Lip-Bu Tan on Why Inference Cost Could Change Everything

At Vista’s 2026 Annual General Meeting, Robert F. Smith sat down with Lip-Bu Tan, CEO of Intel, for a conversation on one of the most critical topics in enterprise technology: the economics of running AI at scale. Thirteen months into his tenure, Lip-Bu is re-architecting Intel’s roadmap around the priorities that matter most in the next era of AI. Below are highlights from our conversation.

Robert: Lip-Bu, few people have a better view of what’s required to power the next generation of AI. As we enter the agentic age, what does this moment demand from the infrastructure layer, and what are the characteristics of the companies that you believe will capture the most value?  

Lip-Bu: The speed of adoption and impact is significantly bigger than the internet. The whole semiconductor foundation and infrastructure needs to massively scale for agentic AI. The training is pretty much set, but the part I think is much bigger is agentic inference and that has profound impact across every layer. The companies that are going to be really successful are providing something very disruptive. In terms of semiconductor design, AI is driving down the time and cost to build a chip. And so those two are defining success and failure, you’re either going to be a leader or not. And it’s the same thing for software: how do you embrace AI, optimize for infrastructure and then drive cost, performance and time?  

Robert: That drive, to optimize for cost, performance, and speed, is what makes AI partnerships so critical. From day one as CEO, you bet on inference over training. Can you tell us what inference is and why it matters in this context.  

Lip-Bu: Customers don’t want to have just one company, one silicon to do it all — and it may not be built or optimized for their particular workload. That’s why we have a partnership with SambaNova, and I’m so delighted that you joined me on that. That data flow architecture can perform with one-tenth or one-fifteenth of the power. It’s very power efficient. By pairing Intel’s processors with SambaNova’s architecture, we can deliver agentic AI solutions that run at a fraction of the power in an air-cool environment, making inference faster and more cost-effective at scale. And all this will be a very good fit for different applications. I’m looking forward to how we can use that to serve your 90+ software companies.  

Robert: Energy is top of mind for our investors. Capacity constraints are already triggering data center moratoriums in some states. How does the energy constraint affect inference at scale, and how do you see it playing out? 

Lip-Bu: In order to massively deploy the whole AI infrastructure, there are some constraints and limitations. Power is one. The other one, short term wise, is memory. I don’t see any relief until 2028. It’s critical. Storage and memory really make data work for AI. And then third, high-speed connectivity is becoming a bottleneck because if you have clusters of central processing units (CPU) or graphics processing units (GPU), how are they connected? The latency can kill you in terms of speed to respond.

To build AI cloud infrastructure at scale, clean energy like solar cannot keep up with demand. The one that can really scale is nuclear. And then right now, we have this oil crisis, so then nuclear and fusion are the way to go. But the key is how to scale it. Energy has become a big issue, and we have to collectively work on that. That’s why I got into SambaNova. We need a new architecture to drive, even though we like to have all this compute capability and scaling, you need to have a more elegant architecture.  

Robert: Let’s shift to the model world. Satya Nadella talked about bringing models to the data, not enterprise data to the model. As a CEO, how do you think about data sovereignty and how do you manage it at Intel? 

Lip-Bu: Data security has become very important. In the past, we have had different networking companies protecting the network. But now, I think even more important is how to protect the data. A lot of big corporations, even governments I spend a lot of time with, are asking: ‘How do you protect your sovereign cloud and protect the data that you have?’ For the enterprise, even one like Intel, we want to adopt a model inside that doesn’t compromise our data. The customer data is the most important, and I need to protect that very tightly. 

Robert: Driving change is challenging for any organization. What change management approach did you need to take for the size, scale and complexity at Intel? 

Lip-Bu: Culture is very important for an organization. Intel 20 or 30 years ago was a very strong and leading company. Over the years, complacency, layers of management — it just compounded year after year. I had to go function by function. The decision making was so slow that decisions you need to make in one week were taking months and years. I changed the culture to a kind of startup culture. You have to make quick decisions.  

Robert: Last question: as a CEO navigating this transition yourself, how do you use AI agents in your own daily work? 

Lip-Bu: We all only have 24 hours. I have 10 priorities for my employees and my board for this year, and now I can use agents to help me monitor where we are. I’m a big fan of Perplexity. Gemini has been very useful. And Anthropic’s Claude is fantastic. Sometimes I try all three and compare which one is more accurate and gives me more insight. You can’t just use one, you need to use two or three. Those help you make the decisions you need to make. 

To help navigate the technical concepts discussed here, Vista has included a glossary of key terms below. This conversation has been condensed and lightly edited for length and clarity and is part of an ongoing series featuring leaders at the pinnacle of business, finance and technology.

To subscribe to future editions of Vista’s Insights for Edge, click here.