AI Gets Cheaper. Demand Gets Bigger: Inside CoreWeave’s AI Infrastructure Strategy

AI's economics are changing fast. Token costs are coming down. Models are becoming more accessible. Enterprise use cases are expanding. And instead of reducing the need for infrastructure, those trends could create substantially more demand for it.

That was one of the central themes in a conversation between Constellation Research founder R "Ray" Wang and Mike Intrator, CEO of CoreWeave, at Fully Connected 2026. The discussion unpacked what happens when AI moves from a frontier technology into a broader economic infrastructure layer.

AI infrastructure is not just another cloud

CoreWeave's thesis starts with a simple idea: AI workloads are different.

The company was built around specialized infrastructure designed specifically for AI, rather than adapting a generalized cloud to the demands of AI workloads.

Intrator compares the difference to that between a minivan and an F1 car. A generalized cloud can do many things well. AI infrastructure needs to be purpose-built for workloads operating at an entirely different scale.

That specialization extends beyond compute. Intrator argues that the software stack, cloud, security and other components all need to come together as one functioning system. AI infrastructure doesn't work if one piece of the puzzle is missing.

Cheaper AI could mean more infrastructure demand

One of the most interesting parts of the conversation was the economics of inference. As token costs decline, the obvious assumption is that infrastructure economics should get tougher. But Intrator sees another dynamic at work: lower costs can expand demand.

That's the Jevons Paradox at work. As AI becomes cheaper to use, more organizations can afford to use it, and existing users can expand their use.

Intrator points to the broader diffusion of AI across the economy as a key driver. Lower cost per token makes it possible to bring AI into more decisions, more workflows, and more use cases. That creates an important distinction for enterprise leaders: cheaper AI does not necessarily mean less infrastructure. It can mean more AI.

The GPU lifecycle may be longer than expected

Another important point is that not every AI workload needs the newest hardware.

As AI use cases mature, workloads can be matched to different infrastructure levels. CoreWeave has seen older GPUs remain economically valuable as customers right-size hardware for specific workloads.

Intrator points to A100 GPUs already contracted through 2029 as an example of how long the economic life of AI infrastructure can extend. That changes how enterprises should think about AI infrastructure. The race isn't necessarily about replacing everything with the newest chip as quickly as possible. It's also about matching the right infrastructure to the right workload.

The model doesn't matter as much as the infrastructure

The conversation also touched on the rise of open-weight models and lower-cost alternatives.

CoreWeave's position is relatively model-agnostic. Intrator argues that both open and closed-model ecosystems can flourish, while the physical infrastructure remains fundamental to delivering AI at scale.

That's an important distinction. The AI model layer will continue to change rapidly. The underlying infrastructure required to train, run, and serve those models remains a more persistent requirement.

AI infrastructure is moving into the enterprise

The next phase is not just about frontier labs.

CoreWeave is increasingly focused on making infrastructure usable across a broader customer base, including enterprises. Intrator cites companies such as Caterpillar and Capital One as examples of organizations that need access to AI infrastructure and the tooling to actually use it across the business.

That's where Forge comes into the conversation.The goal is to bring together the different pieces of CoreWeave's ecosystem and provide the tools customers need to build and operate AI applications across their organizations.

The implication is significant: AI infrastructure is shifting from something consumed primarily by AI specialists to something the broader enterprise needs to operate.

The infrastructure race is far from over

The conversation ended with a deceptively simple question: Where will chips, energy, land and demand be in three years?

Intrator's answer was consistent across all four: up.

His underlying argument is that demand for AI is becoming deeply rooted and sustained as the technology moves from frontier labs into enterprises and broader society. More adoption means more compute, which means more chips, power, land, and capital.

That may be the biggest takeaway from the conversation. The AI infrastructure story isn't simply about building enough capacity for today's models. It's about preparing for what happens when AI becomes cheaper, more accessible, and embedded in more of the economy.

The models will change. The workloads will change. The economics will change....But the infrastructure requirement is only getting bigger.