CoreWeave Closes the AI Loop: Forge Turns GPU Cloud Into a Full-Stack AI Platform

Introduction

For most of its corporate life, CoreWeave sold a fairly narrow product: high-end Nvidia GPU compute, delivered fast, to companies that could not get it anywhere else. That was a good business, and it made the company one of the most closely watched infrastructure plays in the market. But it was also a business with a ceiling, because renting raw acceleration is a commodity in exactly one sense of the word — every AI cloud can eventually buy the same chips, and the premium disappears once supply catches up with demand.

At its Fully Connected conference in San Francisco on September 30, 2026, CoreWeave made its answer explicit. The company announced CoreWeave Forge, a unified development platform that pulls together training, inference, evaluation, observability, and agent development into what the company calls a single connected environment. The pitch, in the words of executive vice president of product and engineering Chen Goldberg, is that everything a business learns from running AI in production has to feed back into how engineers improve it — and that this has to work as one system rather than five vendors' tools stapled together.

More than 4,500 customers, partners, developers, and AI leaders attended the event, and the announcement is the most substantial repositioning CoreWeave has attempted since it began selling GPU time. It is also a direct answer to a question the market has been asking since the 2024 funding wave: what does an AI-native cloud look like when the hardware stops being the differentiator?

What Forge Actually Is

Forge is not a new model and not a new cluster architecture. It is a control plane for the iteration cycle that sits above the compute CoreWeave already rents out. The company's own framing is a loop — run, observe, curate, improve, evaluate, and repeat — and the specific components map onto each stage of that loop.

The most substantial piece is inherited rather than original. CoreWeave acquired Weights & Biases in 2025, and Weights & Biases Models now provides the experiment-tracking layer, handling the comparison of tens of thousands of runs and millions of metrics. Post-training expertise came from OpenPipe, including serverless reinforcement learning, serverless supervised fine-tuning, and model distillation. Notebooks are built on the open-source marimo project, which means a prototype written in a familiar reactive notebook environment carries through into training, evaluation, and production instead of being rewritten by hand.

Two components are new or newly generalized. CoreWeave Agent Lens performs observability on production agents, analyzing tens of millions of traces and converting them into suggested fixes. CoreWeave claims that on an internal benchmark against a general-purpose frontier LLM, Agent Lens detects 20 percent more critical failures at one-tenth the cost — a number worth treating as vendor-supplied rather than independently verified. CoreWeave ARIA, previously in preview, is now generally available as a coding agent that reads large-scale experiment and observability data, works out what actually drove a change, and proposes the next experiment worth running. Sandboxes became generally available as well, providing an isolated environment for each run whether the workload is agent tool use, reinforcement learning, or evaluation.

The structure that matters most is the registry. Every model checkpoint and agent configuration is versioned in portable formats with explicit lineage, which is what makes the loop defensible rather than merely aspirational: without versioned lineage, a "fix" suggested by an observability tool cannot be traced back to the exact production state that motivated it.

Forge ships in three tiers — Free, Pro, and Enterprise. CoreWeave is also standing up a partner network providing validated integrations, with VAST Data, CrowdStrike, and ClickHouse named among the early participants.

The Hardware Bet Continues

The temptation when a hardware company pivots to software is to assume the hardware push has slowed. CoreWeave is doing the opposite. At the same event the company said Nvidia's Vera CPU, purpose-built for AI agents, is in early access with broader testing to follow, and that the Nvidia Vera Rubin NVL72 is now available on the platform.

The first customer on Vera Rubin NVL72 was AI software firm Cognition, which reportedly had workloads running within two days — a turnaround figure that says as much about the software layer as the silicon. CoreWeave also continues to cite record MLPerf inference and training results and its status as the only AI cloud to hold SemiAnalysis's top Platinum ClusterMAX ranking three consecutive times.

This combination is the strategic point. Analyst Dave McCarthy of IDC framed the announcements as CoreWeave trying to separate itself from other specialist AI clouds such as Nebius, Lambda, and Vultr while competing with AWS, Microsoft Azure, and Google Cloud for workloads. His read was that CoreWeave has accepted that landing the large AI labs is not sufficient on its own: "If they're going to build a long-term sustainable business, then they need to be able to appeal to a wider audience, including enterprises," which in his view requires "more capabilities — things that feel more turnkey and less like a DIY project." He also noted the company is not retreating on infrastructure, continuing to invest in Nvidia hardware rather than letting the hardware slide because a software layer now exists.

That is the pattern the wider cloud and edge computing market has been circling for two years. The hyperscalers already bundle model services, observability, and managed training with compute, and the specialists have been left competing on price and delivery speed alone. Forge is an attempt to close that gap without abandoning the one asset the specialists genuinely own.

Why the Loop Is Hard

There is a reason this has not been built before at this scale, and it is worth being precise about it. Nick Patience, vice president and practice lead for AI platforms at the Futurum Group, put the framing neatly: every platform decision in this market involves weighing how much optionality a team is willing to surrender for a connected toolchain, and Forge is built on the premise that teams should not have to make that calculation at all.

Closing the loop from production back into training inside a single environment, while remaining open across models, frameworks, and clouds, is a genuinely difficult engineering problem — considerably harder than either half alone. The failure mode of most enterprise AI programs is not that the model is bad. It is that the organization never closes the distance between what production does and what the next training run learns from it. A production trace does not feed the next training run. A finished experiment does not inform the next evaluation. Every handoff across a tool boundary loses a signal or costs time. The cumulative effect is that improvement happens far more slowly than the raw compute growth would suggest it should.

Forge's claim is that keeping training runs, experiment tracking, evaluations, and agent traces co-located — where the model or agent actually runs — is what changes that arithmetic. CoreWeave is also betting on inference-side reinforcement learning, with RL Rollouts in preview under Dedicated Inference, which hot-loads checkpoints into a live deployment so a training loop can keep running without redeploying. If that works, the interval between "we found a problem in production" and "the fix is serving traffic" collapses from weeks toward hours.

Whether an enterprise will actually trust a platform vendor to sit in that position is the open question. Centralizing observability and checkpoints means centralizing knowledge about how a company's AI systems behave. CoreWeave's answer is that everything stays in open, portable formats with clear lineage, and that the platform remains usable with any model, framework, or cloud — a claim that is easy to make and harder to verify from outside. MasterClass and Canva are cited as already building on Forge, which at least establishes that the reference customers are real enterprises rather than the AI labs CoreWeave has always served.

What It Means for Buyers

For customers, the practical consequence is that evaluation and observability stop being bolt-ons. Choosing an AI cloud has historically meant choosing a vendor for training, a second for evaluation tooling, a third for tracing, and a fourth when an agent misbehaves in production. Forge reduces that to one procurement decision with three tiers, which is precisely the kind of simplification that moves a platform from a specialist's tool to an enterprise default.

The competitive pressure runs in two directions. Hyperscalers have to keep matching software depth while competing against specialists who can be more focused. Other specialist clouds have to answer a platform announcement with either their own software layer or a clearer reason enterprises should not need one. And the hyperscalers' own advantage — bundling — is the thing Forge is deliberately imitating.

One detail is worth watching because it sets the tone: RL Rollouts hot-loading checkpoints into live deployments means an improvement can reach production traffic without a redeployment cycle. If that pattern becomes normal, the boundary between "model development" and "running the service" dissolves further, and the operational questions — who can roll back, who owns an incident, what was actually deployed — become cloud problems rather than research problems. That is a genuine shift, and it will be felt long after this particular launch fades from the conference circuit.

Conclusion

CoreWeave has built its business on a scarcity thesis: AI labs needed GPUs, suppliers could not deliver them fast enough, and CoreWeave could. Forge is the company's admission that the scarcity era is ending and the differentiation has to move somewhere else. It has moved to the software that connects production behavior back to training.

The execution bar is high. Closing the loop across training, inference, evaluation, and observability while staying genuinely open across models, frameworks, and clouds is harder than any single piece of the announcement suggested, and analysts are right to frame the enterprise push as the real test rather than the launch itself. But the direction is now clear across the AI cloud market, not just at CoreWeave: the vendors that sell acceleration will increasingly be judged on how much of the iteration loop they can actually own.

Images

Server racks in a large-scale computing facility. Wikimedia Commons, Rack of Worldwide LHC Computing Grid — CC BY-SA 4.0. Illustrative of the rack-scale infrastructure such platforms run on; not a CoreWeave facility.

A 2.5D accelerator package with the GPU die, stacked HBM memory, and silicon interposer exposed on the package substrate. Wikimedia Commons, AMD Fiji GPU package — CC BY-SA 4.0. Illustrative of the accelerator packaging CoreWeave resells; not a current CoreWeave deployment.

An aisle of enclosed server cabinets on raised-floor tiles in a high-performance computing facility. Wikimedia Commons, Front of server racks at NERSC — CC BY-SA 4.0. Illustrative of managed compute capacity; not a CoreWeave installation.

References

• CoreWeave Forge Connects the Full AI Development Loop — CoreWeave press release, September 30, 2026.
• CoreWeave Targets Enterprises with Forge Platform — Data Center Knowledge (Informa TechTarget), Wylie Wong, September 30, 2026.
• CoreWeave Expands Enterprise Push As AI Cloud Competition Heats Up — Investor's Business Daily.
• CoreWeave Forge: Turn AI Iteration Into Compounding Improvement — CoreWeave blog.
• CoreWeave Forge product page — feature and edition detail.