Nebius stock collapsed 40% during the DeepSeek panic while at least one AI infrastructure firm was having its best sales week

When DeepSeek’s AI model release landed in late January 2025, it set off a wave of selling across publicly traded AI stocks. Nebius Group, the Amsterdam-listed AI infrastructure company trading under the ticker NBIS, took a particularly sharp blow. According to the Globe and Mail and related market commentary, the stock fell nearly 40% in the period immediately following the announcement, part of a broader fear-driven reassessment of whether large AI infrastructure buildouts remained economically justified.

Roman Chernin puts a human frame on that number. He recalls Nebius stock dropping roughly 40% in the span of about a week, though he acknowledged uncertainty about the exact timing, placing it somewhere in February or March, and wavering between 2024 and 2025 before settling on early 2025. That hedged recollection is worth noting: the date he is reaching for is real, but his memory of it is approximate. What the external record confirms is the scale and the period, not the specific week Chernin had in mind.

What the external record cannot confirm, and what gives Chernin’s account its editorial weight, is the second half of his observation. That same week, he says, his company had its best week in sales. The two data points sit in the same sentence: a public-market panic and a private-market surge, running simultaneously in the same infrastructure category.

Nebul stock went down 40% in one week or so uh uh in February I think it was February or March 2024 2025 and anecdotal story the same the same exact week we probably had the best week in sales Roman Chernin

The divergence is not hard to explain in retrospect. DeepSeek’s release triggered fears that cheaper, more efficient AI models would reduce demand for compute-intensive infrastructure. Public markets moved on that fear. Operators buying AI infrastructure capacity, by contrast, may have read efficiency gains as a reason to build faster, not slower. If models cost less to run, the economics of deployment improve, and the business case for acquiring more compute gets easier to make, not harder.

That interpretation is inferential. Chernin does not spell out the mechanism; he reports the coincidence. But the coincidence is pointed enough that it deserves to be reported plainly. A week that looked, from a shareholder’s perspective, like confirmation that the AI infrastructure trade was overextended looked, from a sales perspective, like the opposite signal.

The broader pattern the episode illustrates is the gap between public-market sentiment and private-market demand in fast-moving technology categories. Equity prices in AI infrastructure have been driven heavily by narrative: the narrative that large model training requires enormous, sustained capital investment. DeepSeek complicated that narrative in a single news cycle. But infrastructure demand is not driven by narrative alone. It is driven by operators who are pricing actual workloads and making actual purchasing decisions, often on timelines that have nothing to do with whatever story is running on financial media in a given week.

Chernin’s account is a single data point from a single company in a single week. It does not establish a general law. What it does establish is that the DeepSeek moment, read as an unambiguous negative for AI infrastructure, was not universally negative for AI infrastructure businesses. The stock chart and the sales ledger were telling different stories at the same time. Which one was more predictive of where demand would go from there is a question the evidence here does not answer. But the question is worth asking, and Chernin’s anecdote is a reasonable place to start asking it.