Nebius Buys Israeli Startup Inferize for Up to $150 Million to Squeeze Idle Time Out of Its GPU Fleet

Nebius Group, the Nasdaq-listed AI infrastructure company, has acquired Inferize, an Israeli startup that attacks one of the least glamorous but most expensive problems in AI inference: GPUs sitting idle while models load. The deal was not disclosed in financial terms, but people familiar with the transaction put it at $100 million to $150 million, according to the Israeli technology outlet Calcalist.

Inferize, founded in early 2026 and operating in stealth for roughly ten months, built technology aimed at the cold-start problem. When an AI model receives a request, its weights must be loaded into memory before the first token can be served - and during that initialization window, expensive GPUs produce nothing. The same waste appears when weights refresh during reinforcement learning cycles or when traffic spikes force operators to spin up new instances faster than models can load. Inferize's approach lets inference capacity scale dynamically with actual usage, so operators no longer have to keep a buffer of idle GPUs on standby just to be ready for demand.

The startup is small - 17 employees at its Tel Aviv office - but its founders are repeat specialists. Guy Bortnikov and Lior Gorbonos previously worked together at Granulate, an infrastructure optimization company Intel acquired in 2022, and delivered a working prototype within three months of founding Inferize. TLV Partners led the seed round. "Keeping spare GPUs running is the price of being ready for demand. Removing that cost is what we built Inferize to do," Bortnikov said in the announcement.

Nebius CTO Danila Shtan framed the acquisition around the same economics, noting that serving AI inference well requires the whole system - not just faster GPUs and optimized models - to respond quickly when customer demand shifts. Inferize's engineers will join Nebius's inference team, and the technology will be integrated into Token Factory, the company's managed inference platform, where it sits alongside technology from Eigen AI and the core team and licensed technology of Clarifai.

The purchase is Nebius's second Israeli acquisition in recent months. In February it agreed to buy Tavily, which builds a search layer for AI agents, for $275 million upfront with total compensation potentially reaching $400 million on performance milestones. Nebius also held early-stage acquisition talks with AI21 this year; those ended without a deal and became a commercial partnership instead, while AI21 subsequently cut more than 60% of its workforce to focus on its Maestro agent-optimization platform.

Beyond M&A, Nebius has been deepening its Israeli footprint with data center capacity agreements in Modi'in, Masmiyya and Beit Shemesh, and it was selected to establish the country's national AI supercomputer. The company reported $582.3 million of revenue in its most recent quarter and has been racing to pair its rapidly expanding GPU fleet with a software layer that raises utilization.

The deal reflects where the AI infrastructure market's attention has moved: from training clusters to serving models economically. As inference becomes the dominant workload, the cost of a token is decided less by chip price than by how many of the chips a operator actually keeps busy - and small teams that can move that number are suddenly worth nine figures.