Conference Presentation, Fireside Chat
Nvidia Founder and CEO Jensen Huang on the AI revolution
- Jensen Huang, founder and CEO of NVIDIA (founded 1993), attributes the company's success to a vision of "accelerated computing" designed to augment general-purpose CPUs for specific, hard-to-solve problems.
- NVIDIA's strategy has focused on maintaining strict architecture compatibility to protect software investments, resulting in CUDA maintaining a massive installed base where legacy software accelerates on new hardware.
- The company transitioned from a gaming-centric GPU provider to a broad data center hardware and software platform, driven by the need to solve domain-specific algorithms in graphics, physics, scientific computing, and molecular dynamics.
- Accelerated computing targets the "magic kernels" of software, which often represent only 5–10% of the code but consume 99.999% of runtime; offloading these can yield speedups of 100x to 500x.
- Data center modernization involves densifying sprawling facilities into smaller, liquid-coolable racks; while individual racks may cost millions, they replace thousands of traditional nodes, significantly reducing cabling and energy costs.
- Over the last decade, computing scaling has exceeded Moore's Law by 10,000x (100x CPU scaling + 100x virtualization + 100x cloud), effectively ending the previous era of transistor-driven efficiency gains.
- Generative AI is shifting the IT industry from creating "tools" to creating "skills," such as digital employees, autonomous vehicles, and robot operators, potentially expanding the market beyond the initial trillion-dollar infrastructure wave.
- Customers report a 10x return on investment (ROI) from acceleration; for example, accelerating Spark data processing on GPUs can reduce runtime by 20x, offsetting a doubling of compute costs.
- Infrastructure demand is so high that every dollar spent on NVIDIA hardware translates to five dollars in cloud rental revenue, with current AI infrastructure capacity sold out globally.
- Software engineering at NVIDIA is entirely augmented by AI; every engineer now has "digital companions" for code generation, with Huang predicting the end of humans writing every line of code.
- NVIDIA's competitive moat relies on a full-stack infrastructure approach (seven different chips per system), custom software optimization, and a universal architecture compatible across cloud, on-prem, and edge devices.
- The new Blackwell system offers 4x faster training and 30x faster inference compared to the Hopper architecture, delivering 3x higher revenue per watt for a given power budget.
- NVIDIA's innovation cycle allows for a new super-cluster release annually, leveraging architectural compatibility to ensure software built today runs seamlessly on future generations.
- System deployment speed is a key differentiator; NVIDIA super-clusters can be operational within 19 days of shipping, whereas cobbled-together multi-vendor systems can take a year.
- The supply chain is highly concentrated in Asia, particularly Taiwan, with TSMC acting as the primary fabricator due to its unparalleled agility and process technology margin.
- NVIDIA maintains redundancy and intellectual property to shift fabrication if necessary, though TSMC remains preferred for its performance and cost advantages.
- Huang describes the current demand for AI infrastructure as "intense" and "emotional," noting the immense pressure on the company to meet global needs for competitiveness and revenue generation.
- Blackwell is currently in full production, with shipments scheduled to begin in Q4 and scaling into the following year to meet unprecedented demand.