CoreWeave
CoreWeave
CoreWeave signed two landmark agreements in 48 hours that reshaped the artificial intelligence infrastructure landscape. On April 9, the neocloud provider announced a $21 billion expanded agreement with Meta to supply AI cloud capacity through December 2032. The following day it secured a multiyear deal with Anthropic to run Claude at production scale. The company says nine of the 10 leading AI model providers now run workloads on its platform, a claim that underscores how quickly the neocloud model has moved from niche offering to central infrastructure layer.
The back-to-back deals highlight a counterintuitive reality for technology executives. Companies spending well over $100 billion a year building their own data centers still need external GPU capacity to match the pace of AI demand. The neocloud layer appears to be graduating from a temporary supplement into an increasingly important component of how the AI economy operates.
Meta plans to spend between $115 billion and $135 billion on capital expenditure in 2026, a figure that dwarfs what most Fortune 500 companies generate in total revenue. The company operates its own gigawatt-scale data center campuses. It designs its own AI accelerator chip called MTIA. Yet none of that is sufficient. The $21 billion CoreWeave agreement builds on a prior $14.2 billion contract signed in September 2025 and according to CoreWeave will include early deployments of the Nvidia Vera Rubin platform.
The most likely explanation is that the binding constraint for Meta is not capital but time-to-capacity. Training and deploying large language models at Meta's scale requires clusters of tens of thousands of GPUs available in specific configurations at specific times. Building those clusters internally takes years. Renting them from a provider that has already secured the silicon and power gets workloads running in months.
Anthropic faces a different version of the same problem. The company's annualized revenue run rate surpassed $30 billion in April 2026, up from roughly $9 billion at the end of 2025 according to Bloomberg. More than 1,000 business customers now spend over $1 million annually on Claude services. That growth demands inference capacity at a scale Anthropic cannot build on its own. The company already runs workloads on AWS Trainium, Google tensor processing units and Nvidia GPUs. Adding CoreWeave's Nvidia-native clusters gives Anthropic another production-grade option for serving Claude to an expanding enterprise customer base.
CoreWeave's numbers tell a story of explosive growth built on borrowed capital. The company went public in March 2025 at $40 per share and now carries a revenue backlog exceeding $66.8 billion. Management has guided for $12 billion to $13 billion in revenue for 2026. Its client roster includes OpenAI with a $22.4 billion contract, Meta with $35 billion in total commitments and now Anthropic at undisclosed terms.
That growth comes at a steep cost. Capital expenditure is expected to roughly double to between $30 billion and $35 billion in 2026. Net interest expense for full-year 2025 reached $1.2 billion according to company filings. At current guidance ranges, CoreWeave would be spending between roughly $2.3 and $2.9 for every dollar of revenue it earns. The business model depends on sustained demand growth to justify the debt financing required to build data centers ahead of customer need.
Customer concentration adds another layer of risk. Microsoft accounted for about 67% of CoreWeave's 2025 revenue. The Meta and Anthropic deals help diversify that base, but the company remains dependent on a small number of hyperscale clients whose spending commitments could shift if AI demand plateaus or if their own internal capacity catches up.
CoreWeave is not the only company capitalizing on the build-versus-buy tension. Nebius signed its own agreement with Meta earlier this month for $12 billion of dedicated AI infrastructure capacity over five years, with deployments also expected to feature the Nvidia Vera Rubin platform. The neocloud operator aims to reach up to 1,000 megawatts of connected data center capacity by the end of 2026. Lambda is preparing its own initial public offering after securing a deal with Microsoft and $1.5 billion in funding.
The hyperscalers are not standing still either. AWS continues scaling its Trainium chips. Broadcom and Google are collaborating to provide Anthropic with access to approximately 3.5 gigawatts of TPU-based computing capacity beginning in 2027, an arrangement that gives Anthropic a non-Nvidia path to scale. Meta is iterating on MTIA. Each of these custom silicon programs is designed to reduce dependence on Nvidia GPUs over time, which could erode the pricing advantage that neoclouds currently enjoy.
The 48-hour sequence of deals reveals a shift in how the AI infrastructure market is organizing itself. Even as hyperscalers build custom silicon and expand their own data centers, they and the AI labs they serve are simultaneously signing multi-billion-dollar contracts with specialized GPU cloud providers. The result is a multi-sourcing model where neoclouds like CoreWeave supplement rather than replace in-house capacity.
For enterprise technology leaders evaluating their own AI infrastructure strategies, three dynamics merit close attention. The shift from training to inference as the dominant workload means production reliability and latency matter as much as raw compute throughput. CoreWeave's growing client roster suggests that specialized AI cloud providers are gaining significant leverage in the market, though it remains unclear whether this concentration will persist as hyperscaler alternatives mature. And the financial structure underpinning these providers, with capital expenditure running at multiples of current revenue, means the durability of the neocloud model remains an open question tied directly to whether AI demand continues to accelerate through the end of the decade.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。