EXCLUSIVE: Nvidia CEO Huang on Blackwell Sales, Vera Rubin and China (full interview)

Key Concepts

• Blackwell and Vera Rubin GPUs: Nvidia's latest and upcoming GPU architectures, crucial for AI and cloud computing.
• Supply Chain Management: Nvidia's extensive global network of partners (TSMC, SK Hynix, Micron, Samsung, Foxconn, Quanta, Wistron) for manufacturing and packaging.
• Rack Scale Architecture: A revolutionary system design for deploying GPUs in data centers, featuring the NVLink 72 switch.
• China Market Strategy: Nvidia's current forecast of zero sales in China due to US export restrictions, while advocating for re-engagement.
• Energy Constraints: The growing challenge of securing sufficient power for AI infrastructure expansion.
• Hyperscaler Commitments: Large-scale investments by cloud providers (like those partnering with OpenAI and Anthropic) for AI compute.
• Demand Visibility and Financing: Nvidia's disciplined approach to building out capacity based on confirmed demand and customer financial capabilities.

Blackwell and Vera Rubin GPU Production and Availability

Nvidia is experiencing exceptionally high demand for its Blackwell GPUs, with sales described as "off the charts." Cloud-based video GPUs are sold out. Jensen Huang, CEO of Nvidia, emphasized that while demand is immense, Nvidia has planned its supply chain effectively. They have ample Blackwell GPUs coming and are manufacturing them at a high volume. Similarly, Vera Rubin GPUs are also in production and planned for.

• Supply Chain Strength: Nvidia boasts the largest supply chain in the world, with key partners like TSMC for silicon manufacturing, SK Hynix, Micron, and Samsung for memory, and Foxconn, Quanta, and Wistron for packaging. These partners are performing exceptionally well in supporting Nvidia's needs.
• Planning for Growth: Nvidia has meticulously planned its supply chain to accommodate a "very, very strong year" and the upcoming Vera Rubin generation.

Vera Rubin GPU Roadmap and Architecture

The Vera Rubin GPU is on track for delivery around the Q3 timeframe of next year, continuing Nvidia's annual product cycle.

• Silicon Bring-up: The silicon for Vera Rubin, comprising seven different chips, is back in Nvidia's labs, and the bring-up process is underway across engineering teams. Approximately 20,000 people are involved in bringing Vera Rubin from silicon to systems, software, and algorithms.
• Seamless Transition: The rack scale architecture, which is "completely revolutionary," is designed for seamless transitions between GPU generations. This architecture, which includes the NVLink 72 (fifth generation, unique in the world), was first implemented with Grace Blackwell and Grace Blackwell Ultra. The same rack scale architecture will be utilized for Vera Rubin, ensuring a smooth rollout due to established supply chains and familiarity with the complexity.

Nvidia's China Market Strategy

Nvidia's current forecast for sales in China is zero. This is a direct consequence of US export restrictions.

• H20 Sales: There were no meaningful H20 sales in the past quarter due to a lack of demand, even if Nvidia were permitted to sell them.
• Commitment to "More Competitive Compute": Nvidia remains committed to working with both the United States and China to sell "more competitive compute."
• Importance of the China Market: Jensen Huang stressed that China is a "very important market" for Nvidia, the US, and China itself. Nvidia desires the opportunity to re-engage with the Chinese market with its "excellent products" and compete globally.
• Market Size and Impact: The Chinese market is estimated to be around $50 billion this year. Nvidia believes it's beneficial for American people to compete there, for China to receive Nvidia's technology, and for Chinese software companies and open-source models to be used globally.
• Advocacy for Open Competition: Nvidia will continue to engage with the US and Chinese governments to advocate for permission to compete in the open market. Until then, the assumption for market forecasts in China remains zero.

US Export Controls and Middle East Sales

The US Commerce Department has permitted Nvidia to export up to 35,000 Blackwell chips each to Saudi Arabia and the UAE through G42. However, these exports are subject to US requirements, particularly concerning preventing technology transfers to China via the Middle East.

• Preventing Diversion: Nvidia has a long-standing commitment to preventing diversion of its technology. They have rigorously investigated past speculations about diversion and found no evidence through extensive testing and data center sampling worldwide.
• Compliance Measures: Nvidia will continue to be rigorous in its compliance. Methods to ensure compliance include having operations run by American clouds or implementing technological and process measures to prevent diversion.

Energy Constraints and Infrastructure Buildout

Energy availability is a significant challenge for Nvidia's expansion, potentially more so than chip supply itself, given the company's rapid growth rate.

• Scale of Growth: Nvidia is growing at a rate of approximately 60% year-over-year, with quarterly growth sometimes equaling the size of an entire company.
• Comprehensive Planning: To manage this scale, Nvidia must be "world-class" in its supply chain and also work downstream with energy providers, power generator companies, and land/shell providers to ensure power is ready for deployment.
• Leveraging Network: Nvidia's extensive go-to-market network, present in every cloud service provider and GPU cloud, provides a significant advantage. This vast network allows them to find power solutions of various scales (large, medium, small) in different parts of the world.
• Architecture's Role: Nvidia's architecture is fundamental because it "literally runs every model," from frontier models like OpenAI's and Anthropic's to open-source biological and physical AI models. This universal compatibility ensures high demand and off-take regardless of the cloud provider.

Hyperscaler Commitments and Financing

Nvidia is confident in the financing capabilities of its hyperscale partners, such as those working with OpenAI and Anthropic, who have made tens of billions of dollars in compute commitments.

• Disciplined Investment: Nvidia's approach is "thoughtful" and "disciplined" in aligning capacity build-out with demand visibility and customer financing capabilities. This coherence is crucial before initiating large-scale investments.
• Exponential Demand: Companies like OpenAI and Anthropic are experiencing "exponential growth" in compute demand due to the rapid adoption and use of AI across an increasing number of applications. They are currently limited by available capacity.
• Partnership and Support: Nvidia is working diligently to provide more capacity and optimize their partners' stacks for efficient usage. They are committed to supporting the scaling of these "consequential companies" and are delighted to be partnered with them.