Xiaomi's AI Comeback Reveals a Harsh Truth About the Talent War

Shortly after Xiaomi CEO Lei Jun recruited Fuli Luo, a core researcher from DeepSeek, the company released MiMo-V2-Flash, a 309-billion-parameter model that put them at number two in open-source AI rankings. One hire. A few months.

Meta ran the same playbook when it acquired Scale AI and aggressively brought over OpenAI talent. The pattern is consistent enough to be instructive.

What the Moat Actually Is

Compute infrastructure is now table stakes. If you can buy H100 chips, you can start. The hardware barrier has largely collapsed, and data regulation in permissive environments like China keeps that door open for now. What changes the trajectory of an organization is a person who has already trained a frontier model before.

That framing is uncomfortable because it means competitive advantage is concentrated in a small number of individuals, any of whom can be recruited away. Anthropic and Google remain formidable partly because their retention rates are high enough to protect that concentration. Companies that lose their best researchers watch competitors leapfrog them in months, not years. There’s no structural fix for this; it’s just how thin the margin is at the frontier.

What This Means for AI Startups

The pitch deck framing of “proprietary offline data” or “specialized professional datasets” rarely matches what actually drives outcomes. From what I’ve seen across the different business models:

For model developers, having someone who has trained a frontier model before matters more than data or compute alone. Positioning in strategic markets, like Korea’s sovereign AI push, adds another layer of defensibility, but that’s geography as much as technology.

For B2B services, the founder’s personal network and ability to land early customers is what determines survival. Companies like Legora illustrate this well. The model underneath is almost secondary.

For B2C services, distribution and referral mechanics drive the user metrics that attract follow-on funding. The AI capability is often interchangeable with a competitor’s.

None of these paths offer a durable moat in the traditional sense, which is worth being honest about before raising on the premise of one.

The Uncomfortable Part

Even model developers, the companies with the deepest technical capabilities, are locked in a talent war over the same handful of people. When the organizations closest to the frontier are fighting over a pool that small, it tells you that architecture, datasets, and chip supply are not where competitive advantage lives. They are necessary but not sufficient. The edge is in attracting and keeping the people who know how to push the boundary, and that edge is fragile by definition.