CoreWeave

3 days. 3 tracks. 1 incredible conference. The 3 tracks at #FullyConnected26 are built for you: 1️⃣ Build adaptive AI infrastructure: compute, networking, storage, and the platform tools that turn raw infra into systems teams actually rely on 2️⃣ Push models further: distributed training, RL, experiment tracking, and the workflows that make every iteration faster 3️⃣ Evaluate and monitor agents: end-to-end observability, failure detection, and the feedback loop that makes your agent better every time This is where the AI community compares notes on what's actually working, and where the next wave of agent infrastructure gets built. Which one are you most excited for? Drop a 1️⃣, 2️⃣ or 3️⃣ in the comments, and we'll send you a discount code to register. 🗓️ September 29–October 1 📌 Moscone South, San Francisco 🔗 Register now: https://www.utm.io/upDdF

CoreWeave Sandboxes give your agents a room it can't burn down 🧯 You can now run Claude inside a CoreWeave Sandbox and drive it from the Claude app on your phone. Kick off a long task and put it in your pocket. The agent gets a custom container image with your tools and MCPs, plus Slurm access. You set the network policy, down to no internet at all. Our favorite recent demo of this is from Deok Filho, who ran Claude inside one and built a 2048 game from his phone. The agent did the work. The sandbox made it safe to let it. Try it for yourself, we built a marimo notebook here: https://utm.io/usdXg

"We were totally scrappy...and we still are." That's how our founding product marketing manager, Emma Ramos, describes her CoreWeave journey. Fast forward and layer on bigger customers, more products, and a platform that's evolved to enable bigger innovations...the team building the Essential Cloud for AI is growing. Because growth doesn't have to mean losing what made you move fast in the first place. Check out our open roles, we're hiring. https://www.utm.io/urA13

We shipped two firsts last month, and they tell one story. First came the numbers 📝 We published the first measured silicon results for NVIDIA Vera Rubin NVL72: at similar interactivity, 10x more tokens per second per megawatt than Blackwell, with NVIDIA TensorRT-LLM and NVIDIA Dynamo optimizations enabled. Those are REAL results from live hardware, not projections. Then came the fabric underneath them. We deployed the NVIDIA Spectrum-X SN6600-LD, the industry's first fully liquid-cooled Ethernet switch, and packed 1.64 Pb/s of fully non-blocking capacity into a single rack. Those aren't two separate wins. Reasoning models push enormous token traffic across the network, and every watt the fabric saves is a watt the GPUs get back. You only see numbers like these when compute, networking, and cooling are engineered as one system, which is exactly how we build. More info: https://t.co/qwQHD9cSn2

A cluster can pass every benchmark and still lose real hours to an interruption or a checkpoint that silently did not save. That is where Part 2 of our production AI factory series starts. Dion Harris and I look at what happens after validation: goodput, health signals, scheduling control, recovery, and carrying the same discipline into the next generation.  On long-running distributed training workloads, CoreWeave has demonstrated up to 96% goodput on NVIDIA Hopper GPUs. Mission Control processes over 200 million metrics samples per second across all customer environments. SUNK dynamically reallocates GPUs across training and inference workloads on the same cluster as demand shifts. In June 2026, #CoreWeave completed the industry-first bring-up and validation of #NVIDIA Vera Rubin NVL72. The first measured results showed up to 10x more tokens per second per megawatt than NVIDIA GB200 NVL72 on the same DeepSeek R1 workload at a matched interactivity target. The test used large-scale expert parallelism, NVFP4 precision, multi-token prediction, and disaggregated prefill and decode with NVIDIA TensorRT-LLM and NVIDIA Dynamo. Capacity is the starting point, its the lifecycle decides whether the factory delivers. Read part2 here https://lnkd.in/gZewUg36

Customers shouldn't have to debug a new architecture while in production. Ever. In June 2026, we completed the industry-first bring-up and validation of NVIDIA Vera Rubin NVL72. NVIDIA Vera Rubin unifies six purpose-built chips into one supercomputer and requires rack-scale operations built for a 100% liquid-cooled system. That is why readiness reaches beyond silicon, into power, cooling, sensors, scheduling and recovery. We built Valvey, Racky and the Rack LifeCycle Controller to manage those systems rack by rack. The first measured result produced up to 10x more tokens per second per megawatt than NVIDIA GB200 NVL72 on the same DeepSeek R1 workload at a matched interactivity target. Part two of the story shows how the discipline behind day-two operations carries into the next generation of AI factories: https://www.utm.io/usbFb

MasterClass 🤝 CoreWeave MasterClass is supercharging their AI teaching agents on CoreWeave Cloud. MasterClass leverages W&B Weave to take an in-depth evaluation of their teaching agents across conversations with learners to take performance to the next level across all workflows. This development comes after this morning’s exciting news that Weights & Biases just hit 1 billion runs, a defining milestone for our team and a testament to the longstanding trust of customers throughout the history of Weights and Biases. Read more about the announcement here: https://www.utm.io/usdxl

We just crossed 1,000,000,000 runs 🐝 Every run starts with a question: Can we make this model better? The AI community just asked it for the 1,000,000,000th time on Weights & Biases. Cohere asked it while training business-critical enterprise LLMs. Decart asked it while building Oasis 3, a world model for physical AI. Pinterest asks it hundreds of thousands of times a month across its training jobs. The year since Weights & Biases joined CoreWeave has been about answering faster. W&B Weave was rebuilt for production agents. We debuted CoreWeave ARIA, our AI Research and Iteration Agent for reading experiments and proposing the next run on its own. Serverless Inference, RL, and SFT handling the compute underneath. Loss curves on your phone via the W&B iOS app, runs in your terminal through W&B Leet and an MCP server for your coding agent. AI development works in a loop, and closing that loop takes tools and infrastructure built as one system. A billion runs taught us exactly what that system needs. The next billion will go faster. Congratulations to the community writing them. https://www.utm.io/usc7W

You've seen the signs... Now see what's behind it: 👉 Red-team a production AI agent, then vote to approve, change, or block it 👉 Crack open a real training run on SUNK and see exactly where every second goes: compute, communication, I/O 👉 Hear directly from leaders like Capital One and Caterpillar Inc. on what actually breaks, and holds, running AI in production #FullyConnected26 kicks off September 29 at Moscone South. Registration is open. Don't miss it. 🗓️ September 29–October 1 📌 Moscone South, San Francisco 🔗 Register now: https://bit.ly/4y19P56