CoreWeave

CoreWeave's revenue expanded 24% sequentially to $2.6 billion in the second quarter. Quarterly power capacity addition of 500 megawatts brought total active power to over 1.5 gigawatts. Revenue backlog grew 56% sequentially to $104 billion, not including the $25 billion added in the third quarter.

Why it matters: CoreWeave's healthy customer diversification underpins our thesis that enterprise artificial intelligence cloud adoption will help the company rapidly scale its revenue. We see CoreWeave adding industrials and healthcare companies as new customers while expanding its bookings with AI natives. Software-driven services also became a key growth contributor, with managed inference hitting the $100 million annual recurring revenue benchmark a few months after launching. More exposure to managed services can support better margins for CoreWeave over the long term. Second-quarter adjusted operating margin of 5% was in line with expectations, and management expects quarterly adjusted operating margin to hit the low teens by year-end. CoreWeave is making progress to reach our midcycle adjusted operating margin forecast of above 20% around 2030.

The bottom line: We lift our fair value estimate for no-moat CoreWeave to $115 from $106. Shares look fairly valued following the 15% rally in after-hours trading. Better customer diversity and positive managed software traction both point to stronger fundamentals for CoreWeave. In addition, the firm was able to raise more than $10 billion of financing in the second quarter, bringing the year-to-date total to over $30 billion. Strong market demand and a stable financing pipeline should allow CoreWeave to reach its capacity goal of 1.85 GW by the end of 2026 and 3 GW by the end of 2027.

Between the lines: CoreWeave's partnership with Nvidia remains solid. The company completed the industry's first validation of Vera Rubin NVL72 racks and is likely to launch Vera Rubin chips ahead of competitors.

Management increased full-year revenue guidance by $300 million at the midpoint to $12.4 billion-$13.2 billion. Adjusted operating income guidance of $960 million-$1.15 billion also increased by $55 million at the midpoint. We saw a big hike with CoreWeave's full-year capital expenditure guidance to $35 billion-$39 billion, from $31 billion-$35 billion previously. The increased capital expenditure guidance gave us strong confidence about CoreWeave's data center ramp-up pace. Together with the $30 billion in financing that CoreWeave has secured this year so far, the company is on track to more than double total active power capacity in 2026. For the third quarter, management expects $3.45 billion-$3.6 billion of revenue and an adjusted operating income of $200 million-$260 million.

Our fair value estimate for CoreWeave is $115 per share, which implies an enterprise value/sales multiple of 6 times. Despite our five-year compound annual revenue growth forecast of 70%, we do not expect GAAP earnings per share to turn positive over the next three years. Similarly, free cash flow would not turn positive before 2033. CoreWeave should remain in a stage of heavy capital outlay to expand its cloud service capacity.

Over the next three to five years, bookings from established AI providers like Microsoft, OpenAI, and Meta should allow CoreWeave’s revenue to scale quickly. These three customers have currently committed over $66 billion to CoreWeave through 2032, and we expect them to expand their commitments over time as generative AI finds more use cases across our everyday lives.

We also think that as the price per GPU hour becomes cheaper, it will enable long-tail demand where organizations start to introduce dedicated GPU clouds for proprietary model training or fine-tuning. There are many benefits to owning a proprietary large language model, such as better control over model behavior, context awareness, and security. The long-term total cost of ownership of training and operating a proprietary model can also become cheaper than buying tokens from an external provider when token usage is poised to see exponential growth. However, the high price of proprietary model training or fine-tuning makes it a non-starter for many organizations at the moment. Hiring just one AI scientist would cost a few million dollars a year, and that is a small expense compared with the actual compute bill companies can expect to receive. In the current environment, we see a limited number of examples where a proprietary model would make sense, but more use cases should start to appear as the price of GPU compute goes down, which supports our assumption that CoreWeave’s annual revenue can scale above $100 billion eventually.

We think that CoreWeave’s adjusted operating margin can reach 20%-25%, which is similar to other AI cloud providers. We expect the company to deliver this level of margin around 2030 if it can scale revenue rapidly. That said, our fair value estimate is largely driven by terminal assumptions. We use an 11.4% weighed average cost of capital, an 8.5% stage II earnings before interest growth rate, and a 14.2% return on new invested capital, consistent with the trend we see in the final years of our Stage I forecast and our terminal forecast for other neoclouds.

We do not think CoreWeave has an economic moat at the moment. CoreWeave is one of the four leading neocloud companies (CoreWeave, Nebius, Crusoe, Lambda Labs) that provide essential infrastructure, in most cases Nvidia GPUs, for AI computing workloads. The AI cloud infrastructure market is extremely competitive, with over 200 companies offering similar computing infrastructure. Players like Amazon, Microsoft, Alphabet, and Oracle are much larger than CoreWeave. Even in a scenario where CoreWeave successfully scales its annual revenue to $100 billion and more over the next decade, company-level return on invested capital still struggles to surpass its cost of capital.

AI’s demand for compute is unlike any new technology we have seen before. Major tech companies are all racing to expand their data center capacity for both internal use cases and external needs. Despite the massive investment, demand continues to outstrip supply for all established hyperscalers, which gave rise to neoclouds. These companies are building entirely new computing infrastructure optimized for AI-related workloads that center around GPUs, from data center design to software orchestration.

A direct benefit of neoclouds’ optimized infrastructure is better computing efficiency. The performance of GPU instances from hyperscalers is often limited by their Ethernet-based networking equipment and the mandatory use of virtual machines. When CoreWeave builds its AI-native infrastructure, the company can avoid these bottlenecks by adopting Nvidia’s Quantum InfiniBand network and offering computing resources in a bare-metal format that removes the need for virtualization. The result is an increase in model FLOPS utilization, or MFU, by around 20%. Such efficiency gains directly translate to tens of millions of savings and faster time-to-market for the most demanding frontier model training projects. Hardware-based efficiency gains and lean software orchestration facilitate a durable edge in bare-metal compute delivery efficiency for leading neoclouds like CoreWeave. That said, better efficiency alone does not lead to a structural cost advantage against competitors. We need stronger evidence that CoreWeave can systematically deliver compute more cheaply, leveraging tools like custom silicon, before assigning the company an economic moat based on cost advantage.

Currently, we think the performance difference that CoreWeave can provide should only matter to a very small group of companies and AI labs that are spending hundreds of millions, or billions, of dollars on AI computing every year. These are the customers that can truly benefit from the sizable time and money savings of using CoreWeave’s AI infrastructure. For the majority of enterprises, their AI-related spending is too small to justify the extra cost and work of introducing a new class of specialized hardware for a very narrow set of tasks. As the cost of GPU computing continues to go down and enterprise AI demand sees exponential growth over the next few years, we think CoreWeave’s total addressable market can expand dramatically.

The majority of the MFU gains that CoreWeave achieved are a result of Nvidia’s technology. CoreWeave has a very close relationship with Nvidia, which led to investor scrutiny. As of early 2026, Nvidia is the second-largest shareholder of CoreWeave with approximately 11.5% of the stake. The financial resources from Nvidia allowed CoreWeave to purchase Nvidia’s compute and networking hardware for its data centers. Besides acting as CoreWeave’s supplier and financier, Nvidia is also a customer of CoreWeave that has $1 billion purchase commitments to CoreWeave’s cloud capacity.

We think both CoreWeave and Nvidia are fully committed to this partnership, which gives CoreWeave the priority to deploy the latest generation of Nvidia GPUs as soon as they become available. Nvidia may view the CoreWeave platform as a reference AI infrastructure that showcases its best AI technology, from the GPUs to BlueField DPUs and the Quantum InfiniBand network, and CoreWeave should continue to benefit from Nvidia’s dominance and high customer switching costs in AI as it sees rapid revenue expansion over the next decade.

To a lesser extent, CoreWeave’s proprietary software, led by its Mission Control and managed Kubernetes offerings, also augmented the system’s performance for AI computing jobs. The company is actively expanding its application ecosystem by setting up a venture fund and acquiring Weights & Biases, an AI development platform. We think these are nice initial steps toward an expanded software ecosystem that can eventually lead to intangible assets and network effects. That said, the process of establishing an AI-native software ecosystem similar to today’s AWS Marketplace can take CoreWeave a decade, if not longer. Established AI labs typically use their own software stack for model training, and most software products essential to the AI workload are open source at the moment. We see a niche market where CoreWeave’s software can help the company attract more customers, but we would not count on its software ecosystem to build a moat for the company anytime soon.

CoreWeave’s deep Nvidia integration and outstanding software orchestration make it a go-to choice for customers looking for best-of-breed GPU clouds, supporting premium prices over other neoclouds.

CoreWeave is making an effort to expand its software ecosystem, which can become a new growth point as new offerings become available.

Reduction in the unit price of GPU compute should unlock more use cases around large language models, expanding CoreWeave’s total addressable market.

The long-term demand for AI cloud infrastructure is highly uncertain. If CoreWeave is unable to ramp up revenue rapidly, it might default on its debt, leaving its equity stake worthless.

CoreWeave needs to overcome multiple supply chain challenges to deliver the contracted capacity on time. Any delays in capacity delivery can negatively impact its growth.

CoreWeave’s future expansion plan can fall apart if the company is not able to secure the funds it needs for its cloud buildout.