Deci AI

AutoNAC: search for a new architecture instead of shrinking an old one

Deci AI was founded in 2019 by Yonatan Geifman, Ran El-Yaniv and Jonathan Elial. The usual way to cut the cost of running a model is to shrink an existing architecture, for example with quantization. Deci came at it from a different direction: an automated engine called AutoNAC, based on the research field of neural architecture search, looked for an alternative architecture tailored in advance to the hardware the model is meant to run on. Until then this field had been considered too expensive for commercial use.

YOLO-NAS and a model's full life cycle

Its products included the DeciNets family of models and the object detection model YOLO-NAS, released in 2023, which became widely used in the computer vision community. The platform covered the whole life cycle, building, optimizing and deploying to the cloud, the edge and mobile, and not only the optimization stage itself.

The acquisition by Nvidia, April 2024

In April 2024 Calcalist reported that Nvidia was acquiring Deci for about $300 million. This amount and date rest on that press report alone, since no detailed official acquisition announcement from Nvidia was found: not in its newsroom, not on its technical blog, and not in its 8-K filings with the SEC around the time of the acquisition. This is probably because a deal of this size is too small to require a separate filing or a standalone public announcement. What was actually verified: the company's old website address, deci.ai, has redirected (a 301 redirect) straight to nvidia.com since at least August 2024, both in a live check and through the Wayback Machine archive. This is indirect technical evidence of current actual ownership of the domain, not a document Nvidia wrote to announce the deal, and it says nothing about the amount, the exact date or the terms of the deal itself, which still rest on Calcalist alone.

A billion-dollar week: Run:ai and Deci together

The acquisition took place in the same week that Nvidia announced its acquisition of the Israeli company Run:ai for about $700 million (as reported; the amount was not officially disclosed): two deals totaling about $1 billion in a single week, both in the infrastructure for running models and not in the models themselves. The team was folded into Nvidia's operations in Israel, which grew a great deal in the years that followed.

The cost of running models as a field of competition

The company belongs to a group of Israeli companies that worked not on models but on the cost of running them. As models grew, the question of how much each answer costs became no less central than the question of how smart the model is. In a market where every percent saved on hardware adds up to millions, this is something corporations are willing to pay for.