CoreWeave has climbed to the top of a pack of independent AI infrastructure providers that rent GPUs to major generative AI vendors such as OpenAI, Anthropic, Google, Meta and Perplexity.
Like some of its fellow GPU-as-a-service providers, CoreWeave has been adding its own software tools to the mix.
On Thursday, the New Jersey-based vendor launched a set of unified agentic AI capabilities that help enterprises deploy autonomous AI agents in production and continuously improve them. The capabilities lean on CoreWeave’s May 2025 acquisition of AI development platform vendor Weights & Biases, for $1.7 billion, and integrate serverless reinforcement learning, inference and observability to post-train large AI models for agentic tasks using the Weights & Biases Weave platform.
In this Q&A, Corey Sanders, CoreWeave’s senior vice president, product management, discusses the software release and how CoreWeave plans to use its new full-stack approach to help organizations build, deploy and safely run AI agents and other compute-intensive AI workloads.
To start off with an easy question: CoreWeave and some other companies have been called neocloud vendors, GPU-as-a-service providers and AI infrastructure vendors. What’s your preference?
Corey Sanders: You might think that would be a softball question, but the term I prefer, and I should use with customers and partners, is AI cloud. I think we’re here to stay, and from an AI cloud perspective, the full value of what we’re offering, and in fact, a key component of our announcement on Thursday, is that offering of a complete stack. When you put all that together, it’s definitely much more than just a GPU as a service. I do prefer AI cloud over neocloud or new cloud.
When you started out, you didn’t have the five companies that you acquired in 2025, including Weights & Biases. But you quickly ramped up toward a full stack. Please talk about CoreWeave’s approach to building a full-stack system, progressing beyond just GPUs, and how your latest news fits into that.
Sanders: As you mentioned, even a year and a half ago, the company was much more focused on delivering GPUs and surrounding capabilities, including compute and storage. We’ve continued to evolve the platform and work with customers and partners. Meanwhile, the vertical integration of AI requirements continues to increase, and so certainly the first wave of customers looks to get the best observability, the best uptime, the best reliability, the best security for the GPU workloads that they’re running. But as they continue to grow and expand on us, they then ask for the additional services, the additional management to make it easier to do that training, that inferencing, while still built on that backbone of high reliability, high security, observability and performance. That’s really how the company grew and evolved into this new place. Because we have that full vertical integration.
Can you elaborate on this new package of capabilities you just released? As for the autonomous improvement system for agents, how does that handle critical failures of models in production? And what guardrails do you have to prevent agents from lapsing into unauthorized behaviors or engaging in adversarial user interactions?
Sanders: Assuming that you’re going to get the perfect answer in your first deployment has been consistently proven incorrect in software development and cloud over the years, and it’s no different in AI. The realization is that this was going to be a continuous improvement loop, or an AI loop, as you ask customers to build and learn on the platform, that’s what the group of these interconnected, closed-loop services that we offer do and are talking more extensively about this week. It starts with our inference platform being able to deploy and run the models in either a serverless or dedicated way. But to your point, it also includes having guardrails and controls in place and then having observability with Weights & Biases Weave to be able to understand the usage of the platform, what operations are going well, where improvements could be made and feeding that into our serverless reinforcement learning platform. That platform can make changes, do post-training work against the model.
What about enterprises that are already running agents on traditional infrastructure or on other platforms besides CoreWeave? What does the migration path look like? And how does CoreWeave’s unified platform integrate with existing MLOps systems, data pipelines and governance tools that companies already have invested in?
Sanders: A couple things that are important to note. First of all, Weights & Biases is a tooling platform. It’s actually multi-cloud. You can run your Weave platform anywhere that you may be running these workloads so that observability can come in regardless of the underlying infrastructure platform. As for the migration perspective, the inferencing platform for what we expose, we use OpenAI standard APIs, and this allows consistency. If you’ve deployed this elsewhere and want to deploy it on CoreWeave, the transition to our inferencing platform should be fairly seamless. We believe pretty strongly in that cross-cloud opportunity and believe these tools help get there.
To switch gears a bit, you have a long background at Microsoft, working there for more than 20 years before CoreWeave. Can you talk about diversifying beyond your original customer, OpenAI? Now you have Microsoft as your biggest customer. You also have AI data center lease payments, since you sometimes go through third parties to build your AI data centers. But what if Microsoft someday decides to go with its own hardware infrastructure instead of yours? What do you do in a case like that if you’re depending on those billions of dollars of revenue?
Sanders: First of all, nine of the top 10 leading foundation model providers build on CoreWeave. That is certainly an exciting position to be in. I would say our broader customer base is pretty diverse and growing more diverse every day, especially when you include the full range of AI researchers and developers. Certainly, we have a lot of focus on those model developers because they’re the cutting edge of the AI industry.
Editor’s note: This interview has been edited for style and conciseness.
