OC3 2026 | Tech Leaders Panel with Microsoft, AMD, Intel & Nvidia

(bright music) - Welcome everyone to the by now, traditional tech leaders panel at OC three. I'm Felix Schuster, CEO, and co-founder of Edgeless Systems. And it's fantastic to once again moderate this discussion with some of the key people shaping the future of confidential computing. This year's panel is titled Confidential Computing in 2026, A reality check on adoption, security, and sovereignty. And we have a lot to discuss. Where do we stand with bringing confidential AI to production? How do hardware vulnerabilities affect trust? And what does sovereignty actually mean in a world where workloads span silicon firmware operating systems and whole cloud infrastructures? And how can confidential computing help? To explore these questions and more, we once again have an outstanding panel of industry leaders assembled here. Joining us today are Anand Pashupathy, VP and general manager for Product Assurance and Security at Intel. Ravi Kuppuswamy senior VP of the Compute and Enterprise AI Solutions Group at AMD. Daniel Rohrer, VP Security at Nvidia. And this year for the fourth time in a row already, our long-term guest and supporter, Mark Russinovich, CTO at Microsoft Azure. Together, the companies represented here shape nearly the entire confidential computing stack. From CPUs and GPUs to platforms and AI structure. Amazing. Gentlemen, thank you so much for joining us today. Well, we have a lot to cover. Let's dive right in and start with a reality check. Mark Russinovich, in the past years in this panel, we talked a lot about confidential AI as an emerging paradigm, what turned out to be harder than expected and what turned out to be easier than expected with regard to confidential AI and bringing it to production? - So I think what's been... I wouldn't say harder than anticipated, it is just been a challenge. And that is that we're right now, have confidential AI acceleration that supports models that fit on a single server. We don't have the support for models that span servers. And so that limits a little bit this the kinds of models that can be implemented with confidential computing. Also the kind of training that you can do with confidential computing, with performance overheads that come with those limitations. We don't have the GPU to CPU hardware acceleration either in place yet with the current generation of hardware,