CoreWeave Fully Connected 2026 Keynote: Introducing CoreWeave Forge | Built for Adoption

Welcome to Fully Connected 2026. Built for AI. Thousands of pioneers ready to create what has never existed. Defining a new pace of innovation. Infrastructure adapted, models pushed, agents evaluated. That's performance proven. The world is waiting for the next big breakthrough in AI. And you're here to build it with a partner that's your force multiplier. This is where superintelligence accelerates and you're standing right inside it. This is Fully Connected 2026, built for AI.

Please welcome Chief Strategy Officer and Co-founder Brian Venturo.

Please welcome EVP Product and Engineering, Hen Goldberg.

Thank you, Brian. Hello pioneers. It's amazing to be here with you. I'm a builder myself. I'm an engineer at my core and at CoreWeave I lead product and engineering. And for the next hour we'll go under the hood together. What we are building, how it works, and what we are learning alongside the people in this room and many other rooms around the globe. I've been saying this a lot lately. There has never been a better time to be an engineer. Just think of the systems problems we get to solve. For us product and engineering folks, AI has changed how we plan features, troubleshoot, and of course, how we write code or don't write code. This amount of change isn't easy. For some, it raises an identity crisis. But on the flip side, it presents an opportunity. Just last week, someone on my team showed me something they built in only three weeks. A year ago, it would have taken an entire year. Think about your own work. What are you doing today that you were not doing a year ago? And this goes well beyond engineering. Everyone is building with AI. It's exciting. It also makes me think about what's next. What else could we do? Where else can we innovate? How do we remove those obstacles that are slowing us down? There are so many possibilities ahead of us. So today I want to show you what we are seeing across the industry and I'm not going to do it alone. I'll have a few collaborators join me on stage just like we collaborate in real life. That is what this event is all about. So, let's start with what all of you as builders have been up to.

Who here is building something epic?

I'm building... We're basically building AI models specifically for engineering plants for more robust health and to prevent future disease.

An early warning system for natural disasters. Something like the Nepal flash floods which happened recently. An app that helps small business owners navigate all of the regulations and zoning restrictions to making their shops as ADA accessible as possible.

Text to speech models and voice conversion models for the entertainment industry like the movie business and video games. We're currently building a benchmark that focuses on evaluating agents' capability on doing end-to-end frontier physics research. I built an agent that improves consumer apps' onboarding flows by generating synthetic personas and then taking them through those onboarding flows.

Platform to make estimating much, much faster for construction contractors.

Intelligent self-correcting voice agent for pharmacies.

Harnesses that help improve or create our own models specifically designed for accounting.

This is so cool. Did you notice that for the most part those problems were not new? AI is solving problems that have always existed around us. It's just that we finally have the tools to work on those problems and innovate. New possibilities are opening everywhere in every industry including the places you might not expect. Take Caterpillar. Their customers need to build more infrastructure. They are short of skilled people and keeping those people safe is a top priority. Those challenges have been there for years. Brandon Hutman has been with Caterpillar for 27 years. He has seen it all. Let's hear from him about how they are bringing AI into their existing machines and job sites, drawing on a century of engineering experience and data from machines working in the field.

It starts with a spark. A spark that inspires possibilities. A spark that ignites progress. For a century, Caterpillar has been that spark, powering progress across roads, cities, industries, and economies. Together, we've built the foundation of the modern world. But as the world accelerates at an exponential pace, our customers face new challenges every day. Understanding their needs fuels our purpose. We build a better, more sustainable world. To do that, we must listen, develop, and deliver tailored products and services that drive real outcomes. And those outcomes are just the beginning. Because sparking transformation means more than building equipment. It's getting closer to our customers without losing focus on creating quality products. It's leading the industry with differentiated digital and technology solutions and embracing new ways of working to accelerate innovation. Ultimately, it's about solving our customers' toughest challenges because the future demands it. We've honored our first 100 years and the strength of the legacy that propels us forward. This is our moment with technology at the forefront, customers at the center, and our people leading the way. The future belongs to us. Tomorrow takes courage. Tomorrow takes commitment and a spark to ignite it. Tomorrow takes Caterpillar.

Please join me in welcoming the Vice President of Physical AI Platforms and AI Autonomy at Caterpillar, Brendan Hutman.

Good to see you. I think that there's nothing probably that technical people don't like more than big machines.

It's pretty cool.

It's pretty cool. Thank you so much for joining us. It's amazing, right? Like you've been in this industry of course most of your career and your customers, they have a lot of things going on right now. What are some of the kind of challenges that...

Well, first of all, Hen, congratulations on the event.

Thank you.

It's amazing. Thank you for the opportunity to be a part of it and thank you for the close collaboration on the work that we're doing. So, at Caterpillar, we always start with our customers and for us, it's very simple. Solving their toughest challenges. You heard it in that video. And while we're doing it, building a better world. And if you look at what is happening in the industries that we serve now, there are real macro challenges that our customers are facing. I'll zero in on construction as an example. It's an industry that's growing, significant growth. A lot of it's driven by infrastructure and data centers. But then when you look at overall productivity, it continues to decline year-over-year. One that's very important and very passionate for Caterpillar is job site safety. Less than 5% of the workforce is employed in construction, yet 20% of the fatalities on the work site happen in construction. And then the last piece is just the demographics of the workforce. Our customers continue to face challenges with qualified and skilled labor to be able to operate the machines in the field. So when we look at all of those macro-level events, our customers need Caterpillar. They need our expertise and they need the technology that you saw in the video that we're actively working on.

And going deeper into that technology, you're at the forefront of your industry. What has changed technically? What does AI and physical AI allow you to do now?

You know, it's been interesting. We've been doing autonomy now for over a decade and we've been doing that largely in structured environments. We've been learning very closely with our customers. But if you look at that technology stack, it's largely deterministic and it's largely worked in structured environments such as mine sites. So a mine site as an example would be very structured in the haul roads that machines follow, the loading patterns, getting material in and out of a mine site. You look at construction as an alternative to that. It's almost the polar opposite. Unstructured, dynamic, machines can be doing different jobs on the site daily. When you look at just the nature of how close machines and individuals get with those machines, taking that deterministic technology and bringing it into the construction industry presented a significant challenge. And it was only just three or four years ago where the dream of being able to bring autonomy operations to a construction site almost felt unrealistic. But if you look at what has happened in the industry, obviously you know it started with the large language models, but the move to agentic, the move to real reasoning and then the move to VLA's, actual models that can perceive the world around them, reason on the action to take and then invoke actions in the real world. That's what we're excited about and that's really allowing us to make some breakthroughs.

I love how much you guys are leaning into the technology. So are you guys becoming an AI lab? What's your approach?

No, really, really it's not. We stay pretty humble back to what we're good at. We've got a what was mentioned at the intro. We've got a hundred years of serving our customers in these industries. We see that as a real advantage as the industries go through a transformation. We view what we're doing first and foremost. We're starting with our customers and it's a little different than the way we've done technology development before. We're actually doing it hand in glove with our customers. Bringing them in very early so we can understand how the technology and the autonomy, the other technologies that we're bringing that accompany that fit within their workflows and they can really utilize them to unlock value. So that's one place. The other place is we're not becoming a frontier lab. But what we are doing is we're working very closely with industry leaders. CoreWeave, Nvidia just to name a few. And the innovations that are happening in the industry. What we bring to the market is our hundred years of experience, the vast amount of data that we have on how job sites function, work, and operate. And then the ability to really bring these at scale to the industry in a way that few others have the opportunity to do.

And how has it been going so far?

It's very quick. It's very quick. It's quicker than any technology development that we've seen.

Not surprising.

Yeah, not surprising. And we would say that obviously the work we're doing with CoreWeave has been instrumental in helping us accelerate that work. So nothing to announce today but very soon what we saw in the videos there. We're excited that it's going to be in action on job sites soon.

Amazing. We love working with you guys. Thank you so much for the trust and the partnership and thank you for joining me. Thank you so much.

What you just heard is what we are hearing from all our other customers. You have the data, experience, and understanding of your domain and we focus on getting you the tools to succeed. And because AI is so accessible, more people can take a problem they understand and do something about it. Well, that sounds promising, but a little bit too simple. Here is where the complexity really creeps in. Getting an answer to a problem is very different from building a system that actually gets work done. You have more choices for different parts of a task. And don't even get me started about the complexity in the agentic world. An agent can keep going autonomously. Call a tool, read the result, decide what to do next, and call another tool. Industry data shows a comparison on this. The agent benchmark used roughly 142 times as many input tokens and more than 20 times as many turns as the chat example. And as the agent keeps working, the context it has to carry grows. Longer context needs more memory. More steps create more latency. And the cost of those steps can really add up. And what happens if an agent outcome is not as desired? You need to know exactly where it went wrong. Is it the data or the retrieval or how the agent uses a tool? Getting the right outcome is only part of it. There's also access. And does the agent use that access responsibly and securely? And if there's one thing that we have learned as an industry is that we never get things right the first time and the key is how quickly we learn and implement those changes. The good news, we are not starting from scratch. We've already solved that for infrastructure with infrastructure as code, declarative, versioned, reproducible deployments. We solved that for applications using agile development. And we improve the models through training, iterations, experiment and evaluation. And now to ship new products quickly, we have to bring model development, application development and infrastructure operations together as a single system. You know how we came up with it? We hit this bottleneck ourselves. We wanted to improve inference performance and build a speculative decoding workflow where a smaller model drafts the next tokens and a larger model verifies them. It worked and it got us to the top of Artificial Analysis output speed comparison for Kimi K2 at a time. But achieving that result meant manually stitching together training, evaluation and serving. Aaron from our team was our integration layer and not getting too much sleep those days. One very tired human in the loop. That's not the architecture we wanted to scale. So we started turning that manual process into a repeatable system. We also watched it across more than 1 billion training runs recorded in Weights & Biases. We pulled insights from hundreds of thousands of active GPUs across our fleet. One gives us a view of how teams develop and evaluate models. The other teaches us what it takes to run workloads at scale. It comes down to this. You run the system. You observe what happened. You curate the specific signals that matter, improve the systems based on them, and keep evaluating if the improvements landed or not. That helps you decide what goes next into the next version. That's the AI loop. Run, observe, curate, improve, evaluate, and run again. Researchers, developers, and system engineers. All of us in this room, we all need to work together. Our tools need to work together. Today, I'm excited to announce CoreWeave Forge.

CoreWeave Forge brings every step of the loop into one connected environment. So what you learn in one step carries into the next. So when you put the loop on top of our secure infrastructure, that's the whole CoreWeave cloud. The AI loop is how we get better at building with AI. And CoreWeave is how you go from first agent to best agent. Starting today, you can sign up for CoreWeave self-service and start for free. No platform commitment, no infrastructure commitment. Scan this code and start building on CoreWeave Forge today.

CoreWeave Forge makes every layer of the AI loop fully connected like the conference. Get it? But the one important thing I want to note is that connected does not mean closed. You stay free to build with the models, frameworks, and clouds you choose. Here you can see the services available with Forge from building and training to evaluation and more, but the value is not simply having them in one place. The point is that the services you choose share context and evidence so work moves between them without starting over. Now let's start with the loop. We start with running. Do you like that? Wait until you see observe. Okay, running means inference. We have spent years building the infrastructure for that. The compute, networking, storage and the software that operates them together. It is purpose-built for AI. But no two workloads demand exactly the same thing from the infrastructure. In legal work, you may need to connect information across thousands of pages of contracts, so context memory matters. In banking, fraud detection must occur while you are tapping your card, so latency matters. In manufacturing, missing a defect can be costly, so accuracy is critical. And when you are processing millions of product images for an online marketplace, the cost of each one adds up. Let me show you an example. This is CoreWeave Forge, one connected environment for the AI loop. Now let's go to inference. Here is the fastest way to get something running. Our serverless models with dozens of models pre-built and already serving and the catalog stays current. GLM 4.5 Flash landed this month ready to serve. You pick one, call the endpoint, and you're in production. For this example, we've already deployed GLM 4.5. We wanted isolated infrastructure, so we used dedicated inference. The same experience, but on GPUs reserved for us. Here it is. Running is the easy part. The interesting part is what happens next. We started running and you know our requests were getting slower and cost per token was creeping up. Even though the deployment everything looked healthy, but the success of the product has changed the workload, longer context and more traffic. Okay. So let's see. So I'm going to ask Mission Control agent what's going on.

Okay. Okay, so the first thing it will go to the basics, will say what kind of data I'm going to use. 15 minutes of data and just checking the basic stuff, looking that the replica is healthy. It seems like everything is going well. And now it's going to take a look at the KV cache, the prefix, the latency. Let's see what it's going to find. Okay, there is a root cause. Oh, okay. I need to make a decision. Two options. Option A is going to first of all look at the thresholds of the KV cache that we can see here and the long context were filling it. Okay, that means that the blocks were getting evicted early and we were recomputing things that we've already computed. Now let's take a look at the two options that I get to choose. In the first one there is a suggestion to scale up, and then for the second one, which means of course that I will need to have more GPUs deployed, for the second option it's suggesting that I will actually make some changes in the model and the weights and we can see the details here. Let's apply the second option.

So the KV cache was thrashing, long context were filling it, and we can see the details of the analysis. Okay. So now we are working on that. Okay. This is really awesome. What we can see is that we've rose the throughput from 3,000 tokens per minute per GPUs to 5,000 tokens per GPUs. But wait, now we can see another thing that the agent has discovered.

The agent flags something else. Is it okay? Now it is. We can do two replicas instead of one on the same amount of GPUs. The agent knows this rack, which GPUs share the NVLink domain and what communication costs between them. So we applied that too. We're confirming that we're now going to deploy the second replica and going to wait until it's healthy again. Seems using the same amount of GPUs. Both are ready. Now let's run the test. Okay, this is much better. Throughput rose again from 5,000 tokens per minute per GPU to 20,000 tokens per minute per GPU. This is awesome.

What I want you to notice is that the first improvement changed the next best move. That's what Mission Control is for. Continuous optimization. And here's the scorecard. We can see that the spend stayed flat the whole time. Every GPU produces more tokens per minute. Price per token efficiency has improved by two times and quality is unchanged. That's really a big part of Mission Control's job. It is our operating standard for AI on CoreWeave cloud. Every second it processes more than 200 million metric samples from workload behavior to the physical state of the rack and it connects those signals to show what changed, why it matters and what to do next. That visibility matters beyond performance. CoreWeave Mission Control brings the operational intelligence our engineers use to run the fleet to your workload, connecting what the application is doing with what is happening underneath it. That gives you confidence in the decision and the reason to trust the system. Now let's take a look at the inference from the infrastructure perspective. NVIDIA Vera Rubin NVL72 was designed from the ground up for inference in the agentic era. From bringing up and validating a single rack to clusters, we deployed it really fast. But today is the most important milestone yet. We are announcing the Vera Rubin NVL72 is ready for production with Cognition as our first customer. Here you can see the four Vera Rubin racks. Together they operate as one AI system with every layer from power, cooling, firmware, network and applications all engineered as one. CoreWeave designs and operates the entire stack to deliver a consistent platform that works not only for Vera Rubin but across every GPU family in our cloud. Everything you heard leads to one design principle. We cannot design the cloud around one model, one runtime or one version of the future. The layers must remain composable as the underlying infrastructure continues to evolve because the workload, as we all know, will change again. Today we operate six generations of Nvidia accelerators with a consistent operational model provided through Mission Control. We also work with customers to match workloads to the hardware across generations that fit them. And because of this we are able to run inference for teams whose problems are nothing alike. Grammarly runs consumer scale traffic on CoreWeave inference. Enormous volume, tight latency, and every request is a person waiting on a suggestion. Cognition runs the coding agents you will see later today. Long context, turn after turn, sessions that hold state for hours. Same platform, different requirements. We are ready for your success. Now, let's hear from someone who has been shaping this industry for more than two decades. Please welcome the Vice President of Hyperscaler and HPC at NVIDIA, Ian Buck.

Wow. Well, thank you to Hen and everyone at CoreWeave and thank you for joining today and it's an honor to be on the stage. I've worked with CoreWeave since the beginning and it's a pleasure to see their growth and to grow to this scale and now to bring these amazing products to life. I'm going to talk a little bit about why we built Vera Rubin, our platform, and also what we're doing with CoreWeave in that partnership. For perspective though, I think people still are a bit confused often about AI and what does it mean? For ourselves, we've been doing accelerated computing myself for two decades. But in 2023, that ChatGPT moment really activated AI. It became the fastest growing application ever. Had over 100 million users in just two months. And almost entirely all on a GPU workload. It was an amazing time. But in essence it was a simple human in the loop model. A human gave a prompt. We processed the context of that prompt. We ran it through a large language model and we produce the output. And this would repeat with a human reading like we all experience today in ChatGPT and other chat apps. But fast forward to today, the workload is much more complicated. AI is no longer one model answering one prompt. And an agent doesn't just do one response. It's actually running in a cycle. Humans may provide that initial prompt, but there's an entire language model that's processing the context. Another language model that's observing the prompt and understanding the meaning inside of that context. We then have to do calls out to tools and skills and learn things. These are a combination of both CPU jobs, some of which we're all familiar with, with basic CPU or database operations, compiling code, running checkers, running Python. They can be language skills, they can be CUDA applications, they can be simulations. All those tools and skills need to run on CPUs and GPUs and pulling across all the networking. We then the agents need to then reason, understand the results. Are they accurate? Are they correct? And perform other actions and repeat. While humans are still in the outer loop, the agents are in the inner loop thinking constantly and driving, producing tokens, consuming tokens, not at the speed of humans can read, but it's the speed of the computers can think, process, and ingest. This is a holistic, it's a data center scale problem. It's the kind of problems that CoreWeave knows how to build. Not just the infrastructure which is where it all started but as you can see from Forge, building a platform to allow this whole thing to live, breathe and execute for your problems and across the enterprises.

Nvidia built Vera Rubin for this use case. We have over 10 million active developers programming directly onto our GPUs and delivering an ecosystem of components and capabilities across all the CSPs, all the ODMs, all the OEMs, over 300 supply chain companies were there and built to deliver and bringing Vera Rubin to life. It's fungible, every model, every workload, open, closed, frontier, from the very beginning of CUDA 1.0 to today's frontier models and everything in between. It's available on our platform and we make compatible so the agents can use them so that customers can bring their workloads, their models and get the full benefit of that platform since its inception all the way to today. And finally, it's durable. With every generation of our platform, we make sure that we can always run all the previous jobs just that much faster, that much better straight out of the box. And then with a little bit of work, maybe an agent thinking a little bit harder, you go from a 3x to a 10x to even 100x. And as Hen just showed, just doing a simple even in a demo, a 2x deliverable is not just 2x faster tokens. You just got twice as many GPUs because you engaged the platform, you used Forge and got that much more performance and that many more GPUs for your work. Vera Rubin is a full stack AI factory platform. It's not just GPUs, it's also all the networking. It's all the accelerated storage. It's bringing together the NICs, the memory, the context management, and it's also CPUs. We are now Nvidia as a full platform and data center scale company. We're building not just the best GPUs, but the best networking and the best CPUs to make it all possible.

Nvidia Vera is a CPU that we first built to pair with our Rubin GPUs, but it's also an important CPU as a standalone agent CPU for agents. It was built to combine the sort of capabilities of the pre-cloud where we had very fast CPUs, really great single threaded performance, but not a lot of cores, not a lot of throughput. When the cloud came along, we all wanted to run web browsers in the cloud and allow the business world to scale. Performance wasn't critical. It was just dollars per core and how many cores we could put into a single CPU to maintain costs and allow things to scale. Vera combines those two. It delivers max single threaded performance and at scale with every core, every thread running at full tilt and delivering the fast performance for agents but also have the throughput to maintain cost and throughput. It's actually two, we have doubled the performance with our Olympus core. It has over 3x more core to core bandwidth and so it's one big giant monolithic core instead of tiny little chiplets. And it also delivers 40% lower latency for accessing memory. CPU is often a memory limited application with databases and being able to access data. Vera was designed to have the lowest possible memory latency for both serving workloads on GPUs but also just processing data for agents. And we see that in the benchmarks for sandbox startup time. Just how fast we can fire up a sandbox for an agent. 3x faster than what we're seeing on traditional CPUs. And then in terms of those long running workloads, all those agent workloads, this is a benchmark called Terminal-Bench. It was developed by Stanford. It is a full catalog of all sorts of tasks and tools, a fully open source of all the different things agents are doing and the tools that they're running. Across this benchmark, we're over 1.7x faster than the latest x86 CPUs. And this means the agents just run that much faster, get the results that much sooner and all doing it at scale. And finally, it's also a really amazing CPU for data analytics. This is a benchmark called ClickBench from ClickHouse. ClickHouse develops an analytic data platform that's actually used by most of the AI companies out there for doing things like governance and monitoring all their data and all the actions across all those data lakes and getting real-time updates and events for actions and operations that are happening across ClickBench. We're actually delivering 7x faster data processing across those workloads and Vera actually tops the leaderboard, is the number one CPU for doing ClickBench. So your data on the Vera Rubin platform and all your data processing whether it be agentic or just database or just monitoring is always going to get the best performance with the latest Nvidia GPUs and CPUs to deliver those great results and that quality service that you need.

Hen mentioned six generations of Nvidia GPUs. And what's amazing is they've never ever been turned off. They've been delivering value from day one back in 2019 with those first Volta GPUs all the way today, of course, with the latest Vera Rubin and they all can run all those workloads. In fact, today, while Vera Rubin is all about the multi-agent systems, bringing the latest technology and the latest platform to market at the lowest possible cost with the highest possible throughput, we of course have an army of Blackwell and Blackwell Ultra GPUs to do all the reasoning and run all that pre-training and to run all the workloads and runs every model. Hopper, which came to life in that ChatGPT moment back in 2023, can do all the supervised fine-tuning, can do all the model post-training of course and run all the models even going back to the Ampere generation before that which today is serving all sorts of small model inference, it's doing all the document processing ingest. If I'm a legal firm I want to ingest all my AI documents, look for patterns, recognitions, reason about it so I can help move my business faster and a platform being deployed today to do that. And even Volta, a GPU in 2019, pre-ChatGPT, even at the very beginnings of AI, in production today for doing things like call transcriptions and of course running CUDA applications, running simulations. They're used today to do acoustic wave simulations, for simulating the audio in this room, for coming up with new designs. They actually are using CUDA to actually serve customers today for doing audio processing and simulation to design things like stadiums or products or glasses that all need to understand audio. We can do that of course on CUDA, that's accelerated computing.

CoreWeave and Nvidia have been partners from the very beginning. We built a platform from Volta to Ampere to Hopper to H100s to Vera Rubin to NVL72. We publish and produce key libraries that enable those workloads. Libraries like cuDNN for Excel for AI all the way to inference. None of that actually serves anything. It takes a CoreWeave, a platform forged in order to activate it and make it real. And in partnership with Hen and her team and with Brian, with Mike, I'm really excited to be bringing the next generation of product and solutions to you guys through CoreWeave and bringing it to life. So with that, I want to thank everyone for coming. I want to thank also the opportunity to talk today and I really look forward to seeing what CoreWeave is going to build next. Thank you.

Thank you for the kind words.

Of course.

So, you guys have done an incredible job with six generations of Nvidia accelerators, all still in production today. So, I brought a little special gift here.

Oh, I love Nvidia swag.

We do. We do like our swag. So if you want join me in showing off our latest. This is the Vera CPU tray is coming to CoreWeave.

This one's yours. Ready to go. Yes, you can see it.

So I can't lie. It's definitely on my wish list.