Nebius buys Inferize, which says it restores a warm GPU in 20 seconds, not 23 minutes
The 17-person Tel Aviv startup claims GPU snapshots cut a DeepSeek V4 Pro cold start roughly 70-fold, and reports value the deal at $100 million to $150 million.
Nebius announced on 1 October that it has acquired Inferize, an inference-optimisation startup, to strengthen its Token Factory platform. Terms were not disclosed. CTech estimated the price at $100 million to $150 million and Globes at $100 million to $130 million, runtimewire reported.
Inferize captures a warmed GPU and CPU state, including compiled kernels and CUDA graphs, and restores it on fresh GPUs within seconds, runtimewire reported. The idea is that operators no longer have to keep spare GPUs idling in case traffic spikes. Co-founder Guy Bortnikov said spare capacity is the price of being ready for demand.
The headline number is Inferize's own. Its published benchmark for DeepSeek V4 Pro shows a snapshot restoring in 20 seconds against about 23 minutes for a cold start, a roughly 70-fold gain, runtimewire said. Inferize also claims GPU-cost savings of 10 to 30 percent. Neither figure has been independently tested.
Inferize was founded in January 2026, built a working prototype in three months and has 17 staff in Tel Aviv. Its co-founders Guy Bortnikov and Lior Gorbonos previously worked at Granulate, which Intel bought in 2022. They join Nebius's Token Factory alongside earlier acquisitions Eigen AI and Clarifai.
Nebius shares rose 1.9 percent in premarket trading to $240.52 after the announcement, according to GuruFocus, which tracks the stock.