On April 9, CoreWeave announced a $21 billion expansion of its existing cloud capacity agreement with Meta Platforms, extending through December 2032. The announcement arrived alongside a proposed $3 billion convertible notes offering and a separate $1.25 billion senior notes offering. The day before, CoreWeave's contracted backlog sat at roughly $66 billion. By close of business April 9, it stood at $87.8 billion. The question a technology executive should be asking is not whether CoreWeave is growing. It clearly is. The question is what that growth is built on, and whether the foundation holds.

The Inference Shift Explains the Meta Deal

The AI infrastructure cycle from 2022 through 2024 was defined by model training: assembling massive clusters of graphics processing units (GPUs) to run the compute-intensive process of building foundation models. That work is capital-intensive, time-bounded, and concentrated at a handful of frontier labs. The cycle from 2025 onward is different. Inference, the work of serving trained models to users in real time, is continuous, scales with usage rather than with research calendars, and requires sustained GPU capacity rather than burst capacity.

The Meta contract is structured around inference. The $21 billion buys Meta dedicated capacity for serving models at scale, across multiple CoreWeave locations, including some of the first deployments of NVIDIA's Vera Rubin platform. For CoreWeave, this is the argument made concrete: inference demand is not episodic, and organizations that need sustained, high-performance compute at scale cannot simply queue behind hyperscaler capacity.

The compute requirement just got heavier. On April 8, the day before the CoreWeave deal was announced, Meta released Muse Spark, the first model from its new Superintelligence Labs. I covered the strategic implications in a separate post at shashi.co. The detail that matters here: Muse Spark is closed and proprietary, a deliberate break from the open-weight Llama releases that defined Meta's prior model strategy. A proprietary hosted model serving more than 3.5 billion users across WhatsApp, Instagram, Facebook, and Meta's Ray-Ban glasses does not get downloaded and self-hosted by developers. Every request runs on Meta's infrastructure, or on compute Meta has contracted. The timing of the CoreWeave expansion is not a coincidence.

This Is Financial Engineering as Much as Infrastructure

CoreWeave's model is unusual in the cloud industry. The company acquires NVIDIA GPUs, finances them using the GPUs as collateral, leases them to customers under long-term contracts, and uses those contracted cash flows to service debt and fund the next round of hardware acquisition. Think of it the way a bank finances a fleet of trucks: the trucks secure the loan, the delivery contracts pay it down, and the profits fund more trucks. The loop works while demand grows faster than the cost of capital. The numbers are large enough to warrant careful reading.

Full-year 2025 revenue reached $5.13 billion, up 168% year over year. The company projects 2026 revenue above $12 billion. Adjusted earnings before interest, taxes, depreciation, and amortization margins sit at approximately 60%, which looks healthy until the accounting context is applied: the net loss for 2025 was $1.17 billion, driven by depreciation on the GPU fleet and interest on the debt used to buy it. Long-term debt stood at $21 billion at year-end 2025. The 2026 capital expenditure plan runs between $30 billion and $35 billion. CoreWeave projects spending $2.60 in capital expenditure for every dollar of new revenue in 2026. The April 9 convertible and senior notes offerings layer an additional $4.25 billion onto that structure.

The acquisition of Core Scientific in 2025, structured as an all-stock transaction, was an attempt to address part of the cost structure directly. CoreWeave gained control of 1.3 gigawatts of data center power capacity and access to more than 1 gigawatt of expandable infrastructure, converting former cryptocurrency mining sites into AI compute facilities. The stated rationale was eliminating future lease obligations and achieving annualized cost savings of approximately $500 million by 2027. That is meaningful, but it does not change the fundamental shape of the financing model.

Specialization Is the Moat Claim. Customer Concentration Is the Counterargument.

CoreWeave's competitive positioning rests on two claims: that NVIDIA GPU infrastructure optimized for AI workloads outperforms general-purpose cloud, and that its "preferred partner" status with NVIDIA gives it first access to successive hardware generations. The second claim has been operationally real: CoreWeave was among the first to deploy Blackwell architecture at scale, and the Meta contract includes early Vera Rubin deployments. Being first to next-generation hardware matters in a market where customers are benchmarking inference throughput.

The customer concentration figures complicate the moat story. In 2024, Microsoft represented approximately 62% of CoreWeave's $1.92 billion in revenue. In the second quarter of 2025, a single unnamed customer accounted for 71% of revenue. With the April 9 announcement, Meta and OpenAI together represent nearly 65% of CoreWeave's $87.8 billion contracted backlog. The customer roster has reportedly expanded to roughly two dozen named accounts, which is growth from the near-total Microsoft dependency of 2024. It is not yet diversification by any conventional definition.

The structural concern is one that applies to any infrastructure provider whose largest customers are also building their own alternatives. Microsoft has Azure and its own accelerator roadmap. Meta is developing its own AI chips, the Meta Training and Inference Accelerator, alongside GPU procurement. OpenAI is a major equity partner in the Stargate infrastructure project. Each of these customers has a stated long-term intention to internalize more compute. The CoreWeave contracts lock in revenue through the end of the decade, which provides near-term visibility. It does not resolve the question of what happens at renewal.

Where CoreWeave Fits in the Clawconomy

AI agents need physical hardware to run on, just as websites need servers. CoreWeave occupies that physical compute layer in what I have been calling the Clawconomy: the infrastructure economy built to support AI agents at scale. Its moves in 2025 and 2026 tell a consistent story. The OpenClaw governance frameworks, the NVIDIA hardware partnerships, the Cisco and Dell integrations, and the FedRAMP authorization pursuit all point to a company building the plumbing before the tenants arrive. The inference economy justification for CoreWeave's capital structure rests on one core assumption: that AI model usage grows large enough, fast enough, that general-purpose clouds cannot handle it. If that assumption holds, CoreWeave looks prescient. If it does not, the debt structure has no cushion.

NVIDIA's $2 billion strategic investment in CoreWeave in early 2026 signals what the hardware maker needs from this relationship. NVIDIA benefits from CoreWeave buying and deploying graphics processing units at volume. An equity stake deepens that alignment. CoreWeave gets preferred access to next-generation chips. The deal makes obvious sense for both sides. What it also means is that CoreWeave's entire infrastructure dependency and its primary strategic backer are the same company. That is not a moat. That is a very close relationship with significant mutual dependency.