The Cold Start Problem Just Became a $150M Acquisition

Nebius just acquired Inferize, a nine-month-old startup, specifically to fix cold starts. Terms weren't officially disclosed, but Calcalist, Israel's leading financial paper, puts the deal at $100-150 million — a striking price for a 17-person company that had never even publicly disclosed its $10 million seed round before the acquisition was announced.

The problem that justified the price

The problem Inferize solved is simple to state and expensive to ignore: every time a GPU instance spins up to serve a request, the model has to load into memory before it can do anything. That gap — tens of seconds, sometimes minutes — either makes users wait or forces platforms to keep GPUs idling “warm” just in case, paying for capacity that's doing nothing. Inferize's own pitch was blunt: keeping spare GPUs running is the price of being ready for demand, and that price is a tax on every inference platform's margins.

Not a one-company problem

This isn't an isolated bet. Modal has spent much of this year building GPU memory snapshotting into its platform — saving full CPU and GPU state after a model loads, so later cold starts restore from that snapshot instead of reinitializing from scratch, cutting cold starts roughly tenfold in their own published benchmarks. InferX, for its part, says it's running sub-second cold starts in production using a disk-based persistent KV cache that avoids the reload entirely rather than snapshotting around it. Three different companies, three different technical bets, converging on the same conclusion: idle GPU time is the single biggest lever left in inference unit economics.

What the price actually says

The signal from this deal isn't “cold starts are annoying.” It's that solving them well is now worth $100-150 million, less than a year after the underlying company was founded. When an infrastructure problem moves that fast from open question to M&A target, it tells you where the money actually is in this market — not in bigger models, in the seconds between a request arriving and a GPU being ready for it.