CoreWeave Ships Forge — GPU Compute Landlord Becomes Agent Platform Provider

CoreWeave launched CoreWeave Forge at Fully Connected 2026 in San Francisco, a unified development layer that consolidates Weights & Biases, OpenPipe, and Marimo with new services including Model Distillation, Notebooks, and the ARIA research agent, now generally available. The platform adds Agent Lens, which CoreWeave says improves failure detection by 20 percent and resolves issues at roughly 10x lower cost than a general-purpose frontier LLM benchmark, plus generally available Sandboxes with Kata VM-based isolation billed by the millisecond and a Registry for models, agents, and datasets in open, portable formats. DigitalOcean shipped Agent Droplets the following day, October 1, in public preview at $50 per month with Firecracker microVM isolation and governed access to 16,000 tools, as compute providers bundle agent execution, tool access, and governance.

GPU compute landlords are becoming agent platform providers. CoreWeave signaled the move at Fully Connected 2026 in San Francisco with the launch of CoreWeave Forge https://www.coreweave.com/products/coreweave-forge , a unified development layer that connects the full AI improvement lifecycle – Run, Observe, Curate, Improve, Evaluate – on the same infrastructure that serves the GPUs themselves. The platform consolidates several previously separate products – Weights & Biases, OpenPipe, and Marimo – alongside new services including Model Distillation, Notebooks, and ARIA, a research agent now generally available. But the structurally significant pieces are the ones that move CoreWeave from compute supplier into the agent execution and observability layer: Agent Lens, Sandboxes, and Registry. Agent Lens https://www.coreweave.com/products/coreweave-forge , a new service, traces every step, decision, and tool call a production agent makes, end to end. Monitors score live traffic against baselines with human oversight, surfacing failures before they become patterns. CoreWeave reports that this approach improves failure detection by 20 percent and resolves issues at roughly 10x lower cost compared to a general-purpose frontier LLM benchmark. CoreWeave Sandboxes https://www.coreweave.com/products/coreweave-forge , now generally available, provide isolated CPU and GPU execution environments for agent tool use, reinforcement learning, and evaluation. The serverless variant uses Kata VM-based strong isolation and bills by the millisecond. The on-cluster variant runs inside a customer’s existing CoreWeave Kubernetes Service with no incremental cost. The CoreWeave Registry serves as a system of record for models, agents, and datasets in open, portable formats with full lineage tracking. Together, these three services represent the GPU landlord expanding into the layer where agents actually execute, get observed, and get versioned. The move mirrors what DigitalOcean shipped the following day. DigitalOcean Agent Droplets https://forkast.news/digitalocean-turns-agent-infrastructure-into-a-commodity-sku/ launched October 1 in public preview at $50 per month, offering Firecracker microVM isolation with governed access to 16,000 tools. Same architectural pattern – compute provider bundling agent execution, tool access, and governance – but from the general-purpose cloud layer rather than the GPU layer. That distinction matters. GPU access remains the binding constraint for advanced AI workloads. When the entity controlling the GPU supply also ships the agent sandbox, observability, and asset registry, the compute-supply-chain thesis extends into the agent infrastructure layer. The switching costs are structural, not contractual: the most performant agent workflows will naturally run where the closest GPU sits. CoreWeave’s positioning attempts to manage this tension. Forge is described as cloud-agnostic – working with any model, framework, or cloud – and the Registry stores assets in open, portable formats. But the planned integration of NVIDIA Vera CPUs into Sandboxes, which CoreWeave claims will deliver 3x faster agent sandbox startup versus x86, creates a performance incentive to stay within the CoreWeave ecosystem. Efficiency-driven lock-in operates without explicit contractual barriers. The Forge launch connects to patterns Forkast has been tracking across the agent infrastructure stack. The harness pattern https://forkast.news/the-harness-pattern/ describes how infrastructure providers bundle orchestration layers with compute. The commodity SKU thesis https://forkast.news/agent-infrastructure-is-becoming-a-commodity-sku-and-the-moat-is-moving-upstream/ predicts agent infrastructure becoming standardized, purchaseable units. The execution-layer gateway https://forkast.news/the-execution-layer-gateway-is-where-enterprise-ai-security-actually-lives/ shows where security and access control actually live in the stack. And Kubernetes agent sandbox https://forkast.news/kubernetes-agent-sandbox-is-becoming-the-unit-of-agent-deployment/ is becoming the standard execution substrate. Forge touches each of these layers from the GPU landlord position. Pricing starts at $0 for individual builders, $60 per month for teams under 50 employees, and custom enterprise pricing. A 30-day Pro trial is available. Early customers include MasterClass and Canva. As GPU landlords continue building out agent platforms, the competitive landscape will be defined by who best bridges raw compute and the iterative requirements of production-grade agent deployment.