Nvidia CEO confirms Blackwell GPU architecture will ship in Q4 2024.
Nvidia CEO Jensen Huang addressed investor concerns about the Blackwell GPU architecture's delayed rollout and AI investment returns at the Goldman Sachs Tech Conference on Wednesday. Despite earlier promises that Blackwell would ship in the second half of 2024, Huang confirmed delivery would begin in calendar Q4, following a manufacturing defect that required a mask change during the company's Q2 earnings call. Unveiled at Nvidia's GTC conference last spring, the architecture promises between 2.5x and 5x higher performance and more than twice the memory capacity and bandwidth of the H100-class devices it replaces. Huang attributed shipping delays and stock volatility partly to intense customer demand: "Demand is so great that delivery of our components and our technology and our infrastructure and software is really emotional for people because it directly affects their revenues, it directly affects their competitiveness," he explained. According to Huang, demand for Blackwell parts has exceeded that for the previous-generation Hopper products, which debuted in 2022, before ChatGPT's arrival made generative AI a must-have.
On ROI concerns, Huang argued that the performance gains of GPU acceleration far outweigh infrastructure costs. He claimed that with Spark, "probably the most used data processing engine in the world today," users can see a "20:1 speed-up," which translates to a "10x savings" even if the infrastructure costs twice as much. On generative AI specifically, Huang stated: "The return on that is fantastic because the demand is so great that every dollar that they [service providers] spend with us translates to $5 worth of rentals." However, he acknowledged that ROI on applications and services built atop this infrastructure remains less certain.
Addressing AI use cases, Huang highlighted his firm's deployment of custom AI code assistants and claimed "the days of every line of code being written by software engineers, those are completely over." He also touted generative AI's application to computer graphics, explaining Nvidia's DLSS technology: "We compute one pixel, we infer the other 32." He argued such technologies would be critical for autonomous vehicles, robotics, digital biology, and other emerging fields.
Huang contended that smarter datacenter design could help drive down costs. He noted that Nvidia's modular cluster designs—called SuperPODs—create entire datacenters capable of condensing massive compute into single systems, which is why Nvidia can charge millions of dollars per rack because each "replaces thousands of nodes." However, he cautioned that putting these dense systems—as much as 120 kilowatts per rack—into conventional datacenters is inefficient. Instead, he proposed densifying sprawling 50, 100, or 200 megawatt datacenters into smaller facilities that can leverage liquid cooling for efficiency. With Blackwell, Nvidia's top-specced parts are designed to be cooled by liquids, positioning the company to drive this datacenter modernization.