The Inference Stack From Silicon to Serving | SambaNova, DigitalOcean, Vultr

Recorded live at NeoCloud Summit 2026, Hotel Kabuki, San Francisco. The inference market has changed dramatically in just six months. Use cases are moving from chatbots to agentic workflows, open weight models are closing the gap with frontier models, and new models arrive at a pace that makes infrastructure planning harder than ever. Together, these shifts are changing the balance of power across chips, models, and cloud providers. In this panel, Rob Naidoff of SambaNova moderates a conversation with Syona Sarma of DigitalOcean and Duncan Ng of Vultr on how the inference stack is evolving, from silicon all the way to serving models. Syona explains how the move to agentic AI has driven an order-of-magnitude increase in token volume, and why DigitalOcean now designs at the data center level, treating the whole data center as a system rather than racks of GPUs, CPUs, and storage. Duncan shares how enterprises are accelerating beyond their first proof of concept into voice agents, back office automation, and new workflows, and why they're rethinking how to acquire token production across multiple generations of GPUs and accelerators. The panel also looks at how model velocity is reshaping operations, with what used to count as fast inference now becoming the baseline for performance. TOPICS COVERED • How inference requirements have changed in the last six months • The shift from chatbots to agentic workflows • Treating the data center as a system • Keeping pace with rapid model releases • Day zero support for new models • Tiered architectures for speed vs. TCO • Smaller, fine-tuned models for enterprise workloads • How to evaluate alternative silicon vendors • Underwriting risk and long-term value for alternative hardware • Who captures the most durable share of the inference dollar • Open weight models and the rise of model routing • How enterprises are responding to rising AI spend • Customer demand for faster tokens • Heterogeneous and disaggregated inference • How NeoClouds can differentiate The conversation digs into what it takes to bring alternative silicon into production. Syona points to operational simplicity, day zero enablement, and seamless support for techniques like prefill-decode disaggregation and KV cache offload. Duncan explains why financiers want to know whether hardware will still produce value after its first contract ends. On the question of who captures the inference dollar, the panelists see value spread across infrastructure, platforms, routing, and memory and storage innovation rather than one clear winner. To close, both share how NeoClouds can stand out, from helping customers solve real business problems to building open, integrated platforms where customers can choose the parts of the stack they need. Watch the full panel for a practical look at the economics and engineering behind today's inference stack. :mega: Join us at AI INFRA SUMMIT 6 on December 9, 2026, in San Francisco. Register now → luma.com/aiinfra6