CoreWeave

Variance is a tax on alpha. Here's where it actually comes from. Most quant teams trace reproducibility problems back to the model. The real source is usually one layer down: in the infrastructure coordinating the workload. When backtests, retraining jobs, and feature generation pipelines run in parallel on a hybrid stack, three things have to hold simultaneously for results to be trustworthy: ✅ Deterministic compute: same job, same cluster, same result  ✅ Unified data: no divergence between training inputs and inference inputs ✅ Verifiable execution: a complete audit trail from data ingestion to output General-purpose cloud wasn't designed to guarantee all three under sustained parallel pressure. It was designed to scale. Those are different problems. The architecture that actually addresses this isn't about adding more compute. It's about making the environment behave as one coordinated system rather than a collection of components that happen to be connected. CoreWeave and VAST Data wrote the blueprint, get it here. https://lnkd.in/gZy3CiQk

📣 #FullyConnected26 call for speakers is now open! We're not looking for polished decks and safe takes. We want the real work: what broke, what scaled, what surprised you, and what you'd do differently. We're building three tracks: ✔️ Build adaptive AI infrastructure ✔️ Push models further ✔️ Evaluate and monitor agents Customer stories, live demos, technical depth, not prototypes dressed up as progress. We want to hear from you. Submissions close June 26. Apply here: https://lnkd.in/gJaY-Jvf

What does AI training and the Monaco Grand Prix have in common? The fact that milliseconds matter. Congratulations to Fernando Alonso for securing Aston Martin F1 Team's first point of 2026 this past weekend. Next up: The first home race for our Double World Champion. See you in Barcelona 🏁

POV: Your agent passes every eval, hits production, and fails anyway. You're not shipping a bad model. You're shipping a model with no feedback mechanism. The superintelligence loop is a closed feedback cycle between training and inference, where every production interaction becomes signal for the next improvement. The goal: agents that don't just perform, but improve themselves automatically, without an ML researcher in the loop for every iteration. For years, teams trained offline, evaluated against labeled datasets, fixed what broke, and repeated, only shipping to production when metrics looked good enough. The problem? Labeled datasets can't cover every real-world scenario. Agents hit production and fail anyway. The loop changes that model. Now, we can build and evaluate a strong first version, and then let production data drive continuous improvement. Here are the four things that make it work: 👉 A self-improvement agent: the essential piece that closes the loop, automatically triggering retraining based on production signal 👉 Serverless RL: post-train for reliability without managing infrastructure, at up to 40% lower cost 👉 Production inference at scale: stable, observable behavior so every interaction generates usable signal 👉 Agent observability: Weights & Biases Weave surfaces failure modes and routes production data back into the next training run Agents that compound in capability. Systems that improve themselves. All in a day's work. https://utm.io/up4z9

2,000+ engineers and AI leaders are heading to San Francisco this September. Are you coming?  📅  September 29 – October 1  📍 Moscone South, San Francisco Accelerate the superintelligence loop, with: ♾️ 3 Days ♾️ 3 Tracks  ♾️ Hands-on labs & learning  ♾️ Live demos  ♾️ Keynotes, special guests and more This is where AI gets built, challenged, and proven, in real time. Registration is open. https://utm.io/upDdF #FullyConnected26

What a week 🎉 The first NVIDIA Vera Rubin NVL72 is running on CoreWeave. 72 Rubin GPUs. 36 Vera CPUs. 260 TB/s NVLink fabric. Up to 10× better inference per watt vs. Blackwell. Getting here required more than plugging in hardware. We built: 💙Valvey: software-defined liquid cooling with real-time flow, pressure, and leak detection per rack 💙Racky: unified rack control that manages power, cooling, and environment as a single cloud resource 💙1.6 Tb/s backend bandwidth per GPU across a multi-rail, multi-plane RoCE fabric Congratulations to the entire team who made this a reality! Make sure you tune in on June 30 as we unpack (literally!) this milestone, with SiliconANGLE & theCUBE 👉https://lnkd.in/erWUyyEB

Don't let your agent evals cost you your weekend...🫠 CoreWeave Sandboxes are now a first-class Harbor environment. We validated it on a full terminal-bench 2.0 run. Bring your rollouts. Get all the details on CoreWeave Sandboxes 👉https://lnkd.in/gpXvNc4s

This isn't bad engineering. It's what happens when infrastructure wasn't built for coordination at this scale. Here’s what breaks, why it compounds, and what resilient-by-design actually means. https://utm.io/upY9O Has this ever happened to you? Tell us in the comments ✍

NVIDIA Nemotron 3 Ultra is live on CoreWeave Serverless Inference. 🚀 In agentic AI, what matters is speed of task completion at a given accuracy. Long-running agents make thousands of model calls per task, so the model underneath decides whether a workflow finishes in seconds or stalls out halfway through. Nemotron 3 Ultra is built for exactly that. It's an open, frontier-reasoning MoE: 550B parameters with 55B active, a hybrid Transformer-Mamba architecture, and up to 1M tokens of context. The Mamba layers keep long-context inference efficient, which is what makes it practical for agents that plan, call tools, and reason over long trajectories instead of one-shot prompts. NVIDIA designed it for orchestration, coding agents, deep research, and enterprise automation. On Serverless Inference there's no infra to manage. No clusters to provision, no capacity to babysit. You hit an endpoint, it scales to the workload, and Nemotron 3 Ultra runs right alongside the frontier models you're already using. So you can route the right model to the right step instead of forcing one model to carry the whole agent. Open weights, frontier reasoning, built for long-running agents, ready to call today. Try it: https://lnkd.in/gunGdncw

Most AI training stalls aren't a compute problem. They're an infrastructure problem. Straggler nodes. Sync stalls. Failed runs that eat hours of forward progress. Capacity that's allocated but not actually doing useful work. Join Tara Madhyastha and William Markuske on June 23 for the first episode of Training Tuesdays, where we break down why distributed training throughput breaks down at scale, and what to look for before your next infrastructure decision. We'll cover: ➡️ Why AI training roadmaps stall even when teams have GPUs, budget, and models ready ➡️ Which signals reveal whether infrastructure can sustain model progress, and why synthetic benchmarks miss the failure modes that matter at scale ➡️ How CoreWeave ARENA helps teams validate real workloads before making a broader infrastructure decision 🗓️ June 23, 2026 ⏱️30 minutes 📍https://utm.io/up1AN Bring your questions, its training time.👇