CoreWeave Forge Wants to Own the AI Loop, Not Just the GPUs
CoreWeave spent the last two years becoming the most famous renter of graphics chips on Wall Street. On Tuesday, it told the industry it wants to be something bigger: the company that owns the software every AI team touches from the first training run to the thousandth production query.
The announcement came at Fully Connected, CoreWeave's AI cloud conference in San Francisco on September 30, 2026, which the company said drew more than 4,500 customers, partners, developers, and AI leaders. The product is called CoreWeave Forge — a "development layer" for building and improving AI models and agents — and it is the clearest signal yet that the neocloud wars are shifting from who has the most GPUs to who can make those GPUs easiest to turn into working AI.
Update, Oct. 1: Wall Street's split verdict
Update, Oct. 1: Wall Street is split on whether the software story justifies the stock. Bulls: Truist kept a Buy with a $165 target (~90% upside); UBS started coverage with a Buy; JPMorgan upgraded after the pullback; consensus is a 'Moderate Buy' with a ~$136 mean target (Barchart/MarketBeat). Bears: Rothschild & Redburn initiated with a Sell and $54 target, arguing 'credit markets are beginning to price risks that equities largely ignore' — falling GPU rental prices, hyperscalers building their own hardware, and financing costs eating returns on the debt-heavy model. The leasing-economics critique is the real counter to Forge: if GPU rentals keep cheapening, owning the software loop matters more, not less — and that's exactly CoreWeave's bet.
What Forge actually is: the AI loop, in one place
Forge's pitch is built around what CoreWeave calls the "AI loop": run, observe, curate, improve, evaluate, repeat. The idea is that a team training or deploying a model should be able to push a version into production, watch how it behaves, pull the interesting signals back out, retrain or fine-tune, evaluate the new version, and push again — all inside one connected environment, with nothing lost between steps.
That sounds obvious, but in practice the AI development stack is a patchwork. A production trace captured by one vendor's tool never makes it back into another vendor's training run. An experiment logged on Monday doesn't inform the evaluation run on Friday. Every handoff between tools from separate vendors costs time and drops signal. Forge is CoreWeave's argument that closing that loop is worth more than any single piece of it — and that whoever closes it wins the long-term relationship with AI teams.
Crucially, CoreWeave is not trying to lock anyone in to do it. The company says Forge stays open across the models, frameworks, and other clouds teams already use. That is a deliberate design choice, and a smart one: developers have spent the last decade punishing every platform that tried to fence them in.
Inside the stack: Weights & Biases, OpenPipe, marimo, and a version registry
Forge is less a brand-new invention than a unification. The company says it brings together Weights & Biases Models — the experiment-tracking tooling CoreWeave owns — with post-training expertise from OpenPipe and the open-source marimo notebook project, alongside CoreWeave's own services, in one environment built for continuous improvement.
That combination tells you exactly where CoreWeave thinks the value has moved. Weights & Biases was already the default system of record for training runs at serious AI labs. OpenPipe specializes in the unglamorous but decisive work of post-training — fine-tuning models on real task data so they actually perform in production. Marimo brings reactive, reproducible notebooks into the picture. Stitched together, they cover the full arc: notebook exploration, training runs, experiment tracking, evaluations, and agent traces, all living where the model or agent actually runs.
And then there is CoreWeave Registry, which the company says keeps every checkpoint and agent configuration versioned. If that works as advertised, it is the boring, indispensable part of the whole package — the reason a team can roll back a bad model the way software teams roll back a bad deploy. Version control did for code what Forge is trying to do for models, and the company that supplies it tends to become the default.
The strategy shift, in the company's own words
The most honest sentence of the day came from Corey Sanders, CoreWeave's senior vice president of product management, at a media briefing during the event:
"We have moved from a focus on being the best provider for AI-centric infrastructure to delivering a plethora of AI services to enable customers to build AI applications on our platform."
Read that twice. For two years, CoreWeave's entire identity was speed: get Nvidia's newest chips racked, cooled, and rented faster than anyone else. That business worked spectacularly — but it is also a treadmill. Chips commoditize, competitors catch up, and the only durable advantage is making yourself the place where developers live, not the place where they rent machines. Sanders just said the quiet part out loud: the infrastructure was the wedge; the software is the business.
This is not an untested playbook. It is the AWS playbook from a decade ago — win developers with one great utility, then sell them the whole platform. The question is whether CoreWeave can make the jump from infrastructure operator to software vendor before the software vendors (or the hyperscalers) eat its lunch from the other direction.
The gap Goldberg says Forge closes
Chen Goldberg, CoreWeave's executive vice president of product and engineering, framed the launch around a gap that every ML team will recognize:
"As more people build with AI, they're putting models to work with their own data and workflows. That's where the gaps between model capability and system performance become clear. Engineering teams need to understand those gaps, identify the signals that matter, improve the next version and measure whether the change worked under real operating conditions. All of this has to function as a single, connected system. Forge brings all these capabilities on one platform, so what a business learns from running AI becomes part of how engineers improve it."
He has a point. The industry has spent two years obsessing over model capability — bigger checkpoints, better benchmarks — while the harder problem has quietly become operational: taking a model that aces an eval and making it behave reliably inside a real product, on real data, under real load. That is a systems problem, not a weights problem, and the teams that solve it will matter more than the teams with the highest benchmark scores.
CoreWeave is also leaning on proof by example. The company says MasterClass and Canva are already building on Forge — two brands that sit at the intersection of heavy AI workloads and consumer scale. Take the claim as a company assertion rather than an independent fact, but the choice of partners is deliberate: these are the kinds of companies that burn through inference budgets and have every incentive to make each model version perform better than the last.
Software meets silicon: Vera Rubin reaches production scale
CoreWeave paired the software launch with a hardware milestone that makes the timing legible. In a separate Business Wire announcement at the same event, the company said Nvidia's Vera Rubin NVL72 rack-scale system is now available on CoreWeave Cloud, with Cognition as the first customer anywhere running production workloads on the system.
Customers get the new racks — 72 Rubin GPUs and 36 Vera CPUs per rack — under the same operating model and tooling as their existing Nvidia GB200 and GB300 fleets, with CoreWeave's performance engineering layered on top. Goldberg tied the two launches together explicitly, calling fast bring-up "the payoff from years of engineering our platform across GPU generations" and arguing that agentic workloads — long contexts, repeated model calls, thousands of concurrent tasks — "put pressure on the entire platform."
This is the one-two punch: the hardware generation that finally tilts the economics toward inference and agentic workloads, and a software layer designed to manage exactly those workloads. Whether you buy CoreWeave stock or not, the pairing is the right read on where the industry is going — the next competitive frontier is not just having Rubin racks, but knowing what to do with them.
CoreWeave Forge skeptics: the bear case
Two notes of caution are in order. First, the MLOps and agent-development space is crowded to the point of claustrophobia. Hugging Face, Databricks, Modal, LangChain's LangSmith, and every hyperscaler's in-house tooling are all selling some version of "close the loop." Forge's differentiator is that it runs where the compute runs — but whether that integration is worth more than best-of-breed tooling from specialists is exactly the bet investors will have to price.
Second, much of Tuesday's news is company-claimed. The 4,500-attendee figure, the MasterClass and Canva partnerships, Forge being "available now" — all come from CoreWeave itself. The company also unveiled a partner network with validated integrations, naming VAST Data, CrowdStrike, and ClickHouse as early partners, which at least signals that the platform is being built with real ecosystem plumbing rather than as a demo.
For more on how CoreWeave has been financing its buildout, see our coverage of the $6.4 billion in financing CoreWeave raised from Wall Street earlier this month.
CoreWeave Forge FAQ
What is CoreWeave Forge?
A "development layer" for building and improving AI models and agents — run, observe, curate, improve, evaluate, repeat — in one connected environment.
What tools does it stitch together?
Weights & Biases Models (experiment tracking), OpenPipe (post-training), the marimo notebook project, plus CoreWeave Registry for versioning checkpoints and agent configs.
Is it a lock-in play?
CoreWeave says no — Forge stays open across the models, frameworks and other clouds teams already use.
What's the Vera Rubin connection?
Nvidia's Vera Rubin NVL72 racks (72 Rubin GPUs, 36 Vera CPUs per rack) launched on CoreWeave Cloud the same week, with Cognition as the first production customer.
The bottom line
CoreWeave's message on Tuesday was remarkably coherent for a company making a category jump: own the metal, own the tooling, own the version history, and developers will never leave. The GPU-rental business made CoreWeave famous. Forge is its attempt to make itself indispensable. If the AI loop is really where the industry's value is moving — from weights to workflow, from training to operations — it is exactly the right bet. The hard part, as always, is execution.
Sources: CoreWeave announcements via Business Wire (Sept. 30, 2026) for Forge and Vera Rubin NVL72 availability; Data Center Knowledge (Forge platform reporting, Corey Sanders quote); Unite.AI (launch details); DevOps Digest (Chen Goldberg quote). Partner, attendee, launch-partner, and availability claims are company-reported. The bitcoinversus.tech pairing angle was verified against CoreWeave's primary releases.
Comments
Post a Comment