CoreWeave

NVIDIA Vera Rubin NVL72 is available on CoreWeave! Cognition is already running Devin, its AI software engineer, in production on it. They measured it against their own NVIDIA GB200 NVL72 baseline. Up to 4.8x the total token throughput for SWE-2 inference. 3.8x the output token throughput for reinforcement learning. Weeks from bring-up to production. Same operating model, same tooling their NVIDIA GB200 NVL72 and GB300 NVL72 fleets already run under. Nothing re-plumbed. We brought up our first NVIDIA GPUs in 2017 on V100. Those clusters are still serving production traffic today. Nine years. Five architectures. One platform. Here's how we did it. https://crwv.co/utcq6

A workload that runs clean on eight nodes can behave very differently at 500. When it breaks, the symptom rarely tells you which layer failed. It could be compute, networking, storage, or software. That's when you need an engineer who has seen it before, not another best-practices doc. CoreWeave Direct-to-Expert puts our AI engineers in the channels your team already uses. These are the people who design, build, and run the infrastructure. They work with you from the first design review to production, and one expert stays on your issue until it's fixed. The right engineer, right away. Dive in. https://crwv.co/utcUU

Still catching my breath after Fully Connected, and this piece from Wylie Wong at Data Center Knowledge captures why that week mattered. The announcements weren't a collection of launches, it was a culmination. As Wylie writes, Forge in particular extends our strategy beyond GPU infrastructure to the software layer enterprises need to develop, operate, and refine AI applications. That's the arc: we built the best place to run AI workloads, and customers told us they needed more on top of it. What got us here was a year of rigor and discipline from our technical teams, building a cloud, purpose-built and fully integrated for AI. Metal to model, every layer. Grateful to the everyone who spent time with us that week, to the team who built all of this, and to the customers who pushed us here. https://lnkd.in/gFeAgtkz

Here are the charts behind Cognition's first customer-executed benchmark on NVIDIA Vera Rubin NVL72. They measured it against their own NVIDIA GB200 NVL72 baseline. For inference, NVIDIA Vera Rubin NVL72 delivers up to 4.8x the total token throughput per GPU of NVIDIA GB200 NVL72 at matched interactivity. For reinforcement learning, it delivers 3.8x the output token throughput per GPU. Agentic coding is an unforgiving workload. Long contexts. High concurrency. Token volumes where cost per token decides what you can ship. For Cognition, those numbers mean more concurrent Devin sessions per GPU, drastically accelerated research loops, and lower cost per session, with no loss in generation speed. The whole stack runs on CoreWeave: training, reinforcement learning, and the inference that serves users. All of it on one platform. The benchmark is theirs. So is the workload. https://crwv.co/utcq6

Deok Filho has a short answer for anyone who thinks CoreWeave only rents GPUs, and it's a list of what you can run in CoreWeave Forge. Talking with Alex Volkov on ThursdAl, Deok walked through how developers can use Forge. There's CoreWeave Serverless Inference, Serverless RL and supervised fine-tuning for post-training, CoreWeave Notebooks, and CoreWeave Sandboxes to run agents and RL environments. The case comes down to the loop. Pick a task, eval the model, deploy a new version, and go around again. Deok is a senior product manager on our ML products team, known as the Chief Sandboxer, focusing on CoreWeave Sandboxes. Sandboxes are already connected to the rest of CoreWeave Forge, so the evals and the training runs that move a model around the loop have nothing extra to stand up. Watch the clip, then try it here: https://crwv.co/utcVH

Know more, innovate faster. That's the CoreWeave ARENA. Stories from real founders, coming soon.

🚨Last week, I walked into Moscone Center in San Francisco to get my morning coffee. I was debating whether to get a Matcha or a regular latte, but turns out I ended up chatting with the CEO of CoreWeave. That conversation gave me a front-row seat to how CEOs are thinking about the future of AI and what enterprises should care about in the next 1–2 years. Here are my 3 biggest takeaways: 1️⃣ 𝗔𝗜 𝗶𝘀 𝘀𝗵𝗶𝗳𝘁𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝘀𝗶𝗻𝗴𝗹𝗲 𝗰𝗵𝗮𝘁𝘀 𝘁𝗼 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗹𝗼𝗼𝗽𝘀. I still remember the ChatGPT moment in 2023 from my Harvard dorm. I’d talk to a chatbot. One input. One output. In 2026, the single chat has shifted to AI loops. Every enterprise will run this loop: 𝗥𝘂𝗻 → 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗲 → 𝗜𝗺𝗽𝗿𝗼𝘃𝗲 → 𝗖𝘂𝗿𝗮𝘁𝗲 → 𝗢𝗯𝘀𝗲𝗿𝘃𝗲 The goal? Turn problems in production into tested improvements. When these loops happen at scale, agents will need specialized hardware such as CPUs for tool calls, searching the web, etc. 2️⃣ 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗮𝗻𝗱 𝗲𝗺𝗯𝗲𝗱𝗱𝗲𝗱 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝘄𝗶𝗹𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗰𝗿𝗶𝘁𝗶𝗰𝗮𝗹. Imagine 400,000 runs of your production agents overnight. How do you know what went wrong? What worked? What needs improvement? Companies that learn the maximum from every trace and continuously improve their systems will win. 𝗣𝗼𝘀𝘁-𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗺𝗼𝗮𝘁. 3️⃣ 𝗘𝘃𝗲𝗿𝘆 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝘄𝗶𝗹𝗹 𝗳𝗮𝗰𝗲 𝗮 𝘁𝗿𝗮𝗱𝗲-𝗼𝗳𝗳 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝗤𝘂𝗮𝗹𝗶𝘁𝘆, 𝗦𝗽𝗲𝗲𝗱, 𝗮𝗻𝗱 𝗖𝗼𝘀𝘁. “Could a cheaper model do a good enough job?” AI will run experiments, but humans will take the call between these trade-offs. 𝗛𝘂𝗺𝗮𝗻 𝗷𝘂𝗱𝗴𝗺𝗲𝗻𝘁 𝗶𝘀 𝗵𝗲𝗿𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆. ⚡ CoreWeave Cloud is solving for #1 with the NVIDIA Vera Rubin GPU and CPU architecture for agents. ⚡ CoreWeave Forge is the software stack tackling #2 and #3. Michael Intrator Jon Jones Jean English Brian Venturo Peter Salanki Chen Goldberg and the entire team are ensuring CoreWeave solves for AI production at scale. The next chapter of AI isn’t just about building smarter models. It’s about making them better every time they run. #sponsored #agents #production #scale

Our CEO Michael Intrator joined Fox Business Network to discuss why we're working with AdaniConneX to establish our first presence in India. We'll be deploying NVIDIA Vera Rubin, with plans to open a local office and hire a team on the ground. CoreWeave customers are building and scaling AI globally, and we need to be where they need us. India is an important market, with a deep base of technical talent, developers and enterprises putting AI to work. Demand for compute is global and so is CoreWeave. See the full interview below.

NVIDIA Vera CPU is coming to CoreWeave 🎉 Serving an agent and improving an agent are two different workloads. The second one runs on CPUs: sandboxed rollouts, code execution, evaluation, data pipelines. That half of the AI loop is growing fastest and getting the least attention. NVIDIA Vera is the first CPU built for it, and it's coming to CoreWeave. Rack-scale NVIDIA Vera CPU puts 128 CPUs in a single rack, creating capacity for over 11,000 concurrent environments. Through CoreWeave Sandboxes, those agents run hardware-isolated and sit directly alongside the training jobs they support. No stitching two halves together. The full AI loop, one platform. https://crwv.co/utcu6