If vLLM already solved LLM serving, why did SGLang appear?
Thank you for meeting us at the event
Bengaluru always brings a special kind of energy, and the quality of the questions you asked really kept our team on their toes. Whether we were talking about inference or digging into the specifics of Velocis, our AI cloud, those face-to-face chats were the highlight of our summit.
Thank you for being so generous with your time and for sharing the stories of what you’re building. We left the room feeling inspired by the thoughtfulness and ambition of this community.
What we showed at the booth
Between sessions, the booth stayed busy. We walked people through Velocis end to end: GPU compute on bare metal, virtual machines and Kubernetes, notebooks to build in with billing that stops when you stop, dedicated inference endpoints to serve from, and the cost and access controls that usually decide whether a workload gets signed off.
A lot of whiteboarding, and a lot of good questions we did not have quick answers to.
Neysa on Stage
Our Chief Revenue Officer, Xavier Kurian, took the main stage in Bengaluru to share some reflections on a challenge many of us are starting to feel: the moment our AI ambitions begin to outgrow the infrastructure we started with.
It was a conversation about the shift from the excitement of building a model to the daily reality of serving users. You train a model once. You serve it every second of every day, for everyone who shows up. Xavier’s point was that those two things cost very differently, and that most budgets were quietly built for the first one. He also made the case that running AI in production is a different job from running applications, which is why a purpose-built layer is starting to appear next to the cloud you already have, rather than replacing it.
What we heard on the floor
- The model question was already settled. Almost nobody wanted to debate which model to use. They wanted to talk about what happens after you pick one: routing, tail latency, keeping an endpoint warm, and what all of that costs at volume.
- Very few teams had revisited that choice since going live. Most default every workload to the most capable model available, because nobody gets questioned for using the best one. It is one of the quietest decisions on the whole stack, and it sets your bill for as long as the workload runs.
- Agents have changed the arithmetic. Teams building agentic workflows are seeing token volumes nobody budgeted for. A chatbot sends one prompt and gets one answer back. An agent might make fifty calls to finish a single task, and it does not stop when the office closes.
FAQs
Did Neysa speak at the Data & Gen AI Summit 2026 in Bengaluru?
Yes. Xavier Kurian, our Chief Revenue Officer, took the main stage on 12 August 2026 at Taj MG Road, Bengaluru, on why production AI has outgrown the cloud most of it runs on.
What was the talk about?
Three things. Token prices keep falling while token bills keep rising, because cheap tokens made it worth putting AI everywhere. What blocks production AI is usually governance and cost, not model quality. And running AI is a different job from running applications, which is why a purpose-built layer is appearing next to general-purpose cloud.
What is a Neocloud?
Infrastructure built for AI that runs alongside the cloud you already have, rather than replacing it. It is designed for GPU workloads from the start rather than adapted from general-purpose compute. Pricing is transparent. The tools are the standard ones your team already knows, like SSH, Docker, CUDA and Kubernetes. And current GPU generations are available without a long procurement cycle.
What is Neysa Velocis?
Velocis is our full-stack AI cloud. You get GPU compute plus everything you need to train, fine-tune and serve AI in production, with the cost visibility and access controls that finance, security and compliance want before they say yes.
Can I get the deck?
Book a meeting and we will walk you through it, and send you the parts that fit your workload.