender in the tournament, but has slipped from about 25% odds pretournament after barely squeaking by Siena in the round of 64. 

*Event contracts are offered through Robinhood Derivatives, LLC — probabilities referenced or sourced from KalshiEx LLC or ForecastEx LLC.

Earnings expected from GameStop