Conference Presentation, Fireside Chat
Nvidia Founder and CEO Jensen Huang on the AI revolution
- Accelerated computing is projected to expand beyond computer graphics to address problems beyond the capacity of general-purpose CPUs, entering new markets selectively based on specific acceleration needs rather than pursuing a universal accelerator approach.
- The company commits to maintaining architecture compatibility since 1993, ensuring current software runs on future architectures to protect developer investments, with a goal of delivering three times more revenue per gigawatt or three times performance per unit of spend annually.
- Offloading just 5% of runtime to accelerators is expected to yield 100x application speedups, with specific benchmarks citing 500x image processing improvements and significant gains in machine learning and Spark data processing, potentially delivering 20x speedups that reduce costs by 10x even if compute costs double.
- Data center strategies will shift toward liquid cooling in compact facilities over the next 10 years to accommodate high-density computing, while new AI superclusters will be released annually with seven coordinated chips to improve upon previous generations.
- Workforce dynamics are predicted to evolve as software generation scales by a factor of one million over the last decade, introducing roles such as "digital chauffeurs" and "autonomous" workers where software engineers are supported 24-7 by companion digital engineers, ending the era of manual line-by-line coding.
- Supply chain plans include scaling capacity in the coming year and the year after, with the flexibility to fab components in alternative locations to mitigate geopolitical risks, potentially at varying performance or cost levels.
- Blackwell systems are scheduled for shipping in Q4 with scaling continuing into the following year, responding to extraordinary demand where delivery of components is described as emotionally significant to the market.
- Generative AI investment is projected to generate an instant ROI where every dollar spent translates to five dollars in rental value, driven by the expectation that almost all machine learning workloads have evolved to leverage these accelerators.