Summary

• CoreWeave Forge combines model serving, production observability, data curation, post-training, evaluation, and agent-development tools within a single environment.
• The platform builds on CoreWeave’s acquisition of Weights & Biases and represents a move beyond selling specialised GPU infrastructure towards owning more of the AI development workflow.
• CoreWeave says Forge remains open across models, frameworks, other clouds, and private infrastructure, although tighter integration also gives customers more reasons to remain inside its platform.

Specialist AI cloud providers are moving beyond the business of renting expensive accelerators, with CoreWeave launching a development platform that reaches from production inference and observability into evaluation, data curation, post-training, and the repeated work of improving AI agents after they have been deployed.

CoreWeave has launched Forge, a software layer intended to connect the stages through which models and agents are run, observed, improved, tested, and redeployed. MasterClass and Canva are already using the platform, according to the company.

The launch moves CoreWeave further from its origins as a specialised provider of GPU computing infrastructure. Rather than competing only on accelerator availability, network performance, or the speed at which new Nvidia systems can be deployed, the company is trying to own more of the workflow developers use before and after those processors do their work.

Forge brings together inference, agent observability, notebooks, model evaluation, training tools, sandboxes, model registry functions, and technology inherited from Weights & Biases, which CoreWeave acquired in 2025. The company describes the resulting workflow as a loop in which production behaviour creates the data used to identify failures, improve a model or agent, evaluate the new version, and deploy it again.

The AI cloud market is moving up the stack

CoreWeave built its position in the AI market around infrastructure optimised for graphics processors and other accelerated computing hardware, challenging general-purpose hyperscale clouds with environments designed more narrowly around training and inference. That distinction becomes less defensible if competitors can all secure similar accelerators and customers then assemble the software above them elsewhere.

Forge addresses that problem by making the developer layer part of the cloud proposition. A customer can serve models through CoreWeave, collect traces from production, turn selected failures into datasets, fine-tune or otherwise modify a system, run evaluations against previous versions, and record the resulting model lineage without moving repeatedly between unrelated products.

There is a genuine operational problem underneath the platform packaging. Production AI applications generate information that research teams need if they are to understand where models fail, yet observability, evaluation, training, application development, and infrastructure are often owned by different teams and handled through separate tools.

Agents make those hand-offs more awkward because a single user request can trigger several models, external tools, retrieval systems, and software actions. Understanding why an agent failed can therefore require tracing what it saw, which tool it selected, what each model returned, how much the workflow cost, and whether the final behaviour met whatever quality or safety standard an organisation had set.

Weights & Biases becomes part of the cloud strategy

CoreWeave’s 2025 acquisition of Weights & Biases provided much of the software foundation for that move. The acquired platform already handled model-development functions including experiment tracking, evaluation, monitoring, and model management, giving CoreWeave a route into developer workflows that would have been difficult to build purely through infrastructure products.

Forge now combines those capabilities with CoreWeave’s inference and training services. Its product set includes Agent Lens for tracing and analysing agent behaviour, ARIA for assisting with experiments and code changes, isolated execution sandboxes, managed notebooks, model distillation, serverless and dedicated inference, post-training services, and a registry for recording datasets, checkpoints, metrics, and deployments.

The breadth is strategically more important than any individual feature. Hyperscalers such as Amazon Web Services, Microsoft Azure, and Google Cloud already combine infrastructure with extensive software platforms, while model providers and developer-tool companies are expanding from the opposite direction. Specialist AI clouds consequently face pressure to provide enough of the surrounding environment that customers do not treat them merely as interchangeable GPU capacity.

CoreWeave is following that route while continuing to invest heavily in the physical layer. Its current infrastructure spans North America and Europe, and the company said in June that a Swedish expansion brought its European estate to eight sites. It also operates facilities in the UK, where previously announced investment commitments reached £2.5 billion.

Openness becomes part of the competitive pitch

A more integrated platform creates an obvious tension around lock-in, which CoreWeave is trying to address directly. The company says Forge supports different models and frameworks and can work with workloads running on other clouds or private data centres rather than requiring every stage to sit inside CoreWeave infrastructure.

It also says the datasets and evaluation suites generated through the platform remain portable and accessible through open interfaces. That is commercially important for organisations reluctant to place their entire AI development process inside another proprietary environment after spending years trying to reduce dependence on individual cloud providers.

Portability claims still need to be tested against actual deployment. A platform may expose open interfaces while becoming difficult to replace because teams depend on its workflow, metadata, monitoring, identity controls, integrations, or operational knowledge. The more successful Forge becomes at connecting those stages, the more consequential migration away from it could become.

CoreWeave itself acknowledges the commercial preference beneath the openness argument, saying Forge performs best on its own cloud even while supporting infrastructure elsewhere. The company therefore benefits if customers begin with portable tooling but gradually run more inference, training, and supporting services on CoreWeave infrastructure.

Production AI shifts competition away from raw compute

The timing reflects a broader change in AI spending. Training large models created enormous demand for clusters of accelerators, but enterprise adoption depends increasingly on the repeated cost and operational complexity of running models in production, evaluating their behaviour, keeping agents reliable, and making improvements without breaking existing applications.

Inference can also create a much longer commercial relationship than a single training run. Applications may call models continuously for years, while production data feeds new rounds of testing and improvement. A provider that controls several stages in that cycle can participate in more of the customer’s spending than one supplying processors alone.

European customers are part of that contest because CoreWeave has spent heavily building physical capacity in the region, including UK data centres and newer Scandinavian infrastructure. Forge gives the company a software layer that can be sold alongside those facilities while also reaching customers whose workloads remain distributed across several clouds and private environments.

The competitive test will be whether integration saves enough engineering effort to outweigh the attractions of assembling tools independently. Large AI teams often favour specialised products precisely because they can swap individual components without changing an entire platform, while smaller enterprise teams may prefer fewer vendors and less infrastructure work even if that creates a deeper dependency.

CoreWeave’s launch therefore marks a shift in what an AI cloud wants to be. Access to scarce processors built the first generation of specialist providers, but accelerators alone are becoming the foundation rather than the complete product. Forge is an attempt to turn CoreWeave from somewhere an organisation runs AI into somewhere it repeatedly builds, measures, fixes, and operates it — a considerably more ambitious position in a market already occupied by the largest cloud companies in the world.