RAISE Summit 2026: Discussing The Future of AI Infrastructure with Pat Gelsinger
Paris greeted this year's RAISE Summit with an unforgettable heat wave. Every dinner felt hot. Versailles Palace was spectacular, but not exactly cool. And nothing, not the palace, not the stage, not even a tuxedo, stopped me from sweating through it.
Fortunately, the conversations were worth it. Over the past two years, I've had the privilege of sharing the RAISE Summit stage with some remarkable people. But returning this year to sit down with Pat Gelsinger felt different. Speaking alongside Pat in front of an audience at the summit that drew nearly 9,000 attendees was one of the highlights of my year - and an opportunity to discuss where AI infrastructure is headed and the challenges that will define its future.
Pat and I have gotten to know each other over the past year, since his visit to our offices in Tel Aviv last summer. He has been generous with his time, and just as generous publicly, posting about our results and the progress our team has made. Pat is one of the great CPU architects of our industry. He has spent over four decades shaping the infrastructure the computing industry runs on, most recently as CEO of Intel and, before that, CEO of VMware. Having someone with his technical depth, historical perspective, and architectural judgment pressure-test our thinking has been a huge asset. We were lucky to have him with us.
Last year at RAISE, I used this stage to make an argument. Compute required a different approach, one where software and hardware are designed together, rather than treated as separate problems. This year, I did not have to make that argument alone. It was the starting premise of a conversation with someone who has spent his career building the systems the rest of the industry depends on.
Our conversation returned again and again to a pattern I think about constantly. The industry keeps responding to AI's compute demands by building more specialized chips: one for training, another for inference, separate silicon for prefill and decode. Pat pushed me on whether that fragmentation counts as progress, or whether it signals the industry is solving the wrong problem. We landed in the same place. Building more specialized silicon says less about progress and more about discomfort with something genuinely adaptable. AI is not moving toward a GPU-centric world. It is moving toward heterogeneous systems, where CPUs, accelerators, memory, and software work together. The architecture that wins will be the one flexible enough to evolve with the workload, not the one that forces the workload to conform.
This is the problem NextSilicon exists to solve. We are not trying to build the fastest chip for a single model. We are building a compute platform built to last, one that can run today's large models and tomorrow's larger ones. That matters because models are not just getting bigger. They are changing shape. They are becoming multimodal. They are using longer context. They are combining many neural networks into complex chains. They are moving toward high-thinking workloads, coding agents, reasoning systems, and models that will not be measured only in hundreds of billions of parameters, but in the 1T to 30T parameter range.
That changes what infrastructure is optimizing for. The question is no longer just fast inference - it’s about token economy: how many useful tokens can you produce, at what power, at what cost, and with what flexibility as the models keep changing. A system that looks good on one benchmark but breaks when the workload moves is not future-proof. The winning platform will be the one that can keep adapting as AI itself keeps evolving.
That is where Maverick matters. Maverick is built for a world where the next model is not just a larger version of the last one. It is built for a world of multimodality, retrieval, vector search, long context, orchestration, and agentic chains that involve many models working together. In that world, raw accelerator performance matters, but it is not enough. The architecture has to move data efficiently. It has to handle irregular compute. It has to keep up when the workload changes. It has to make the economics of tokens work.
And that is also why Arbel matters. Last June, we announced plans to productize Arbel, our own RISC-V processor, because the CPU is not going away. Pat has been right about this for decades. GPUs were never designed to displace CPUs. They were designed to accelerate them. For years, AI systems could get away with treating the CPU as a thin host, with many GPUs hanging off one CPU. That balance is shifting. What used to look like a 16-GPU-to-1-CPU world is moving toward something much closer to 1:1.
But Arbel is not yet another cloud CPU with a Zillion cores. That is not the point. Arbel is designed for the AI harness: the tools, orchestration, retrieval, scheduling, runtime, data movement, and ordinary compute that surround the neural network itself. Agentic AI is not one matrix multiplication loop. It is a chain of decisions, searches, calls, memory accesses, and models. The CPU becomes central again because the system around the model becomes just as important as the model - Arbel gives us control over that layer. It gives customers a path off vendor roadmaps that were built for a different era, and it lets our platform run more of the full AI workload without forcing customers to rewrite their software every time the workload changes.
None of this happens without RAISE. My thanks to Henri Delahaye and the entire team for building a platform that gets more ambitious every year. When I spoke here in 2025, RAISE brought together a little over 5,000 people. This year, that number was closer to 9,000, including a special address from President Emmanuel Macron himself. Watching that growth in real time tells me something. The conversation NextSilicon has been part of for years is no longer a niche argument, it is the conversation the entire AI industry is now having.
I left Paris energized, the same way I did last year, but for a different reason. Last year, I was making a case. This year, I got to make it alongside someone who has spent his career proving one thing. The infrastructure decisions we make today shape what becomes possible a decade from now. That is the conversation I want NextSilicon to keep having. With Pat, with RAISE, and with everyone else willing to rethink what compute needs to become.
Stay Connected
Stay tuned as we share more details in future blogs and announcements via our Nextletter.
About the Author:
Elad Raz is the founder and CEO of NextSilicon, a company pioneering a radically new approach to HPC architecture that drives the industry forward by solving its biggest, most fundamental problems.
