CoreWeave Conference: “Insatiable” AI Demand Leaves 2026 Compute Capacity Broadly Sold Out

Key Points

• Demand is "insatiable" and 2026 is broadly sold out — customers have broadened from AI labs to hyperscalers and large enterprises, signing longer take‑or‑pay contracts now weighted toward five‑ to six‑year terms and spanning multiple GPU generations (A100, H100, H200, Blackwell).
• Massive capex and near‑term margin pressure — CoreWeave is guiding $30–$35 billion of capital spend (midpoint $32.5B) financed at the asset‑company level, and says margins are currently squeezed by rapid growth with a Q1 trough before expected expansion to roughly a 25% contribution margin after the initial three‑month ramp.
• Operational differentiation and software monetization opportunity — management argues GPU infrastructure is not fungible, cites an expanded relationship with NVIDIA (including a reported $2B incremental investment) and is considering selling its software stack and expanding storage/peripheral services as margin‑accretive growth avenues.

CoreWeave NASDAQ: CRWV co-founder and chief development officer Brannin McBee said demand for the company’s AI infrastructure remains “overwhelming” and “insatiable,” describing 2026 as “broadly sold out” in terms of billable compute capacity. Speaking with Morgan Stanley’s Keith Weiss, McBee said the demand backdrop has broadened from an early focus on AI labs to include hyperscaler cloud customers and, increasingly, large enterprises.

Demand broadens across customer types and contract terms lengthen

McBee said CoreWeave’s demand has expanded beyond the “starting cohort” of AI labs beginning in 2022, first reaching hyperscaler cloud clients and more recently accelerating within the enterprise segment. He pointed to the pace of enterprise adoption as “truly fascinating” and noted that customer behavior has been shifting toward longer commitments.

CoreWeave provides multi-year “take-or-pay” contracts, and McBee said the typical contract duration has increased over time. He described a progression from roughly three-year contracts about 24 months ago, to four-year contracts 12 months ago, to a backlog now weighted toward five-year contracts, with some extending up to six years.

McBee also highlighted that demand is not limited to the newest hardware. He said customers are requesting a range of GPU generations—including A100s, H100s, H200s, and Blackwell—and that these requests are often driven by workloads engineered for specific systems, with inference described as a key driver.

CoreWeave emphasizes operational differentiation and software “stack”

Weiss pressed McBee on whether CoreWeave’s growth is simply a function of supply availability or whether the company has durable differentiation. McBee argued that GPU infrastructure “is not fungible,” saying that a given GPU in one cloud environment is not necessarily equivalent to the same GPU elsewhere due to operational execution and performance at scale.

McBee said CoreWeave’s platform was built intentionally for “parallelizable workloads,” which he contrasted with more traditional “serializable” workloads. He described CoreWeave’s offering as a highly performant way to operate “supercompute scale” systems and said the company’s operational infrastructure is designed to bring large clusters online and stabilize them.

He also pointed to third-party benchmarking, referencing SemiAnalysis as an example of research that evaluates platforms for running AI infrastructure, and said customers continue to choose CoreWeave repeatedly across AI labs, hyperscalers, and enterprise customers.

On product expansion beyond compute, McBee said “peripheral” demand is rising with newer workload types such as agentic applications, which he said are increasing pressure on CPU demand and storage. He noted CoreWeave disclosed an “greater than 80%” storage attach rate among customers generating more than $1 million in revenue, and said its storage offering is “well north of $100 million” in ARR.

NVIDIA relationship and potential software monetization

McBee discussed an expanded relationship with NVIDIA that Weiss said included a $2 billion incremental investment. McBee said the broader agreement was designed to accelerate CoreWeave’s growth in line with AI adoption and said the relationship also reflects an acknowledgment of the strength of CoreWeave’s software stack for operating GPU infrastructure.

McBee said the company has discussed the possibility of selling its software solution to other entities, particularly those that want to own GPUs on their balance sheets or have data sovereignty requirements. He characterized this as a potentially “margin-accretive” opportunity.

Capex, financing structure, and near-term margin pressure

Weiss noted CoreWeave’s capital spending guidance of $30 billion to $35 billion and asked about financing. McBee said CoreWeave has been a “market leader” in financing AI infrastructure and described a structure with a parent company and an “asset co,” where assets and related capital spending are housed.

McBee said the capital spending—citing a midpoint of $32.5 billion—sits at the asset company level, which can access financing facilities supported by what he described as strong demand to “participate in the paper.” He added that CoreWeave is “very excited to announce” an advancement in asset-level structuring, without addressing specific headlines. He said this reflects the company’s execution track record and the durability of contracts and data center agreements.

On margins, McBee said the company experiences a roughly three-month investment period when bringing capacity online. After the infrastructure is online and stable, he said deployments generate about a 25% contribution margin to the parent for months three through 60 of a contract. He described current margins as pressured due to “extreme growth,” including nearly 30% quarter-over-quarter growth in active power from Q3 to Q4, and said Q1 represents a “trough” in margin performance, with expectations for expansion thereafter.

Supply chain constraints, power, and GPU useful life

McBee said supply chain execution remains “immensely difficult,” describing modern AI data centers as massive engineering projects requiring thousands of workers and specialized skills such as electrical engineering. He distinguished between overall grid power availability and the challenge of delivering “active power”—power that is effectively delivered into racks and servers.

CoreWeave primarily leases data center capacity while doing some self-development, and McBee said the company had “43 active sites in operation.” He acknowledged that CoreWeave was surprised by a site issue disclosed in Q3, but said it is now “firmly back on track” and was delivered “a little bit early” versus expectations.

Asked about rising memory prices, McBee said memory is a small portion of node costs compared with GPUs, and he emphasized that higher component and electricity costs ultimately get passed through to customers, while the company’s main focus remains securing components and managing supply chain risk.

On long-term capacity, Weiss cited a goal of procuring 5 gigawatts of incremental power by 2030. McBee said power is “out there,” with the key challenge again being data center supply chain and delivery. He said CoreWeave procures capacity based on customer demand, with planning conversations typically occurring 12 to 18 months ahead of deployments.

McBee reiterated CoreWeave’s position on GPU useful life, saying six years remains the appropriate depreciation assumption and that the company is seeing empirical support that usage could extend beyond that. He cited A100s as the oldest infrastructure in scale and said A100 pricing increased in 2025 and held pricing power within about 10% over the year, which he viewed as a signal of sustained demand for older-generation GPUs, particularly tied to inference growth.

Finally, on the possibility of using non-NVIDIA silicon, McBee said CoreWeave is “client-led” and does not build capacity speculatively. He said customers are currently only requesting NVIDIA infrastructure, even though CoreWeave considers itself hardware-agnostic in what it can operate.

McBee also outlined a broader software strategy that mirrors historical cloud buildouts: first optimize foundational infrastructure, then add peripherals and application layers. He referenced CoreWeave’s acquisitions—Weights & Biases, OpenPipe, Monolith, and Merino—as part of that layered approach, while noting it is difficult to predict how quickly those components will scale relative to the core GPU services business.

About CoreWeave NASDAQ: CRWV

CoreWeave is a U.S.-based provider of GPU-accelerated cloud infrastructure designed to support compute-intensive workloads such as artificial intelligence, machine learning, visual effects rendering and other high-performance computing applications. The company supplies access to large fleets of modern GPUs and complementary infrastructure that enable customers to train and deploy large models, run inference at scale, and process graphics-heavy workloads with low latency and high throughput.

CoreWeave’s product offering includes on-demand and dedicated GPU instances, bare-metal servers, private clusters and managed services tailored for enterprise and developer use.

This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.