CoreWeave Fully Connected 2026 | Urvashi Chowdhary, CoreWeave

Clips
More from CoreWeave Fully Connected 2026

Urvashi Chowdhary

VP of Product

CoreWeave

play_circle_outline Forge & Aria: Accelerating AI Research with Natural-Language Iteration and Free Open-Source Model APIs

play_circle_outline Inference optimizations: vLLM, KV-cache routing, quantization, speculative decoders.

play_circle_outline Open-Source Unified Deployments: Kubernetes, Dedicated, and Serverless Inference

play_circle_outline Accessibility: developer-friendly, no-contract access enabling broader AI innovation.

Info
Transcript

Urvashi Chowdhary, CoreWeave

Urvashi Chowdhary

VP of Product CoreWeave

In this interview from Fully Connected 2026, Urvashi Chowdhary, vice president of product at CoreWeave, joins theCUBE's John Furrier and Dave Vellante to discuss how CoreWeave is building the AI improvement loop that connects training, inference and developer tools into a single production-scale system. Chowdhary unpacks Forge, CoreWeave's new platform that lets developers call leading open-source models via a simple API key, then distill, fine-tune and observe results in one continuous cycle. She explains why reliable infrastructure alone isn't enough — the ... Read more

explore Keep Exploring

How can users or enterprises get started with CoreWeave’s Forge, what services and capabilities does it offer (training, inference, model improvement/observability), and how much does it cost or how do you engage with it? add

How is inference being supported and optimized at CoreWeave to deliver high-performance, cost-effective AI workloads? add

What inference architecture choices has CoreWeave made that differentiate it from other providers? add

What news or announcements from CoreWeave Fully Connected should people pay attention to this week? Is Forge now open to everyone? add

bolt Powered by CUBE AI

Urvashi Chowdhary, CoreWeave

search

(INTRO)

John Furrier

>> Welcome back to theCUBE's live coverage here in San Francisco for CoreWeave's first conference, Fully Connected. I'm John Furrier with Dave Vellante. AI is taking the world by storm, but the new layers emerging in the infrastructure, neoclouds are now AI infrastructure clouds as they have use cases for real specialty workloads around AI native among other things. A lot of great things. Urvashi Chowdhary here, Vice President Product and AI Services at CoreWeave, former Amazon Web Services, Thanks for coming on.

>> Thank you so much for having me today.

John Furrier

>> Okay, so let's work backwards from the customer.

Dave Vellante

>> Everything.

John Furrier

>> First principles. You guys have great visibility in the market. Public business performance is great. The products, the customers we're talking to, they like it. And Forge is one of the big news.

>> Yes.

John Furrier

>> That came together. Explain where you guys are at for people who are hungry for the new category of asset-light developers. I'm coding away, I want to use some GPUs. Where do I go? I got product-market fit maybe. Or we talk about the services you have.

>> Absolutely. Super excited to share more about it. I think if you look at the AI developer's journey, they're looking to solve a problem and they want to do it as quickly as they can with the best performance and the cost to scale. So we've been really focused on building up the layers of our stack, building on top of the reliable infrastructure to build more managed services, whether it's for training, post-training, inference. And then also with Forge completing the AI improvement loop. So you can not only run your workloads, but you can also observe them. You can iterate fast, you can evaluate them, and you can continually improve how you get to those outcomes.

John Furrier

>> Prior to this conference, we've been following CoreWeave, like all the other fast-growing AI clouds. A lot of CapEx involved, a lot of financing. That's well-discussed on Wall Street where we have theCUBE at NYSE. But my question is always, how can I use it? How could we, theCUBE, or a customer, use it? The enterprise is hungry for agentic. Yes, that's a nice market. What's the service? Forge is there. How do I engage? How much is it going to cost? Scope the motion with the customer.

>> So Forge is free for users to get started today. There's a QR code that we flash in our keynote as well as you go to the website. You can get started right away. You can quickly access the leading open source models for inference and start calling those models with a simple API key. Once you want to improve your models, you can quickly distill them, fine-tune them. And then as you build your experiments, you can track them, you can observe your agents. So the entire loop from running your models and applications to observing them and iterating on them, along with Aria, which is our Automated Research Assistant, it can help you iterate through the loop much faster with simple natural language.

John Furrier

>> So training created the boom.

>> Yes.

Dave Vellante

>> Yeah.

John Furrier

>> Training the models. Inference is gonna determine the AI economy. So Dave had stats, he did a survey that inference has been creeping up and will take over and grow. Yeah, it was, and that was a huge survey. Take us through the survey.

Dave Vellante

>> Well, the survey, the N was small, but it was deep. We went deep.

>> Interesting.

Dave Vellante

>> It was 15 customers. 7 were heavy users. 6 were recent users. And then 2 decided not to go with CoreWeave for a variety of reasons. Of the ones that were using, their inference was maybe 10% in year 1, 90% training, shifted to 60/40 in year 2, 60% training. And then they expect it's going to be 90% inference going forward. The curves. And by the way, training keeps growing.

>> Yes.

Dave Vellante

>> It's just inference is growing fast.

>> Everything gets bigger and inference continues to grow at a faster pace.

Dave Vellante

>> And your purview is the entire portfolio, is that right?

>> Yes, I look after our training, post-training, and inference workloads.

Dave Vellante

>> So, okay, so I'm interested in how you're making inference perform better, be more cost-effective. We've been saying all week, sort of observing the similarities and differences with AWS in the early days. And there's some similarities in that you're expanding your portfolio beyond compute.

>> Yes.

Dave Vellante

>> Storage and networking and software. And presumably those go into making inference more effective. Michael Intrator talked about how we don't think about compute infrastructure for inference and infrastructure for training. We think of AI infrastructure. So your job is to make that all sort of work together. So can you explain that? What goes into making inference perform the way you want and be cost effective?

>> Inference is my favorite topic. I spend so much time thinking about it and working in this space. I think for CoreWeave, we have a really solid foundation for AI workloads because our cloud is specialized for AI applications, for high throughput, high performance with the latest generation of hardware connected with fast storage and networking. And so I think the foundations were really strong. And then what we've continued to do is build on top of those foundations at the next layer up. So if you want to run inference, you want models to be served with the best price-performance profile, you want to tune performance with the optimizations in the inference stack, whether it's in the engine like vLLM, or it's KV cache-aware routing and management of distributed inference with pre-fill decode disaggregation. So we are building the next layer up and building a lot of expertise and optimizations in being able to operate inference stack, and then also optimizing the next layer up for models. So our team invests a lot of time in quantizing models, building custom speculative decoders for popular open-source models so that when customers run with serverless inference or dedicated inference, they get the best performance out of the box.

Dave Vellante

>> So, so I know it's all about the compute, but is it right? There's obviously GPU, there's CPU, there's storage, there's networking. There is, as we said, the software above all this. Take us inside that portfolio. When you install, when you, you know, stand up, when you light up, no, turn on an AI factory. Right. The storage that's in there, that is obviously there's Spectrum-X from NVIDIA. What else is there? Are you— where's the CoreWeave value add? Is it in the optimization?

>> Maybe I can give a specific example for one of the trends we've observed recently is that customers are doing a lot more reinforcement learning where they're trying to improve models for agentic use cases with rewards and verifiers. And the bottleneck is actually inference when they're running RL rollouts. So something that we have built is support for RL rollouts. And one of our customers, You.com, I'm gonna share their example. We've built a collaboration with NVIDIA team, with their Dynamo stack, and with the customer where we enabled RL rollouts in a way that the customer was able to, improve their quality by 23%, and then reduce the number of tool calls they were making by 30%. So ultimately, what that, all of the optimizations that I talked about earlier in the 3 layers, what it led to was better performance at a lower cost. And then specifically around the optimizations we're doing, bringing in storage and other parts of the infrastructure in place, When you're doing RL rollouts, you're continually creating new model checkpoints and versions and you want those to roll out into your inference setup so you can scale it independently. And we were able to speed that up by 15x, which means your training runs fast and your inference is scaling while you're continuing to train your model quickly. So that was something that brought together storage, networking, compute, inference stack, altogether.

Dave Vellante

>> And if I understand it, this is one of the differences, at least now relative to vis-à-vis AWS, you're not, configuring, putting together your own flash drives and building your own storage arrays, building your own networks. You're optimizing those.

>> Yes.

Dave Vellante

>> Through your knowledge of how the system works, the software optimization. That's the focus of CoreWeave.

>> Correct. And we, we are investing in some infra-level optimizations as well, like you saw with the Vera, Valvey, Racky investments that we are doing, or in our Mission Control or in our inference stack with the performance optimization. So I think we're being very surgical about where we think we can deliver a lot of value for customers.

Dave Vellante

>> Thank you for that.

John Furrier

>> Dave had a question a couple interviews ago with CoreWeave leaders, and he made the reference of AWS, Parallel, which we've been comparing because it feels early, everyone's coming together. But the one question he asked I thought was good. I want to ask you because you worked at AWS. Amazon brought the two-pizza team. They brought the working backwards from the customer. We covered that on theCUBE with Andy.

Dave Vellante

>> No compression algorithm for experience.

John Furrier

>> All those little Andy-isms.

Dave Vellante

>> Be curious.

John Furrier

>> Learnings from the pioneering work that AWS did.

Dave Vellante

>> And it changed the industry.

John Furrier

>> Changed the industry. Do similar things happen with CoreWeave. So what's the equivalent learnings from the work you're doing. That's not the equivalent of a two-pizza team because you have operational excellence developing that you're learning with your team and the whole CoreWeave team and ecosystem. What's the new refresh for this wave of this new layer of infrastructure? Are there techniques? Talk about that. You mentioned Racky, which you guys built. No one else could do that.

Dave Vellante

>> The AI version of the leadership principle.

John Furrier

>> What is the CoreWeave version of these little knickknack phrases

>> like—Phrases. I think there's a few. CoreWeave is extremely customer-focused and obsessed, which I love. I think something that we value a lot is ownership within our team. So if there's a problem, people step up and we're not afraid to take on the hard challenges. So one thing that constantly inspires me is people who are willing to take risks and take on the hardest problems and get it done and have so much fun doing it. So we have a let's execute, let's go, let's move with speed. I think speed is our advantage. So the CoreWeave way is keeping some of those attributes alive and it spans across how we build our products, how we operate as a team. Hopefully that comes across as how we talk to customers.

Dave Vellante

>> How do you coin the term? I don't know, 3 years ago moving at the speed of NVIDIA. That really does seem to be one of the fundamental principles of the AI era. It's still forming.

John Furrier

>> Because you have operators. A lot of your customers are operating.

>> Yes.

John Furrier

>> They want to turn on, not just stand up. Cloud was just buy by the drink and connect to an API. This era is different. A little bit more integration, a lot more system.

>> Yes. I think we need to think of systems to help customers get to their outcomes and for customers that are training and at the frontier of the next model, like time to market and that speed is important. For inference, speed is also important when we think about performance and scale.

John Furrier

>> Okay, so here's a question for you. What inference architecture bets did CoreWeave make that others aren't? How would you answer that?

>> I think I talked earlier about how we've kind of built optimizations across the stack. One thing that we've been very intentional about is leveraging open source tools and technologies, contributing back to open systems so customers have flexibility, and then also building our services on top of each other. So, the different options that we have with Kubernetes and Dedicated Inference and Serverless Inference, they're all built on top of each other. So we learn from each layer and bring those optimizations to the next.

Dave Vellante

>> That's the difference, John, right? Remember Amazon, the early days, was criticized for its lack of contribution back to open. They had to hire Adrian Cockcroft to actually solve that problem.

John Furrier

>> Yeah, that kind of played out flat.

Dave Vellante

>> But you're open source from the start.

John Furrier

>> They're back in now. They're more intentional. But the question is, you mentioned earlier, tool calls.

>> Yes.

John Furrier

>> Tool calls has come up as not a good thing. Less tool calls, the better, because they burn tokens. So you're going to have lots of agents, lots of tool calls. A lot of model interactions. Yeah, a lot, a lot more agents, fleets of agents and a lot of model interactions. How do you guys view that feature bug part of the system and how do you use your architecture to handle that growth? Because I think that's going to grow on both sides.

>> Absolutely. I think tools are important, but at the same time, what customers really need— and this is where for Forge we added Agent Lens— you need observability. You want to know are the right tools being called? Are they working as you expect in your chain of tools that are being called? Which one is bottlenecked? So I think that observability with Agent Lens, along with secure execution with Sandboxes, that's also something that we launched earlier, complement really well. So you want to be able to trace everything, and then you want to, if you know what's wrong, then you can optimize and improve it.

Dave Vellante

>> The other sort of principle, I guess, if I can use that, sticking with the theme of leadership principles, is the infinity loop.

>> Yes.

Dave Vellante

>> We're seeing a lot of those in this era. CrowdStrike uses an infinity loop. You guys use it in a different context. But this idea of, broadly, a user surface, a client surface, interacting with the backend system of intelligence and then learning from each other in this constant improvement, iteration, where machines are actually heavily involved in that, as are humans. And maybe that percentage is shifting, but that loop seems to be something that's fundamental to this era.

>> Absolutely. And you called out rightly that it's very hard to just be a human in that loop. Yeah, we have built systems that can analyze the data at scale, give you the insights that are required. Make it actionable. You can also completely automate the execution of the loop if you like. So we want to be able to give customers control where they need it, but without being the bottleneck themselves. And our goal is with Forge and the AI improvement loop that we can accelerate it.

Dave Vellante

>> I heard Walter Isaacson this morning on TV talking about how the PC era— and Walter, he's been on theCUBE, great respect for him. But he said, well, the PC era was about bringing us closer to technology and AI is kind of moving us further away from the technology, allowing machines to do the work. And he said, actually, that's not a good thing. I think it's a great thing because technology never works the way you want it to work. And so there's sort of the hope, the vision, the promise of AI is we can speak in natural language. Anybody can be a great technologist. Again, another similarity to AWS, it democratized, you know, turned the data center into an API. companies like CoreWeave, you're turning intelligence into something that virtually everybody has access to.

John Furrier

>> You think about CoreWeave and AI Factory Vision, it's an entire system producing intelligence.

>> Yes.

John Furrier

>> And that's, I think, what cloud offered was hosting your SaaS app that was producing access to compute, and higher-level services versus building a data center. Here it's different mission.

Dave Vellante

>> Yes.

John Furrier

>> But it's all involving all the same componentry but tied together.

>> Absolutely. We want to create accessibility for these leading technologies and do it in a way that even if you're an individual developer signing up today on your own, you still get the best performance, you still get the best reliability, and you don't have to compromise. So I think that accessibility ultimately opens up a lot more possibilities and options for our customers to innovate, which is exciting.

John Furrier

>> Final question for you. What news should people pay attention to coming out of CoreWeave Fully Connected this week?

>> I think everybody should sign up and give Forge a go. Give us feedback. Tell us how it's working, how it's improving their AI lifecycle and development process. So I would encourage everybody to give that a shot.

John Furrier

>> Open to everybody. Is Forge becoming open to everyone?

>> It is open to everybody, correct. Yeah, you don't need a contract or anything to get started, which is exciting.

John Furrier

>> You seem excited. Great to have you on. What are you most excited about right now?

>> I think I'm most excited about seeing how customers will innovate faster with everything that we've built. I think the possibilities with the AI Loop, being able to observe, improve, automate your development cycle.

John Furrier

>> I think tying the loops was my favorite. Also, one, I won't say sleeper story, but maybe misunderstood story is the RL

>> Rollouts.

John Furrier

>> Yes.That's a pretty big deal.

>> And that's the RL Rollouts capability that I was mentioning earlier with the customer example that we launched. We're excited to see how customers will use it. We did CAB earlier in the week and reinforcement learning was top of mind for many of our customers,

John Furrier

>> so.Great innovation. Thank you for coming on theCUBE. Really appreciate it. Thank you so much for your

Dave Vellante

>> time.Good to

John Furrier

>> have you.I'm John Furrier with Dave Vellante, live coverage in San Francisco as we break down more. This is the first AI infrastructure cloud show ever, Fully Connected with CoreWeave. Again, the new category, a new layer in the infrastructure is developing, very specialized for the AI native world. Doing our best to bring the coverage to you. Thanks for watching.