CoreWeave Fully Connected 2026 | Silas Alberti, Cognition & Chen Goldberg, CoreWeave

Clips
More from CoreWeave Fully Connected 2026

Silas Alberti

Head of Research & Founding Team

Cognition

Chen Goldberg

EVP of Product & Engineering

CoreWeave

play_circle_outline Agentic systems and Devin (Cognition): agents as coworker-like developers

play_circle_outline CoreWeave Forge and Agent Lens: Build and Debug Agent Workflows with Turn-Level Observability

play_circle_outline Convergence of training and inference: thinking of unified AI infrastructure

play_circle_outline Reliability, uptime, and GPU availability for large distributed training runs

play_circle_outline Use cases and economics: cyber defense, latency vs context window trade-offs

Info
Transcript

Silas Alberti, Cognition & Chen Goldberg, CoreWeave

Silas Alberti

Head of Research & Founding Team Cognition

Chen Goldberg

EVP of Product & Engineering CoreWeave

In this interview from Fully Connected 2026, Silas Alberti, head of research and founding team at Cognition AI, joins Chen Goldberg, executive vice president of product and engineering at CoreWeave, to talk with theCUBE's John Furrier and Dave Vellante about the shift from standalone GPU capacity to fully coordinated AI systems. Goldberg unveils CoreWeave Forge, a new platform that unifies the AI development loop — eval, inference, curation and improvement — into a single developer experience. Alberti explains how Cognition's agentic coding system, Devin, dep... Read more

explore Keep Exploring

Was it always your vision for AI agents (like Devin) to act as coworker‑like collaborators, and is that vision now becoming a reality and being deployed at scale (including in vibe coding and knowledge‑worker app development)? add

What exactly is CoreWeave Forge, and can you explain in more detail what it is (software layer/platform/experience) and how developers should get started using it? add

Do you regard infrastructure for training and inference as a single, unified AI infrastructure (rather than separate infrastructures), and how do you reconcile that view with decoupling training from inference across data centers? add

- What keeps customers on the CoreWeave platform beyond GPU availability? - How do you handle drift and synchronization in distributed training/inference (e.g., syncing rollouts with the trainer)? add

As models and applications move up the Pareto curve (becoming more capable and cost‑effective), what benefits will customers see in their work, and what new use cases will that enable? add

bolt Powered by CUBE AI

Silas Alberti, Cognition & Chen Goldberg, CoreWeave

search

(INTRO)

John Furrier

>> Welcome back, everyone, to theCUBE's live coverage here in San Francisco for CoreWeave's Fully Connected, their first annual conference, as they continue to have great business performance and of course, the latest and greatest on Vera Rubin, various software capabilities being announced here. I'm joined by Dave Vellante, my co-host here on theCUBE. We got Chen Goldberg, who's the EVP of Product and Engineering at CoreWeave. Chen, great to see you.

Chen Goldberg

>> Amazing to be here again.

John Furrier

>> Silas Alberti, he's the SVP of Research and on the founding team at Cognition. A fast-growing company doing amazing things, kind of taking an approach like a learning system. Guys, congratulations on what you guys do.

>> Thank you. I'm super excited to be here. Thank you for having me.

John Furrier

>> Very impressed with your venture and what you guys are doing because we've been saying on theCUBE now for like 4 years now, Dave, it's like a brain, Dave. The enterprise needs their own brain and then they could do all that work. And, you know, that's a little bit of hyperbole. But when you look at the agentic market now, coding has brought in the reality that there's a lot of innovation when you inject intelligence, but now you got to translate that into a working system. Multi-step reasoning, agents are taking action. This is like, are they in preschool? Are they growing? You know, like, take us through what you guys do, why this is important.

>> No, it's always been our vision, right? Two years ago, we announced our vision of AI agents. And at the time, people were like is it just next action prediction, there was like chats and there was like in your IDE, you'd have a little sidebar and you could talk to an LLM. But we always had this vision of Devin, the agent that almost functions as your coworker. And yeah, I think this vision is now reality, honestly. And I think people are deploying this at scale.

John Furrier

>> And if you look at a lot of the where vibe coding has brought in kind of real app development, it's gone beyond that to Well, I want to, I'm using all these tools. I got Notion over here. I got doing this. So you start to see development go down to just knowledge workers. They're literally crafting applications that's going to help their life and think like they do. This becomes the new architecture and tech is through.

Chen Goldberg

>> And I want to, it goes beyond vibe coding, right? I think what's amazing about Devin and those systems that are thinking is that now our own team, the way we build CoreWeave is using those coding agents to build systems that scale, that are reliable. We call it the human in the loop. And if before maybe the human in the loop was tired, and now they have new tools that can solve those kind of problems.

John Furrier

>> But the point is, everyone wants to do this, okay? But a lot of people won't admit it, maybe publicly, maybe privately, they'll say, where do I get started? There's a whole shift. We're talking about a generation of IT, generation of business people, generation of developers, they want to get started. What, how do they do it? What's the architecture requirements to set this up? Because this is a new architecture. You guys have proven that you can build a stack in the GPU world now, system world.

Chen Goldberg

>> I think the first thing that I would like to call out is that unlike previous technology transformation where you felt like there was some more maybe resistance, like we've been talking about the cloud native and Kubernetes, There was some resistance for people actually picking up and leaning in. But today on stage, I had Caterpillar, okay, Brandon Hootman from Caterpillar, a 100-year-old company, saying how they're going to use AI with their data. So I think first of all, this is a big change that is happening. How people should start, it's actually our announcement from today. CoreWeave Forge. It is our mission to make the technology accessible and also the infrastructure accessible. So CoreWeave Forge is available today and it is tailored for those AI Loop, those learning systems of how you build agents and connecting all the different steps. You don't have to do everything at once. You can always start with serverless inference, find your model and start. But from there you can evolve to the different steps of the AI Loop.

Dave Vellante

>> So you had the infinity loop on stage. Yes. You started with eval, you had observe, run, which is inference, curate, and improve.

Chen Goldberg

>> Improve. Yes.

Dave Vellante

>> So what exactly is Forge? It's a software layer. It's an experience. Explain it in more detail if you could.

Chen Goldberg

>> It's a platform. It's an experience for developers. The interesting thing is that we came up with it by our own experience. So we had a team that were building an internal inference service improving inference performance. So we had to build a speculative decoding workflow as part of that. And it was hard. We had to stitch everything together. So what we have done, we've taken some of the assets that we already have, like Weights & Biases Models. we've announced our 1 billion runs. But together with that, we're also announcing some new capabilities like Agent Lens. Okay, Agent Lens allows you to get tracing from the turns that your agent is doing. Because let's say you have 40,000 turns running and only 200 of them are not the right ones. There's still a human or a customer getting the wrong experience. So we've announced that, we've announced a Model Distillation that allows you to improve the model as well, and also RL rollout. So together it's a system that allows people to get better results, better cost, all together.

Dave Vellante

>> Silas, so many questions for you as the customer. Michael Intrator at the Goldman conference said we don't think about infrastructure for training and separate infrastructure for inference. We just think about AI infrastructure. Is that how you think about it? Because I think when I think about what you do with SWE, you decouple training from inference across your data centers. But so how do you think about it and help me square that circle?

>> Yeah, First of all, think about compute, and it's true. I think compute is increasingly training and inference and both, right? So for example, if you look at a reinforcement learning training run, this workload actually does include both a lot of training but also the inference. And to the disaggregated point, the cool thing is these days you can actually do distributed training across multiple data centers. And sometimes we've been training across multiple countries or even multiple continents. For one of our training runs, we've been, I think it was like 4 countries across 3 continents.

Dave Vellante

>> So a lot of customers tell us that the GPU availability got them to CoreWeave. They talk about it takes 6 months to get it from a hyperscaler and get it in weeks from CoreWeave. I don't know if that's the case for you, but if so, great. But what keeps you on the platform?

>> There's many layers to this. I think a big one is reliability also. I think that if you run a big training run, you're not just looking at the GPUs, but you also want uptime. People say goodput. And you—

John Furrier

>> I slipped there.

>> The training run runs across thousands and thousands of GPUs. And if just one replica goes down, the whole training run goes down. So I think a very important metric is getting to this 99.99% reliability.

John Furrier

>> How do you deal with drift?

Dave Vellante

>> So the question is—

John Furrier

>> How do you handle the drift and the syncing up on this distributed inference?

Chen Goldberg

>> That's a good question.

>> It's a very good question. So what we look at is, and this is a little bit of a mathematical concept, but we're looking at the KL divergence between the trainer and the rollouts. So one of the reasons why we love to hire mathematicians is because there's actually a lot of very cool algorithmic improvements you can do to your training run. And as you improve the algorithm, it becomes more robust to being asynchronous.

John Furrier

>> Last time we had the chance to chat with Chen was when they launched the NVIDIA Vera Rubin NVL72, they were first to get it. You guys wanted access, you got access to the Vera Rubin. I have to ask, why did you want it? What was the interest? What was the mindset and what did you do?

>> It's a very, very good question. I think there's two answers. First one is obviously with each generation, price performance just goes up, right? So we can do more with the same amount of compute. But I think the more nuanced answer is that with each generation also you think about what are the architectures that we, 'cause you co-optimize the architecture with the hardware. And so being early and actually being able to study the kernels and the dynamics of this new platform allows us to prioritize our research investments, honestly.

Dave Vellante

>> So do you expect that there will be things that you can do with Vera Rubin that weren't possible with previous generations?

>> I think the big thing that each new hardware generation unlocks is the efficient training of like a new level of scale, right? Because I think at a certain point of scale, it just doesn't really become economical anymore on a given hardware generation. So I expect that as Vera Rubin rolls out more widely, we'll see a general increase in the scale of parameters and the size of the training runs that both us and our competitors will achieve.

Chen Goldberg

>> And specifically with Vera Rubin, what you can see that NVIDIA has done there, where they were really focused on inference workload, like what has changed? They have the Vera CPU, which we also announced today, becoming available on CoreWeave Cloud. And the CPUs are becoming extremely important in those workloads because the things that Silas was just talking about from an RL perspective and the tools that you need to run. And also we've increased the memory bandwidth and the network bandwidth, right? So all of those are making that supercomputer be more efficient, close to each other, and allowing us to get those results that Silas is talking about.

John Furrier

>> I wrote a post last week around this thesis, the AI factory is the computer. And the thesis was we might not even call these data centers anymore. They might be industrial factories. But one of the thesis was that moving from the GPU model era to the system where the entire system is producing intelligence, as a system. You just kind of mentioned memory. These are new now, new architectures. So the question is, are you optimizing for the system or the individual components or both? How would you— look at that because you're basically looking at it as a system, but you just said it basically.

Chen Goldberg

>> Yes, we are doing all of the above. Well, our investments in previous generations have been paying off. Okay, I think that's amazing. Ian from NVIDIA was talking about it on stage that we have right now 6 generations of GPUs and every time we learn and improve. So we started with just making sure that the GPUs are healthy. Then we looked at the machine and the NVLink and now we are looking at the rack level and now we are, testing performance across it and actually running multiple racks across different regions. So every time we go, we improve on that.

John Furrier

>> That's called multi-generational systems. I need to ask this question because this comes up a lot. In older generations, when I was your age, they would throw out the old stuff and bring in the new. The new, this is a key factor of the learnings because they're not going away and the Pareto curve certainly maps out to that. Talk about this multi-generational hardware component and why the software stacks work with it well and things that make you guys more competitive.

Chen Goldberg

>> It's very important that the entire stack will be agnostic to the infrastructure or to the hardware, right? Because the other thing that I said today on stage was that we know that we actually don't know how the future will look like. So how do you build a system that can continue and evolve and be composable and interoperable. So with that in mind, what we are building is really a great foundation that connects all the data. And what it gives, Silas was talking about a new generation is coming, I might need to make some changes with my applications, right? But you also wanna make sure that the other things will continue to work. You don't wanna change everything at once. So we have Mission Control that connects all the data, the brain of our operation. With it, but those different generations are suited for different type of workloads and also allowing people to keep the workloads running.

John Furrier

>> Silas, if you don't mind me asking, because we'd love this mixture of experts vibe we have with theCUBE, because you're an expert, open weights has become a huge discussion. And if inference is the economic unit, that's the value, it's like a brain would want to not fire off a bunch of compute. Weights are like neural pathways. Once you can optimize. How do you guys think about that? 'Cause this comes up a lot in some of these tech conversations. Well, if you're gonna build a neural network, like the brain, these neural pathways can be highly efficient. When I put my hand on the stove, I don't think to pull it off, it just comes off, right? Just that. So you're starting to get into that kind of, I guess, brain analogy. But how does that translate into the mechanisms? How do you guys think about leveraging open weights and just how to make the models and the inference portion very fast, like the hand on the stove versus real thinking. I want to solve a problem, tap the right side of my brain, that kind of thing. How do you guys think about that?

>> Yeah, I think OpenWeights obviously has been a great kind of tailwind for us as we start investing in our own research program and building our own models. And I think you were talking about how do we make it fast? What is nice about customizing models? I think we've always also been about speed. So our models are some of the fastest in the industry. And I think it's because we like optimize the entire system end to end. I think also another nice chart that I showed today in the keynote is the Pareto frontier moving to the top left. So as we train the models, we don't just make them smarter, we also make them cheaper and faster. And that's great for everyone, for us and also our

John Furrier

>> customers.What's the secret sauce on that?

>> I think there's actually some secret sauce there, which is the way we do length penalties. So I think if you want to push the Pareto frontier, not just to the top right, but to the top left. So you want the model to not just become smarter, but more efficient. You need to do some machine learning science and figure out how to incentivize the model to become more resourceful with its tokens.

Dave Vellante

>> And the spectrum of Pareto determines sort of what infrastructure you utilize underneath. Are you deliberate about that or do you sort of let AI decide? How do you adjudicate that?

>> It's a good question. I think definitely different hardware can sometimes have different Pareto trade-offs. But I think even on a given system through the way you configure your inference stack, you can achieve different trade-off points between speed and cost and performance.

Chen Goldberg

>> Yeah.

John Furrier

>> I guess my question now is, okay, we see what's happening, we kind of get what it means. What's next? What's that mean for the developer out there? I'm going to build an app, I want to start a company or I'm an enterprise, I want to integrate and inject intelligence and have the system, certainly hybrid cloud is distributed computing. You can have factory nodes everywhere. How do you think about that and how do you build it for the enterprise? I got to get my team mobilized around this.

Chen Goldberg

>> I think that there are two parts of that thing. I think the first thing is that enterprises should think about what are their core assets? What do they bring to the table? It's their data, it's their expertise, it's their experience, it's the customers. What problems are worth solving? That will be first. The second thing is that we are happy to help them work through those iterations. Just start, run something, right? And move ahead. The key part, by the way, that I think is also really interesting with open weights is that some customers are saying, hey, we only want to use open weights models. Our data, we're going to do maybe our own fine-tuning or training, model distillation. Doing more of that. And open-weight models just open more possibilities. So we are excited about that as well.

John Furrier

>> And applications for the customers when they start to go shift up in the Pareto curves, what benefits will people see in their work? If people want to do their best work, what happens?

>> I think cost is not just cost saving, but also unlocks entirely new use cases, right? For example, we're looking at a lot of customers that now want to use agents for cyber defense, right, protecting all their systems, and they do a cyber scan and they find $30,000. And I think with the current price of the frontier models, it's just not economically feasible to fix all these costs.

Chen Goldberg

>> It's more painful. It's not funny.

John Furrier

>> Yeah, I'm laughing because I can just feel the pain now.

Chen Goldberg

>> Yes.

John Furrier

>> A nervous pain laugh.

>> And so it's like, it's critical that we are able to bring these costs down an order of magnitude to make cyber defense possible.

Dave Vellante

>> So we've established that you've got the latest and greatest, fastest. That's attractive to customers. We've established that the quality of the experience kind of keeps you there. I'm curious, Silas, as to, what you want from Chen and CoreWeave and what you're working on. we haven't talked about your compliance, your security. I think indications are you're ahead there as well from our customer conversations. But what is it that you want from CoreWeave? I mean, Chen's right here. What do you want her to work on? And I'm interested, Chen, on what your priorities are.

>> I think the biggest thing for us in the next year is getting to the next level of scale, right? So I think we need to work together on building bigger connected data centers, scaling our fleet across the world, and also having a graceful transition into the new generation, 'cause each generation is a challenge. And I couldn't trust anyone more than CoreWeave to figure out how to run Vera Rubin reliably.

Dave Vellante

>> Sounds like an engineering challenge.

Chen Goldberg

>> Yes, and we feel it's a privilege to work with teams like the one that Silas has at Cognition and innovating with them. So definitely we continue to scale. It's actually fun because now we have the efficiencies of scale with the amount of data that we are collecting and the automation that we can do. It used to be harder for us, but now if you take a Vera Rubin rack, it takes us days actually from getting the rack to handing it over to a customer. That's amazing. But to your point around enterprise, you will also see in our roadmap investing a lot more in the table stakes that customers expect. We've already started doing that. We've announced, for example, our Remote Key Encryption. So things that people will need when they are working in a multi-cloud fashion. And the last thing, of course, we announced our Partner Network. So we are working with the ecosystem to close those gaps.

John Furrier

>> I love your line in the keynote, it's the best time to be an engineer. It's simply, it's one of the best times And it's interesting, data centers were once that glass house, the building that had all the servers in them. Then they became training clusters, popped out some great models. Now they're turning into living, learning, always-on systems. So this learning concept's why I love what you guys do.

Dave Vellante

>> Great GMs,

John Furrier

>> John.pump out that stuff.

Dave Vellante

>> That's the exciting part.

John Furrier

>> A lot of intelligence coming out of the factory line. It better be good cars, good intelligence, not bad intelligence. it is like producing cars. you're pumping out intelligence. If I'm a customer, they better be good. So what's this learning system? Explain this concept because this is where you guys are now and this is where the customers are going to want.

Chen Goldberg

>> Want in.

>> And I think actually the reinforcement learning workload is a very interesting one, right? Because you can just keep training, right? And so I think what you see more and more people talk about continual learning, you talk about training on real-world feedback. And I think what we're seeing, our runs are always on. And so while we still ship releases, package them up a little bit, I think the reality is we're always training, we're always like trying to find the next data and the next reward signals to improve our models.

Chen Goldberg

>> And there's also some of the new announcements that we've made in this conference. So specifically, we talked before about the model distillation, which allows you to think about a teacher model and a student model, right? And the student learns from the teacher and just gets better results with a smaller size. But also exactly that it's always on. Now we have problems because we want to update those weights and those checkpoints while the system is running. So we've also introduced RL, reinforcement learning rollouts, that does that without having to take down the system.

John Furrier

>> So that reduces either errors or disruption.

Chen Goldberg

>> It keeps the system meeting whatever SLA you decided. I also used that on the keynote, different systems need to optimize to different things. Some need to optimize for latency. Let's say if it's a fraud detection, okay, but maybe if I'm—

John Furrier

>> That's something I hope.

Chen Goldberg

>> It's a legal system that reviews contracts, then I care about the context window, and maybe inference cost. So each system has to optimize for that.

John Furrier

>> Well, it's great to see you again and congrats, great keynote. Thanks for coming on theCUBE. We really appreciate it, we're live streaming. Silas, I'll give you the last word. Put a plug in or talk about some of the cool things you're working on. What's the coolest thing you're working on right now? Put a plug in for Cognition. You guys are doing some incredible work, some really strong work at pioneering. What are some of the cool things you're working on you'd like to share?

>> One thing I'm very excited about is Devin (Cognition), right? So we're talking about the software development lifecycle. Now agents are not just planning the code, writing the code, reviewing the code. But also becoming real production maintainers. So now if there's a production issue, Devin (Cognition) can jump on it, respond to it. And by the time you wake up, there's already a PR and you can merge it.

John Furrier

>> All right. Again, this is, it's almost fantasy land. It feels so good. We smile because it's fun. It's geeking out here on theCUBE.

Dave Vellante

>> Congratulations.

>> Seriously.Thank you so much. All

John Furrier

>> right.I'm John Furrier with Dave Vellante, live streaming here from Fully Connected, the Fully Connected Intelligence Conference it's turning out to be the factory is producing intelligence. That's the economic value of the system. It's no longer about just the one device, the GPU or CPU, it's the whole system. That's the unit of value, it's AI factory. And of course the result is tokens per watt and intelligence. That's where the revenue is. A lot of good stuff's happening. We'll be right back with more on theCUBE after this short break.