Nebius buys idle-GPU startup Inferize for an estimated $100 million to $150 million

Spread across Inferize's 17 employees, Calcalist's estimate of the deal comes to $5.9 million to $8.8 million a head ^[1]. Spread across the roughly 10 months the company existed, it is $10 million to $15 million a month ^[2]. Both figures depend on Calcalist's range ^[2].

Guy Bortnikov, Inferize's chief executive, described the problem as a cost. "Keeping spare GPUs running is the price of being ready for demand. Removing that cost is what we built Inferize to do, and Nebius is where it can go straight into the platform," he said ^[9]. Calcalist describes the choice large inference operators face: pay for spare GPUs that may sit unused for long periods, or run close to real demand and risk reacting too slowly when workloads spike ^[15]. Nebius's release calls the cost of the first option an "idle GPU tax" ^[8].

Nebius disclosed neither the price ^[17] nor a utilization figure for Token Factory ^[8]. The value of the deal turns on three numbers outside the record: how many GPUs serve Token Factory, what share of them is held in reserve, and what a reserved GPU-hour costs Nebius. Without them, the $100 million to $150 million cannot be checked against a saving ^[2].

Two readings fit the facts, and a third sits beside them. In the first, Nebius bought a team. Bortnikov and co-founder Lior Gorbonos were on the founding team of Granulate, the infrastructure-optimization company Intel bought in 2022 ^[5]. In the second, it bought a product whose worth is set by how much reserve capacity Token Factory can stop holding ^[7]. The third is location. Nebius has signed for data center capacity in Israel and was chosen to build a national AI supercomputer there ^[14], and Inferize's staff work in Tel Aviv ^[3].

I think the team reading accounts for most of the price. Danila Shtan, Nebius's chief technology officer, said: "The team's contribution will extend well beyond this first integration" ^[10]. The deal also comes in at 36% to 55% of the initial $275 million Nebius agreed for Tavily ^[3], and Inferize had been operating in stealth ^[3]. If Nebius later reports a Token Factory utilization gain large enough to cover $150 million of avoided GPU spend, the product reading wins and the per-head figure undercounts what it bought.

Calcalist places the deal in a wider move: "As companies move from training models to deploying them for customers, the cost and speed of inference are becoming increasingly important" ^[16]. A single deal by one buyer cannot show that AI clouds now compete on filling GPUs instead of buying them. Nebius is doing both, or rather, it is planning facilities in Modi'in, Masmiyya and Beit Shemesh in the same country where it is buying the software ^[14]. The one model company it negotiated with this year became a commercial partner. Talks with AI21 ended without a transaction, and AI21 later cut more than 60% of its workforce ^[13]. Counting Tavily at its $400 million ceiling, Nebius's two Israeli software purchases since February come to as much as $550 million ^[4].