CoreWeave Fully Connected 2026 Keynote: Introducing CoreWeave Forge | Built for Adoption
CoreWeave's Fully Connected 2026 Day 1 keynote centers on the AI loop of run, observe, curate, improve and evaluate, and the launch of CoreWeave Forge. It covers production agent operations, agent observability and RL rollouts, Vera Rubin NVL72 and the Vera CPU, and enterprise evaluation practices from Capital One, Caterpillar and Cognition.
The keynote opens with a framing argument that the real goal of AI infrastructure is to shrink the distance between human curiosity and discovery, using the image of a treasure hunter following clues across languages and centuries. CoreWeave is positioned as a “force multiplier” for pioneers, and the event is framed around two days: “built for adoption” today and “built for ambition” tomorrow.
The engineering starting line
Hen Goldberg, EVP Product and Engineering, argues that deploying an agent is only step one. Agentic workloads multiply token consumption, context length, latency and cost, and when an outcome is wrong, teams need to know whether the fault lies in data, retrieval, or tool use. She draws a parallel to infrastructure-as-code and agile development: the industry already learned how to make infrastructure and applications repeatable, and now model development, application development and infrastructure operations must be unified as a single system.
The origin of the approach is a concrete bottleneck. CoreWeave built a speculative decoding workflow, where a small model drafts tokens and a larger model verifies them, and reached the top of an output-speed comparison for Kimi K2. Achieving this required manually stitching together training, evaluation and serving, with one engineer acting as the integration layer. That motivated the “AI loop”: run, observe, curate, improve, evaluate, and run again. Forge is presented as the connected environment for that loop, but Goldberg stresses that “connected does not mean closed” and that users remain free to choose their own models, frameworks and clouds.
Running, and Mission Control in practice
The running stage is demonstrated with serverless and dedicated inference for GLM models. A live demo shows Mission Control, an agent that investigates a degraded deployment. Throughput is improved from roughly 3,000 to 5,000 tokens per minute per GPU by addressing KV cache thrashing caused by long-context traffic that evicted blocks and forced recomputation, and then from 5,000 to 20,000 by moving to two replicas on the same GPU budget using knowledge of NVLink domains. The claimed result is a doubling of price-per-token efficiency at unchanged quality. Mission Control is described as processing more than 200 million metric samples per second across the fleet, correlating workload behavior with physical rack state.
On hardware, Vera Rubin NVL72 is announced as ready for production, with Cognition as the first customer. Goldberg emphasizes composability: the stack from power and cooling to firmware and networking must evolve without locking the cloud to one model or runtime. CoreWeave currently operates six generations of NVIDIA accelerators under a consistent operational model.
NVIDIA’s perspective on agentic compute
Ian Buck, NVIDIA’s VP of hyperscale and HPC, reframes the workload. The chat era was a simple human-in-the-loop prompt and response. Agents run an inner loop in which language models reason, call tools (CPU-bound tasks such as compilers, databases and Python, as well as GPU simulations and CUDA applications), evaluate results, and repeat, consuming tokens at machine speed. He calls this a data-center-scale problem and argues that Vera Rubin is a full-stack AI factory platform spanning GPUs, networking, storage and CPUs.
The Vera CPU is positioned as a dedicated agent CPU that combines high single-threaded performance with core density. Cited figures include doubled performance per core relative to the prior generation, roughly 40% lower memory latency, sandbox startup 3x faster than traditional CPUs, 1.7x faster on the Terminal Bench agent benchmark against the latest x86 CPUs, and 7x faster data processing on ClickBench. Buck also argues that CUDA generations remain productive years later, with Hopper handling fine-tuning and Ampere serving small-model inference.
Observe, curate, improve, evaluate
Corey Sanders, SVP of product, presents the second half of the loop. He opens with practitioner testimony about trust boundaries with MCP providers, the difficulty of getting variance in distillation data, the risk of breaking real robots when testing policies in the physical world, and the subtlety of agent misbehavior that surfaces only after many tokens.
The new products map onto the loop. CoreWeave Agent Lens handles observation and curation: it clusters failures into patterns and tracks traces, feedback and context, so a team reviews the cases that deserve attention rather than thousands of conversations. Model distillation uses curated production traces to produce a smaller specialized model. RL Rollouts addresses a systems bottleneck in reinforcement learning: rather than redeploying serving for every checkpoint, weights are loaded into the running deployment while the trainer keeps requesting rollouts.
Maulin Patel of Capital One provides the enterprise evaluation perspective. Because the bank serves over 100 million customers, evaluation must be stronger than simple heuristics. Capital One moved beyond BLEU and ROUGE toward LLM-as-judge with chain-of-thought rubrics and RAG metrics such as faithfulness. For multi-turn agents, the failure modes he identifies are context drift, behavioral traps (such as claiming a fund transfer without calling the API), circular loops, and compounding cost and latency across trajectories. Their response includes task-completion metrics that verify actual database state changes, hybrid graders combining deterministic code checks with LLM judges, expert calibration of automated judges, and regression gates in CI/CD. He summarizes three takeaways: evaluation is the real competitive moat because models commoditize; non-determinism must be tamed with layered grading; and embedded governance accelerates rather than slows developers. Capital One open-sourced its WoolHunter vulnerability scanner under Apache 2.0, which uses attacker-first analysis and only flags proven-exploitable bugs to reduce false positives.
The live demo
Sanders walks through an IT support ticketing app that misroutes and mis-prioritizes requests. He connects it to Forge, uses an AI research assistant named Arya to generate observability code via a GitHub diff, and then uses Agent Lens to surface an insight that about 1.6% of conversations had incorrect policy priority assignment. Tags are applied both automatically and by hand. Arya produces a performance report and recommends serverless RL as the post-training strategy. The RL runs execute in CoreWeave sandboxes, and the best checkpoint is compared against the original in an evaluation view. Quality improves, with latency increasing by roughly a second, which he accepts as a deliberate tradeoff. The model lineage is then traced and the new artifact is redeployed to the run stage, closing the loop.
Cognition, the partner network, and the business
Steven Alberti of Cognition reports that Vera Rubin delivered a 4.8x throughput improvement over GB200 on production inference at equal speed and 3.8x in RL rollouts. He also describes SWE-2, a model that approaches frontier coding performance at roughly 60% lower cost, and discusses a Pareto-frontier training approach where additional compute pushes the cost-performance curve toward cheaper and smarter models.
The partnership announcements include CoreWeave Arena for validating workloads on production infrastructure, direct access to CoreWeave engineers, open-source commitments spanning Kubernetes, CNCF, vLLM and Marimo, and a Partner Network connecting security, monitoring, data and orchestration tools, with Exa, Parallel Web Systems and You.com added for search.
In the closing conversation, CEO Mike Intrator describes CoreWeave’s origin in serving frontier labs with a blank sheet and no legacy business, which built discipline around training and serving very large models. He frames the company’s role as removing barriers between an idea and an action, and emphasizes that the market is moving faster than anyone predicted. Brian Venturo closes by returning to the treasure-hunter metaphor and reiterating the message: try it, see what is real, and make the next one better.