I had the chance to spend a couple days with CoreWeave at its Fully Connected event in San Francisco. While the company had some big announcements, the message I walked away with was around an aggressive expansion into the enterprise market.

CoreWeave has made some significant investments in the full AI stack – from datacenter infrastructure to the systems and racks that run inside, to the AI software stack. And it is clear the company wants the market to know that while it went public on training, it will grow on inference.

Did it accomplish that mission? I’m going to dig into this.

Raw Capacity Is a Commodity

Neocloud. GPU cloud. AI infrastructure provider. There are a few terms we have for this new class of cloud providers that have cropped up over the last few years. I prefer GPU cloud. Why? Because they all have their roots in delivering GPU capacity to the model builders. Case in point, in its most recent quarterly filing, CoreWeave disclosed three customers accounted for 72% of its $2.57 billion in revenue. While the names aren’t disclosed, we know some of its top customers are Microsoft, OpenAI, Meta, and Anthropic.

For the past few years, the GPU cloud story has been pretty simple. Get power. Get accelerators. Sell capacity to the labs training the biggest models. Build as fast as you can and sell all of it. And really, it hasn’t been selling as much as taking orders. If a provider proved competent in securing the resources, the model builders kind of took things from there.

This business model has worked. CoreWeave’s latest results represent a 112% year-over-year gain. More impressively, it boasts a revenue backlog of about $104 billion.

But raw capacity is a commodity. Training demand won’t go away, but I expect its growth to level off. And when it does, a company that only sells GPU hours ends up in a race to the bottom on price. Yes, execution still separates providers. Reliability, utilization, and support all matter. But execution alone is a thin moat when there are so many options.

So, where does the durable business come from after the training gold rush? I’ve been asking this question for the last three years. The answer? Inference. And the tools wrapped around it. It’s where AI is used and where it drives business outcomes.

This is where CoreWeave is moving. The company has developed the software and tools to deliver a vertically integrated inference stack. And it complemented this with a go-to-market investment to drive adoption. The short-term results have been encouraging, as its managed inference platform grew from $1 million to more than $100 million in booked ARR within a few months, with a target of at least $250 million by year-end. Even at target, that’s a small slice of the $18.5-$19.5 billion run rate that CoreWeave expects to exit 2026 with.

The product shift is visible. So are the first green shoots of a healthier revenue mix. It’s worth going a little deeper on what CoreWeave is doing to enable this inference shift.

Above and Below the Waterline

I think about CoreWeave’s strategy in two layers.

Below the waterline is infrastructure. Datacenters, power, racks, networking, storage, and the software that operates them. This is table stakes for every GPU cloud, and CoreWeave does it really well. Maybe the best. Part of that comes from being one of the earliest GPU clouds in the market, and the experience shows. The company understands the balancing act of a supply chain spanning GPUs and racks to transformers and gas turbines.

Part of it also comes from getting the most out of that infrastructure. When you have six generations of NVIDIA GPUs and different types of CPUs deployed everywhere, how do you make sure each piece of silicon is utilized to its fullest extent? Again – this is where maturity in market pays off. Over the past nine years or so, CoreWeave has built the tooling that does exactly this. Mission Control gives CoreWeave infrastructure teams and its customers a consistent operating layer across every generation of GPU sitting in the datacenter. Think of it as a control plane of sorts that handles observability and works alongside SUNK (Slurm on Kubernetes), which does what most of us consider infrastructure orchestration.

Mission Control is an example of a tool that sits below the waterline but has surfaced to customers. It was born out of how CoreWeave’s own engineers have managed the fleet at scale for years. Productizing it gives customers the same tooling, which builds confidence in running production AI on CoreWeave.

One of the cool capabilities I saw in Mission Control is its ability to bring infrastructure telemetry into a customer’s AI-assisted development environment. Say your application slowed down after yesterday’s deployment. Connect Claude Code to Mission Control via MCP, and it can query metrics and logs alongside your code and configuration to help determine what changed – and whether the problem is in the application, the infrastructure, or how the two interact.

Back to AI infrastructure. One of the announcements at Fully Connected was CoreWeave making NVIDIA’s Vera Rubin NVL72 available with Cognition as its first customer (online within two days). Does being first really matter? For me, first out of the gate with a new platform isn’t really that big of a deal – a couple months from now everybody will have it. What’s impressive is how quickly the CoreWeave engineers can go from the shipping dock to operational. And how they’ve been able to do it generation after generation.

OK, so what sits above the waterline is the environment. These are the tools that let teams build, run, and improve AI without assembling it all themselves. That’s Forge.

What Forge Actually Is

Mission Control tells you whether the infrastructure is healthy. Forge answers a different question: is the model or agent doing what it should? Getting AI into production is hard. Keeping it working as users, data, and demands change is harder. When something goes wrong, teams need to figure out why, make a change, and know whether that change helped. Today, that usually means jumping between tools and piecing the evidence together by hand. That’s the problem Forge goes after.

Forge brings together Weights & Biases, OpenPipe’s post-training capabilities, and marimo notebooks with CoreWeave’s own training and inference services. Weights & Biases is the anchor, and since so many AI teams already use it to track experiments, CoreWeave isn’t starting from scratch.

CoreWeave organizes Forge around an AI loop: run, observe, curate, improve, and evaluate. In plain terms, put the application into production. See how it behaves. Pull out the failures that matter. Make changes and test them before deploying again.

Agent Lens (in public preview) handles the middle of that loop. It finds related failures in production traces and turns them into test cases. From there, the fix might be a better prompt or a change to how an agent uses its tools. Or it might take fine-tuning or reinforcement learning. Forge puts all of those options in one environment. A registry tracks each model version along with the data and runs behind it.

To me, the value is in how those pieces work together. Can a team take something that failed in production, understand it, fix it, and verify the result without leaving that environment? That’s the promise. If Forge delivers, teams spend less time on plumbing and more time improving the AI.

CoreWeave also says Forge works across models, frameworks, and clouds. I think that’s smart. A customer can start using the software without moving its workloads to CoreWeave. That gives CoreWeave a way into development teams it might never reach with an infrastructure sale. Win the developers first, and the infrastructure conversation can follow.

But openness raises a fair question. Part of CoreWeave’s pitch is that Forge and its infrastructure work better together. So how much of the value comes from Forge itself, and how much depends on running it on CoreWeave?

I like the approach. The demo I saw showed the pieces working together and addressing a problem that’s real. Next, I want to see what it takes for customers to get the same result. How long does implementation take? What expertise does it require? Does the connected environment actually make teams more productive? That’s where Forge has to prove itself beyond the demo.

Does This Set CoreWeave Apart?

I wouldn’t call the AI loop or Forge unique. AWS, Microsoft, and Google have all gone after the same problem with their own evaluation, observability, and optimization tools. Everybody sees the problem, and everybody’s solving for it in similar ways.

And honestly, I don’t think CoreWeave needs to out-differentiate anyone here. What it needs to do is show enterprise customers it can deliver the same capabilities, with the same ease of use and the same reliability of outcomes. Parity on the things enterprises care about is what gets CoreWeave into the conversation.

Where CoreWeave can differentiate is in a couple of other ways. The first is vertical integration. CoreWeave built an AI cloud for AI workloads from the ground up. AI is not a workload that has emerged and now needs to be supported. There’s not a decision made – from datacenter construction to infrastructure to software – that isn’t chosen specifically to make AI workloads run fast and efficiently.

The second is focus. The hyperscalers serve every workload on earth. CoreWeave wants to do one thing really well – AI. It isn’t trying to compete on hosting databases in virtualized environments. And on this front, I think CoreWeave is ahead of its GPU cloud peers. No other neocloud I track pairs AI infrastructure with a development platform this deep.

So how do you know whether either advantage holds up? Look at what customers experience day to day. Are their teams spending less time stitching together infrastructure and tooling? Are they diagnosing problems faster and improving models more quickly? Are they getting better cost at the same quality? If the answer is yes, CoreWeave’s differentiation is real and not just a positioning exercise.

 

For the bigger players, the fight looks different. On architecture, CoreWeave has the edge. Its infrastructure was built for AI from day one, and it shows. But the hyperscalers bring something much harder to replicate: decades of enterprise relationships, along with the buying rhythms and partner channels that come with them. Enterprises already run their data, storage, and security services on these platforms, and they lean on consulting partners who’ve supported their IT for twenty years. Add a broad choice of silicon, frameworks, and models, and you’ve got an ecosystem that’s woven into how enterprises operate. CoreWeave can stand up AI infrastructure in a matter of months. Building that kind of trust and reach takes years.

Builders, and Then Everyone Else

This is CoreWeave’s challenge. The company will attract AI-native startups and early-adopter enterprises with teams of AI talent ready to consume the infrastructure (and stack) that CoreWeave provides. But what about everybody else? By my estimate, that’s 80-90% of enterprises still figuring out this “AI thing.”

CoreWeave was very deliberate in its keynote, making the enterprise its focus. In fact, the first thing the company did was bring out Caterpillar to talk about the company’s use of CoreWeave. And Capital One. These are two of several enterprises building on CoreWeave – a good signal, though not yet a trend. And frankly, the two on stage fall squarely in the “early adopter” camp, to borrow from Geoffrey Moore’s Crossing the Chasm. Capital One even describes itself as a technology company that does banking. It has the engineering depth to build and run the kind of environment CoreWeave demonstrated. It’s a builder.

The dividing line is technical capability. CoreWeave frames its shift as moving from building AI to building with AI. I love that line. It’s accurate. And it’s reducing the work that builders have to do themselves.

Quickly following the builders will be a third phase. Let’s call it “consuming AI” – or AI consumers. This is the commercial enterprise that wants to put AI to work without a team of AI engineers. Maybe without a single one. It wants outcomes. How the AI gets built and tuned is someone else’s problem.

Going back to Moore, let’s put it in Crossing the Chasm terms. Early adopters will assemble the pieces themselves. The pragmatists on the other side want the whole product. References from peers. A known implementation cost. Integration with the systems they already run. Someone accountable when it breaks. Simpler developer tools help. They won’t close the chasm alone.

CoreWeave is taking steps. It positions ARIA, its research agent, to make the loop more accessible to both expert and non-expert users. And ARENA lets teams benchmark their workloads on CoreWeave with its engineers before committing. Both help technical teams move faster. Neither hands a business without an AI team a complete, supported solution.

The market’s not there yet, and I am in no way implying that CoreWeave is unable to win in the commercial enterprise space. In fact, it’s taken important first steps by bringing on the right team across marketing and sales. Just a note that as this market shift happens – as we cross the chasm – CoreWeave will be competing against players its target customers already know and trust (the hyperscalers). Starting to build those commercial enterprise roots today is critical.

The Most Underappreciated Announcement

Forge was the star of Fully Connected, and it’s a great demo. But I think the company’s announcement about the partner network will matter more than anything in the long run.

Here’s why. CoreWeave doesn’t have to become a solutions vendor to reach the AI consuming enterprise. In fact, that customer may never buy from CoreWeave directly. It may buy an application, a managed service, or an industry solution from a partner running on CoreWeave. The demand still lands on CoreWeave’s platform.

And that’s what the CoreWeave Partner Network is about. It spans infrastructure, data services, ISVs, and models and inference. And it includes data platforms like VAST Data, observability from Datadog, security from CrowdStrike, and enterprise platforms from IBM and Red Hat. CoreWeave says each integration is tested under production conditions.

This is a good beginning. And today it’s mostly a technology integration network for builders. The next step is partners who package and deliver outcomes. That means vertical ISVs, systems integrators, and managed service providers. And we will know it’s working when we see the revenue shift that shows these types of partners are delivering a growing share of revenue.

One other thing I would like to see is diversity in silicon. I understand CoreWeave is in an interesting position. NVIDIA is more than a friend and partner. It’s a strategic investor. It currently holds about 47 million shares of CoreWeave stock, along with a $6.3 billion commitment to purchase unsold capacity through April 2032.

This isn’t unusual. AMD has made similar moves, from backstopping capacity at neoclouds like Crusoe to offering attractive incentives for model builders to train on its silicon.

However good any single vendor’s silicon is, enterprise AI is not a homogeneous environment. Different accelerators will perform different functions really well. If an AI cloud wants to succeed, it has to embrace this diverse silicon landscape. Enterprises want one-stop shopping, and so do the partners building solutions for them. Nobody wants the complexity of going to different clouds for different types of accelerated compute. This is where the hyperscalers’ breadth is a real advantage.

If CoreWeave gets this partnership and silicon strategy right, it has a real path to the AI consumer. Miss it, and that path is really difficult.

A Shakeout Is Coming

Step back and look at the whole neocloud/GPU cloud market. I think a shakeout is coming.

Sound familiar? We saw something like this in the dot-com era. A real inflection point in technology, lots of capital, and valuations built on forecasts and promises. I lived through it.

The internet transformed the economy anyway. But many of the companies that funded the build-out never earned a return.

The same test applies here. Does customer demand justify the capacity being built? Do the economics hold as pricing and financing costs shift? Some neoclouds will become true AI production platforms. Some will get acquired. Some will simply fade away. The market on the other side will be healthier for it.

CoreWeave is doing what it needs to do to land on the right side of that line. It has the infrastructure. It now has the environment. It’s building the ecosystem. And Forge is becoming a big part of the answer. Execution decides the rest.

What I’m looking for over the next year or so is simple. It’s the green shoots that show a healthy trajectory for a company expanding into the enterprise:

• Paid Forge adoption and repeat use. Paid conversion and renewals tell the story.
• Quantified customer outcomes. Faster fixes, lower cost, and better quality, with numbers attached.
• Adoption beyond sophisticated builders. A repeatable pattern of enterprise logos being added.
• Partner breadth and delivery. CoreWeave doesn’t need to become another AWS, but it does need to support the big players in the enterprise datacenter, along with integrators and vertical ISVs delivering real projects.
• Managed inference growth. Does it hit its year-end target?
• Customer concentration. Is the share of revenue from the top three customers coming down?

The Bottom Line

CoreWeave used Fully Connected to show it’s more than a GPU landlord. Forge is a serious platform for serious builders, and it sits on infrastructure built for AI from the ground up.

So did CoreWeave make the case that it will grow on inference? On product, I think it did. Forge, managed inference, and the infrastructure underneath them add up to a credible stack for builders. On the go-to-market front, it’s still early. The enterprise team, the partner network, and the customer logos are all moving in the right direction, but it will take time for that to show up in the numbers.

The next act is reaching the enterprises that want to use AI without assembling the stack themselves. Partners give CoreWeave a route to those customers. Repeatable deployments will prove the route works. If those deployments show up over the next year, CoreWeave won’t just land on the right side of the GPU cloud shakeout. It’ll help define what an AI cloud looks like on the other side of it.