Jensen Hwang Nvidia’s CEO identified that “the world is unsuitable” against the prediction that the demand for GPUs might be reduced because of deep chic. The demand for artificial intelligence (AI) moves from training to reasoning and emphasized that more chips might be needed.
Hwang introduced the following -generation AI platform ‘Blackwell Ultra’ within the ‘GTC 2025’ keynote held in Sanje, California, USA, and introduced it as an optimized product for AI reasoning.
He called the event “AI’s Super Bowl.” A couple of hours before the event, 1000’s of individuals gathered to cheer.
“AI has made an enormous leap. Now, inference and agent AI are demanding way more computing performance. We now have designed Blackwell Ultra for this moment. This multipurpose platform makes the inference of preliminary, post -training and AI more easily and efficiently.”
Blackwell Ultra is predicated on the Black Well Architecture, which appeared last yr, and includes the ‘GB300 NVL72’ rack scale solution and ‘HGX B300 NVL16’ system. It’s explained that it offers 1.5 times more AI performance in comparison with the previous generation ‘GB200 NVL72’.
Specifically, he emphasized that test time computing, or reasoning, is strengthened. It’s essentially the most advanced product when it comes to time and speed, adding to accelerating applications resembling AI agents and physical AI. It also announced a brand new software called Dynamo to support this.
Hwang said, “DipCicks, which appeared in January, are unsuitable that the demand for top -performance GPUs will decrease.” Because of this of reasoning, the calculation required for agent AI is greater than 100 times greater than I believed last yr.
In reality, 4 public cloud suppliers, including Amazon, Microsoft, Google, and Oracle, bought 1.3 million Hopper chips last yr, but they currently purchased 3.6 million Black Well Chip. The demand is “huge.”
Nevertheless, the announcement on the day just isn’t much different from the recent podcasts or official blogs. NVIDIA shares fell 3.2% on at the present time.
Following the Blackwell, a brand new architecture was introduced. Vera Rubin, equipped with several terabytes of memory and a custom design CPU called Vera, might be released within the second half of 2026. Rubin Ultra, which upgraded this, will appear within the second half of 2027.
He also unveiled an inference model that is healthier than deep chic. The model called ‘NVIDIA LLAMA NEMOTRON REASONING’ announced that it would outperform deep chic-R1 when it comes to accuracy and speed. That is suitable for the AI agent construction.
He also unveiled ‘DGX Spark’ and ‘DGX Station’, which were introduced as “AI computers.” It’s an AI personal supercomputer that enables users to construct, positive -tune and run various AI models of AI in the sting without cloud access.
As well as, General Motors and AI said they are going to expand cooperation to strengthen robots and factories and automatic driving.
Meanwhile, NVIDIA announced plans to determine a quantum computing lab in Boston. In cooperation with leading hardware and software manufacturers, it would develop quantum computing architecture and algorithms.
By Dae -jun Lim, reporter ydj@aitimes.com
