In the BG2 Pod podcast published September 25, 2025, Nvidia CEO Jensen Huang talked about the recent $100 Billion investment in OpenAI and future of datacenter. The podcast can be accessed on Youtube: https://www.youtube.com/watch?v=pE6sw_E9Gh0
Here are some important points he made in the show.
The economics of datacenter has an incentive to always upgrade or select the solutions that has the highest performance per watt. Nvidia tries to compete by being the first to deliver the highest performance per watt solution. The bottleneck of a datacenter is the available electricity. For example a 1 Gigawatt datacenter is constraint to only operate within that electric capacity. Through innovation in chip design, chip process miniaturization, networking, and software, Nvidia has been able to obtain performance improvements with each generation. Jensen provided comparison of Hopper to Blackwell getting 30x in performance/watt in one year from NVlink72. He claims this will repeat again with Rubin and then again Feynman each year. To compete effectively competitors need to match Nvidia’s performance and improvement cadence. The hyperscalers will not use a competing solution even if it cost less than Nvidia if it can’t beat Nvidia’s performance per watt. The production output of a datacenter are tokens. If Nvidia can produce 30x the token than competitor, there is no reason to switch to a difference solution if it save 30% on cost but don’t offer that level of performance. Jensen also points out that his teams are optimizing the workloads and looking at the whole problem and not just focusing on GPUs. They optimized Ethernet using Spectrum X networking , the CPU, the GPU, NVLink scale up and optimizing outside the box and co optimizing with the software. Visit Nvidia white paper for more detail on how Spectrum X is better than traditional plain Ethernet. https://resources.nvidia.com/en-us-networking-ai/nvidia-spectrum-x