CoreWeave Fully Connected 2026 | Saurabh Sharma, You.com & Corey Sanders, CoreWeave

Clips
More from CoreWeave Fully Connected 2026

Saurabh Sharma

CPO

You.com

Corey Sanders

Senior Vice President Product

CoreWeave

play_circle_outline Agents need up-to-date web context for real-world tasks

play_circle_outline Index-fed RL adaptation and hot weight P2P syncs to eliminate model cold starts

play_circle_outline Nemotron 3.5 Lightning: Eight-Hour Fine-Tuning and the Shift to System-Era AI with Tools and Data

Info
Transcript

Saurabh Sharma, You.com & Corey Sanders, CoreWeave

Saurabh Sharma

CPO You.com

Corey Sanders

Senior Vice President Product CoreWeave

In this interview from Fully Connected 2026, Saurabh Sharma, chief product officer of You.com, joins Corey Sanders, senior vice president of product at CoreWeave, to talk with theCUBE's John Furrier and Dave Vellante about how post-training and reinforcement learning rollouts are becoming the new competitive edge in the AI race. Sharma explains how the explosion of agentic AI has driven surging demand for real-time web search, with You.com's independent web index feeding open-weight models the context they need to act. Sanders details CoreWeave's new RL Rollo... Read more

explore Keep Exploring

How does the company provide web search and real-world context to AI agents and open-weight models, and what partnerships support that work? add

What are the key factors and infrastructure techniques needed to make web-search–augmented, open-weight AI models work effectively at large scale? add

As frontier models become freely available, will post‑training (e.g., fine‑tuning or RL rollouts) become the primary competitive advantage for enterprises building large‑scale agent systems? add

bolt Powered by CUBE AI

Saurabh Sharma, You.com & Corey Sanders, CoreWeave

search

(INTRO)

Dave Vellante

>> Hey, welcome back to San Francisco, everybody. My name is Dave Vellante. I'm here with John Furrier from Fully Connected, CoreWeave's big customer event here in Moscone. Saurabh Sharma is here. He's the Chief Product Officer of You.com, and Corey Sanders, Senior Vice President of Product at CoreWeave, is back on

>> theCUBE.Yes, good to be

Dave Vellante

>> back.Great to see you in New Jersey.

>> That's right. I looked better there than I do here. Thank

>> you.

John Furrier

>> Thank you.Oh, we got product people, so that means you guys have the keys to the kingdom. So we got 14 minutes. Tell us all the secrets.

>> What's on the roadmap?

John Furrier

>> What's happening? How you guys

Dave Vellante

>> are—Start with You.com. what you guys are doing in intelligent search and AI, it's amazing. So give us the rundown.

>> Yeah, what we've seen is that as agents have taken over, agents have a voracious appetite for what's happening in the real world, whether it is stock prices, whether it is the latest news, whatever is out there. Agents need to understand that so that they can carry on their task, whatever it might be. And so we've been very lucky to be part of that journey. As agents have scaled up, their needs for search have scaled up, and we're there to provide those services for agents.

Dave Vellante

>> Yeah, so you guys do your own models, right? Maybe you could describe that a little bit, how you're partnering with these guys.

>> Yeah, well, we see, we've built our own independent web index, one of the largest in the world, to feed context into models, right? And so what we're seeing is with the rise of open-weight models, there's a tremendous desire to pair those open-weight models with web search to give those open-weight models the context about the real world. And working with CoreWeave, working with NVIDIA, what we've seen is that we can post-train these models to use web search, which is one of the most important tools in the agentic stack to provide great context to models, open-weight models.

John Furrier

>> I love that approach because the first generation was get everything trained up. Post-training, you mentioned, is key. The workloads, they all look different. Some are deterministic, some are deterministic on paper, but non-deterministic in the real world. Real world is unstructured. So take us through this era we're in now, because you mentioned APIs, you mentioned web. There's a lot of cross-domain action going on.

>> Absolutely.

John Furrier

>> What's the key to making that work at large scale?

>> That's a great question. I think that there's a couple of things that I think about. A lot of times customers will start and they'll say, well, okay, web search is important. We don't know how much our users will need it, but let's make sure our models are equipped with web search. You get the latest news, that type of thing. Every single time that this happens, within 3 months they come back and say web search has been enormously helpful, enormously important, because the way users are using these agents is shifting all the time. And what they're finding is that having more up-to-date information about what's going on has always been critical to expand the type of workload beyond what might have been originally envisioned to do more and more with the agents that they have in

John Furrier

>> mind.You mean from an optimization standpoint?

>> From an optimization standpoint, but also the way that users start to understand how to use these tools. We've all used these AI tools. We start with one thing, we find out 10 other ways to use them. And that's what our customers are finding out as

John Furrier

>> well.All right, Corey, question for you is, okay, that's the requirement. You got to make it all work, right? You got a lot of components now acting as a system. I wrote a post last week, hey, we moved to the system era. The whole thing is a computer, right? You got to make it all work in the system.

>> Yeah, yeah, yeah. So there's a couple of things to consider. And I agree with everything said about the importance of search as part of the AI workloads. we're seeing this from customers all over the place and doing it in the post-training. You know, one of the things we announced this week was RL Rollouts, right? The ability to effectively actively improve a model using data that a customer may have, including search data. And so a bunch of the things that we've done to try and make this work really well on our infrastructure, there's two big parts of RL Rollouts. There's sort of the model generation and then there's the weight synchronization. And so model generation, we optimize the infrastructure. We're very good at optimizing infrastructure. It's sort of what we do. For the weight synchronization, this is an area that we've done a bunch of work to be able to bring the weights in sort of a hot start versus starting cold every time. And then we've also worked with NVIDIA on the Dynamo ModelExpress to move weights in a much more seamless way. So instead of the weight always coming from object storage, we now get weights from other peers nearby. And a bunch of these optimizations, they all add up and they make it much easier for people to be able to do that type of thing.

John Furrier

>> I think the RL rollouts are huge. And because you're doing it at scale, it's not like you turn off everything and say, let's roll out.

>> Exactly. That's why the hot load is so important in that environment.

John Furrier

>> What is the need there? Because to make that happen, we think this is going to be a competitive moat— post-training will be always happening.

>> Correct.

John Furrier

>> And inference is skyrocketing. That's where the money is.

>> That's right.

John Furrier

>> RL is where the intelligence gets produced. So take us through the infrastructure requirements to make RL rollouts work. Take— that's important. Just break it down real quick.

>> In addition to the things I walked through on sort of improving how the data's moved. There actually is core infrastructure like our object storage, which we've built around trying to get data into the GPU as fast and as easy as possible. One of the things that we've implemented and added on to that is what we call sort of cross-region write. So we used to have cross-region reads, and when you do a whole bunch of massive scale training, you just want to make sure the GPUs have the data they need when they're running. For RL rollouts, we actually need to write it back, right? There's a moment where you say, okay, I've got my data, now I need to write it back for other people to use. And so this cross-region write was a big infrastructure implementation to try and make this type of post-training work better and faster and easier. You write like it's a local machine. We treat it like it's a global system.

Dave Vellante

>> How do you deal with the latency on that? Can you explain? Yeah.

>> So, the key thing is this, you write local and then effectively it takes time for it to—

Dave Vellante

>> It's an asynchronous operation.

>> It is an asynchronous operation, but anyone can pull from it. And so you end up sort of

Dave Vellante

>> forced—What happens if something goes wrong in that asynchronous write? How do you recover?

>> Oh, yeah, of course.

Dave Vellante

>> We're

John Furrier

>> Corliss.We're

>> wrong.Yeah, there are ways to sort of validate that you've done the write, that basically the right writer of the data has done

Dave Vellante

>> it.And that's part of the secret sauce.

John Furrier

>> The answer is

>> loops.That's loops. That's always the

John Furrier

>> answer.It's always the

Dave Vellante

>> locks.There's got to be locks. That's non-trivial though. I agree with that, right? Correct me if I'm wrong.

>> No, no, you're

Dave Vellante

>> right.You're right. But anybody can write

>> fast.Yes, that's

Dave Vellante

>> right.That's what happens when something

>> goes—Part of it is the synchronization challenge and understanding sort of who wins and all of that.

John Furrier

>> So exactly right.Saurabh, from your perspective, obviously the customer, you're building all these large-scale systems. So as the frontier models become freely available, is the post-training the new competitive advantage for the

>> enterprise?100%, right? What we're seeing is that the ceiling is no longer model intelligence. The models are getting more intelligent, but it's their ability to use the tools that dictates the agent's success, right? To provide context and take action from the model's intelligence. And so what we found with RL rollouts is that we are able to train a model like Nemotron 3.5 Lightning in 8 hours, right? That is the speed that CoreWeave's infrastructure provides and be able to use our tools significantly better, right? What does that mean for an end agent builder? What that means is accuracy goes up. What that means is that their total cost of ownership goes down because the agent and the model understands how to use the tool. It's kind of like a home improvement project, right? If you know how to use your drill before you actually start the project, it's gonna be that much better.

John Furrier

>> Oh yeah. What you're getting at, the outcome is that the workload-specific intelligence is peaked at the highest level.

>> Yes.

John Furrier

>> That's what you're getting at.

>> That's right. And now the challenge becomes providing the model and the agent with the right context. And the right tools to take action.

John Furrier

>> So this, I'm gonna make a statement. I want you to say if you agree or not. That's an easy one. I think the model is becoming the starting point. Post-training creates a differentiation.

>> I agree with that 100%.

>> the whole concept of the AI loop.

>> Yeah.

>> Is all about that. We wouldn't need an AI loop if all you needed was the model.

John Furrier

>> we believe that. We think that's facts based on our data and what you guys are doing. What happens next? This is the progression. Got workload-specific intelligence.

>> What our customers are seeing is that when they take a model like Nemotron 3.5 Lightning, when they apply post-training to it with You.com, they are able to run their workloads with higher accuracy and lower total cost of ownership. And so if you can do this in 8 hours, right, this is no longer something that only the frontier labs can do. Any company can do this. And that's what we proved out with RL rollouts, with Nemotron 3.5 Lightning, and with You.com.

John Furrier

>> All right, put a plug in for what you're working on as we end out. What are you guys doing for customers? Give a quick state of the art.

>> Yeah. What we see is that total cost of ownership is the topic, right? As model intelligence peaks, how do you get your workload done most effectively? That's what You.com is laser focused on, how to deliver that total cost of ownership advantage, working with CoreWeave, working with NVIDIA to deliver that as well.

John Furrier

>> Corey, close us out 'cause it's home game for you. Tell people why this event is so important. What's the big takeaway from this year, the first ever CoreWeave conference?

>> Yeah, look, I think we've talked about it a lot already here, which is it's so exciting to see what customers are building on the platform and what they're now able to do that even 6 months ago they weren't able to do. And our focus is around building the tools and the partnerships to make that possible.

John Furrier

>> And your product vision, first principle would be what?

>> Helping customers solve their biggest problems.

John Furrier

>> All right. We're here at theCUBE live in San Francisco. I'm John Furrier, Dave Vellante breaking it down. It's the AI intelligence era. It's a system. That's producing intelligence, a lot of things going on. At the end of the day, the workloads, the data, all key to making that system work. We're doing our part, bringing you the data here on theCUBE. Thanks for watching.