NVIDIA (NVDA) Q4 FY26 earnings call: guidance and Q&A

Given the high cloud CapEx approaching $700 billion and potential cash flow compression, how confident is NVIDIA in customers' ability to sustain CapEx growth, and can NVIDIA grow if CapEx doesn't?

Jensen Huang expressed confidence in customers' cash flow growth, stating that in the new AI world, 'compute is revenues.' The inflection of agentic AI and profitable tokens drives extreme urgency for compute, shifting classical computing CapEx to AI, which requires significantly more capacity.

“In this new world of AI, compute is revenues. Without compute, there's no way to generate tokens. Without tokens, there's no way to grow revenues.”

asked by Vivek Arya · answered by Jen-Hsun Huang

What is the role of NVIDIA's strategic investments in companies like Anthropic and OpenAI, and how does the balance sheet serve as a tool to grow NVIDIA's position in the ecosystem?

Jensen Huang stated that NVIDIA's investments are squarely focused on expanding and deepening its ecosystem reach across the entire AI stack, from models to networking and DPUs. The goal is to ensure all AI ecosystems, whether for language, physical AI, or robotics, are built on NVIDIA's foundational CUDA platform.

“Our investments are focused very squarely, strategically on expanding and deepening our ecosystem reach.”

asked by Joseph Moore · answered by Jen-Hsun Huang

Networking revenue is rising, with 3.6x YoY growth in Q4. What is the current and foreseen run rate for Spectrum-X Ethernet, especially with new platforms launching?

Jensen Huang explained that networking is an extension of NVIDIA's AI infrastructure, with NVLink turbocharging the business. He noted that NVIDIA is likely the largest Ethernet networking company globally due to the success of Spectrum-X Ethernet. The effectiveness of NVIDIA's networking (10-20% better utilization) translates directly to significant financial benefits for AI factories.

“I think that we're probably the largest Ethernet networking company in the world today and surely will be soon.”

asked by Harlan Sur · answered by Jen-Hsun Huang

How should we think about NVIDIA's future roadmap regarding customized silicon by workload or customer, particularly with the move to dilate architecture and the Grok acquisition?

Jensen Huang emphasized avoiding 'dilates' to minimize latency and power, preferring reticle-limited dies in architectures like Grace Blackwell and Rubin. He stated that CUDA's architectural compatibility allows extensive software optimization across generations. The Grok acquisition will extend NVIDIA's architecture with low-latency inference technology, similar to how Mellanox was integrated.

“All of our GPUs are architecturally compatible which means that when I'm working on optimizing models today for Blackwell, all of that work and all that dedication to optimizing software stacks and new models also benefit Hopper and also benefit Ampere.”

asked by Christopher Muse · answered by Jen-Hsun Huang

What are the expectations for sequential growth throughout the year, especially with Rubin ramping in the back half, and can Gaming still achieve year-over-year growth in FY27 given supply constraints?

Colette Kress expects sequential growth, with Blackwell continuing to sell alongside the Rubin ramp in H2, noting strong demand for Rubin from nearly every customer. For Gaming, supply will be 'very tight' for a couple of quarters, making it too early to confirm year-over-year growth for FY27, despite strong end demand.

“As much as we would love to have additional more supply, we do believe for a couple of quarters, it is going to be very tight.”

asked by Stacy Rasgon · answered by Colette Kress

With more investment dollars shifting to inference workloads, what is the importance of CUDA?

Jensen Huang highlighted CUDA's critical role in inference, enabling technologies like TensorRT LLM and new parallelization algorithms for NVLink 72, which delivers 50x performance per watt and 35x lower cost per token. He stressed that 'inference equals revenues' for customers due to exponential token generation by agentic systems, and performance per watt directly translates to dollars per watt in power-limited data centers.

“Without CUDA, we wouldn't know what to do with inference. The entire stack from TensorRT LLM that we introduced a few years ago, which is still the most performant inference stack in the world, optimizing it for NVLink requires us to discover and invent new parallelization algorithms that sits on top of CUDA...”

asked by Atif Malik · answered by Jen-Hsun Huang

What drives the sustainability of mid-70s gross margins long-term, and are there innovations in memory consumption that can help maintain these levels?

Jensen Huang stated that the single most important lever for gross margins is delivering generational leaps in performance per watt and performance per dollar that dramatically exceed Moore's Law. This pace of innovation and extreme co-design allows NVIDIA to deliver significant value to customers, which in turn sustains their gross margins.

“The single most important lever of our gross margins is actually delivering generational leads to our customers. That is the single most important thing.”

asked by Benjamin Reitzes · answered by Jen-Hsun Huang

NVIDIA previously outlined $3 trillion to $4 trillion in data center CapEx by 2030. What application areas will drive this, and does the company still feel confident about this envelope?

Jensen Huang remains confident in the $3 trillion to $4 trillion envelope, explaining that the future of computing is token-driven, requiring 1,000 times more computation than classical methods. He emphasized that 'AI factories' will generate tokens for revenue, with agentic AI currently at an inflection point and physical AI as the next major wave. This new computing paradigm, where compute equals revenues, is irreversible.

“The amount of token generation capability that the world needs is a lot more than $700 billion, and I'm fairly confident that we're going to continue to generate tokens.”

asked by James Schneider · answered by Jen-Hsun Huang