Why the bot decided this
Paul Kedrosky warned that cheap AI agents could spread rapidly while expensive data centers and frontier models become overbuilt, challenging the AI infrastructure boom. Source: Benzinga.
The article
EXCLUSIVE: Paul Kedrosky Warns AI Data Center Boom Could Fail Even as AI Adoption Soars Artificial intelligence could become a massive commercial success while leaving investors with billions tied up in unnecessary data centers, technology investor Paul Kedrosky told Benzinga in an exclusive interview. He pointed to a house painter he knows who uses Muse, an AI agent from Meta (NASDAQ:META), to handle emails and customer inquiries instead of hiring an office assistant. Kedrosky uses Muse himself to manage a schedule he says he’s bad at keeping track of. “It’s filled that gap and it’s done it really, really well,” he said. He believes the painter represents AI’s next wave of customers: ordinary businesses that want cheap, useful tools rather than expensive models requiring enormous computing power. AI’s Early Customers May Be Misleading Investors “One of the biggest mistakes you can make in venture capital is looking at your first three or four customers and assuming they’re representative of what the mass market will look like,” Kedrosky said. AI’s earliest customers included software developers, who consume enormous numbers of tokens, the units AI models process when generating responses. “Early adopters are weirdos, right?” he said. Kedrosky argues that ordinary customers have very different requirements. A painter answering emails doesn’t need the same computing power as a programmer generating thousands of lines of code. “They’re not looking for frontier models. They want something cheap and fast that runs at the edge,” he said. If millions of businesses automate everyday tasks using smaller, cheaper models, AI adoption could soar while demand for expensive computing capacity falls short of expectations. That could also weigh on chip suppliers like Nvidia Corp. (NASDAQ:NVDA). Meta’s Strategy Could Undercut OpenAI and Anthropic Kedrosky believes the shift threatens neoclouds such as CoreWeave Inc. (NASDAQ:CRWV), which rent out computing capacity, as well as OpenAI and Anthropic, which rely on customers paying for powerful models. He argues Meta could instead offer cheap intelligence and earn fees when agents complete transactions. “Muse is closer to what the future looks like than is this kind of monolithic premium-driven pricing umbrella associated with the frontier companies,” he said. Meta Bets on Cheap AI While Spending Heavily Mark Zuckerberg is pursuing a different business model from the frontier labs. Rather than relying on selling access to powerful AI models, Meta aims to make agents widely available and generate revenue from shopping, transactions and other services, as outlined in its plans for Muse. More than 10 million people have signed up for Meta’s agent since its September launch, with approximately 25% using it daily, Zuckerberg said Wednesday. Meta is still spending heavily, forecasting $135 billion to $145 billion in capital expenditures this year. But its existing advertising business provides financial support that standalone AI developers lack, and Zuckerberg expects to cut Muse’s computing costs more than tenfold over the next year. Traders see limited near-term risk of an industrywide collapse. On Polymarket, they assigned roughly 9% odds Friday to an AI downturn by Dec. 31, 2026, rising to 21% by June 30, 2027. The contracts require at least three specified distress events within 90 days, rather than simply disappointing data-center returns. Kedrosky believes AI’s usefulness won’t prevent widespread overbuilding. “We are, like in every prior capex moment, vastly overbuilding because of the uncertainty of what end demand looked like,” he said. Image: Shutterstock Read Also: 'Intel Plus Elon' Could Pose Bigger Threat to TSMC Than Musk Alone, Chip Analyst Says