Last night, Nvidia founder and CEO Jensen Huang put technology developed by Mellanox and Deci, two Israeli companies that it acquired over the years, at the center of the company’s annual product launch event in San Jose. The main products were the Spectrum-X and Quantum-X (no connection to quantum computing) chips, developed in Yokne’am and Tel Aviv, that lower power consumption in data centers through optic technology that obviates the need for existing communications switches.
Such chips can make construction of data centers considerably cheaper for Nvidia’s largest customers, such as OpenAI, Microsoft, Oracle, and Elon Musk’s xAI, positioning Nvidia better in relation to Broadcom, its rival in both communications and AI chips.
The other main Israeli launch was the product of the $300 million acquisition of Deci last year. Before it was acquired, Deci was one of the Israeli companies developing large language models, the models that are at the basis of artificial intelligence. Now, Nvidia has revealed why it bought the company: a series of basic models for the field of AI agents, more complex AI software capable of solving more difficult problems. These are known as large reasoning models, and are aimed at companies that seek to develop agentic AI that will carry out complex tasks. The model, based on Meta’s open source Llama training models, is called Llama Nemotron, and will be marketed as open source in order to attract customers to the company’s for-pay products.
Above everything there hovered the spirit of Chinese company DeepSeek, which rocked Wall Street and Nvidia’s stock in particular, wiping almost $600 billion off its value, after it launched its R1 model that is of comparable quality to some of OpenAI’s best models and that won praise from leading industry figures such as Sam Altman, Satya Nadella, and even Nvidia itself.
In a series of slides, Huang presented DeepSeek and its like as a powerful growth engine for his company, on the grounds that its launch highlighted the need for many processors. Nvidia claims that the world of AI is in transition from the stage in which graphics processors are mainly bought for the purposes of training models to one in which most processors are bought in order to operate existing models and to provide service to a growing number of users. DeepSeek is a guest at the Nvidia event in San Jose, and has a stand there.
As a response to the DeepSeek phenomenon, Huang presented the chip developed before the launch of the Chinese model: Blackwell Ultra (GB300), the graphics processor with the largest memory in the industry (288 gigabyte) that provides inference power 35 times that of the company’s Hopper series (H100) and 1.5 times the power of the previous Blackwell generation (GB200).
RELATED ARTICLES
Exclusive: Nvidia to invest over $500m in new Israeli computing facility
Nvidia acquires Israeli AI infrastructure co Run:ai
Nvidia to buy Israeli deep learning co Deci AI – report
Nvidia’s Israel AI spending spree has downside
Another answer to DeepSeek presented by Huang was a product partially developed in Israel, a software environment called Dynamo that enables companies like OpenAI and xAI to manage the running of reasoning models and spread it over a larger number of graphics processors – up to a thousand of them. According to Nvidia, Dynamo will make it possible to deliver the same results with 30-times more efficient power consumption on the basis of the same computing power.
Nvidia will try to sell its new Blackwell chips in the form of a complete server rack with 72 graphics processors, BlueField communications switches, ConnectX Ethernet adaptors, and the new Quantum-X and Spectrum-X chips, all of them apart from the Blackwell processors products of Nvidia Israel.
The aim is to represent an alternative to companies such as Dell and HP, or AMD and Broadcom, that sell competing systems at lower prices. Nvidia hopes that the approach of being a single supplier, even if costlier, will give it a competitive advantage.
Challenge to Mobileye
Mobileye’s share price fell 3.5% yesterday shortly after Huang announced a new driving safety product called Halos, which he said was an AI-based product for controlling vehicle systems and preventing accidents, which is similar to Mobileye’s business and to a parallel system from Qualcomm, another competing chip company.
In a move that should worry Mobileye and other competitors, Huang presented an agreement with General Motors that covers all aspects of the company’s development of an autonomous vehicle. Nvidia already has agreements with Mercedes, Jaguar, and Land Rover relating to conventional vehicles.
The writer is a guest of Nvidia at the GTC AI Conference in San Jose.
Published by Globes, Israel business news – en.globes.co.il – on March 19, 2025.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2025.
