AI infrastructure companies are discovering that renting computing power alone isn't enough to build sustainable, profitable businesses; the real opportunity lies in managed inference services that let customers pay only for the AI models they actually use. CoreWeave's latest earnings report reveals this transformation in action, with the company's managed inference offering growing from virtually nothing to over $100 million in booked annual recurring revenue (ARR) in just a few months.

What Is Managed Inference and Why Does It Matter?

Managed inference is essentially a hosting service for AI models in production. Instead of customers buying raw GPU capacity and managing their own model deployments, they can use a platform that handles the technical complexity. Think of it like the difference between renting a kitchen and renting prepared meals. CoreWeave's managed inference platform lets companies deploy open-weight models, fine-tuned versions of existing models, and AI coding agents without building their own infrastructure from scratch.

The appeal is straightforward: customers get flexibility in how they consume computing resources. They can choose between serverless deployments, where they pay per token processed, or dedicated capacity for predictable, high-volume workloads. Early customers like Grammarly and You.com are already running production traffic through the platform, including AI coding agents and custom model deployments.

How Are Infrastructure Providers Shifting Their Business Models?

CoreWeave's transition from pure infrastructure to a broader AI platform reveals a critical insight: the economics of GPU rental alone are getting squeezed. Here's how the company is adapting:

• Managed Inference as Expansion: After selling customers their initial GPU capacity, CoreWeave now offers a higher-margin path to deepen customer spending through token-based pricing on inference workloads, allowing the company to monetize ongoing usage rather than just upfront capacity sales.
• Observability and Monitoring Tools: CoreWeave is bundling software services that help customers understand how their AI models are performing, reducing operational friction and increasing switching costs.
• Agent Development Frameworks: The company is building tools specifically for agentic AI, a rapidly growing category where AI systems take autonomous actions, creating stickier customer relationships.
• Cross-Cloud Services: CoreWeave is positioning itself as a neutral platform that can work across multiple cloud providers, giving customers more flexibility and reducing vendor lock-in concerns.

What's Driving This Shift in the Market?

The fundamental driver is scarcity. CoreWeave ended the second quarter of 2026 with 1.5 gigawatts of active computing power and contracted power reaching 3.7 gigawatts, with demand from multiple customers for each GPU brought online. This capacity constraint gives infrastructure providers significant pricing power, but it also creates an opportunity to move upmarket.

When capacity is scarce and expensive, customers become more willing to pay for software and services that help them use that capacity more efficiently. Managed inference fits perfectly into this dynamic. It allows CoreWeave to capture value not just from the hardware rental, but from the software layer that sits on top of it.

How Is This Affecting CoreWeave's Financial Performance?

CoreWeave reported Q2 2026 revenue of $2.58 billion, up 112 percent year over year, beating Wall Street consensus expectations. However, the company's adjusted operating income margin fell to 5 percent from 16 percent a year earlier, reflecting heavy investment in new capacity and software capabilities.

The company's guidance suggests this margin compression is temporary. CoreWeave expects adjusted operating margin to reach the low teens by Q4 2026 as the company brings new capacity online and improves utilization. New customer contracts signed in Q2 are expected to carry contribution margins 5 to 10 percentage points above recent quarters, suggesting that the combination of pricing power and software attachment is already improving unit economics.

What Does This Mean for Enterprise AI Adoption?

CoreWeave's customer wins in Q2 2026 paint a picture of AI infrastructure moving beyond AI labs and into mainstream enterprises. The company added customers including Bentley Systems, Caterpillar, Grammarly, and Isomorphic Labs, while expanding relationships with Cognition, Databricks, and Runway ML. Caterpillar, for example, is using CoreWeave's infrastructure to train and deploy AI models for autonomous construction equipment, a physical AI application that requires continuous inference at scale.

This broadening customer base reduces CoreWeave's dependence on a narrow set of AI-native customers and suggests that managed inference services will become increasingly important as enterprises move AI from experimental projects into production systems. When AI models are running 24/7 in production, the operational complexity and cost of managing that infrastructure becomes a major concern.

What Role Does Hardware Innovation Play?

While CoreWeave focuses on the software and services layer, the underlying hardware infrastructure is also evolving to support inference workloads more efficiently. SanDisk and SK hynix recently released the High Bandwidth Flash (HBF) technical specification through the Open Compute Project, a standardized approach to adding high-capacity, high-bandwidth memory closer to AI compute cores. This innovation addresses a fundamental challenge in AI inference: moving data between memory and processors efficiently.

The HBF specification was developed with input from Google and Tenstorrent, suggesting that hyperscale AI infrastructure providers are actively shaping the hardware standards that will underpin the next generation of inference systems. As Alper Ilkbahar, Chief Technology Officer at SanDisk, noted, "AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment".

Alper Ilkbahar, Chief Technology Officer at SanDisk

"AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment. This specification helps give system designers a practical path to bring high-capacity, high-bandwidth memory closer to compute, while enabling more flexible architectures," said Alper Ilkbahar, Chief Technology Officer at SanDisk.

Alper Ilkbahar, Chief Technology Officer at SanDisk

What Should Enterprises and Developers Watch For?

The shift toward managed inference and software-attached services suggests that the AI infrastructure market is maturing. Companies that can combine reliable, low-latency compute capacity with intelligent software tools for model deployment, monitoring, and optimization will likely capture more durable economics than pure capacity providers. For enterprises, this means more options for deploying AI models without building entire infrastructure teams in-house.

CoreWeave's success with managed inference also signals that the economics of AI inference are becoming more favorable. As inference workloads scale, the cost per token processed should decline, making it more economical for enterprises to run AI models continuously rather than treating them as occasional batch jobs. This shift could accelerate AI adoption across industries, from manufacturing to financial services to life sciences.