「英伟达概念股」CoreWeave联创访谈：AI需求似乎每天都在加剧

Original title: An Interview with CoreWeave Executives: AI Demand Seems to 'Intensify' Every Day

Original author: Tae Kim

Original translation: Peggy, BlockBeats

Editor's note: This interview offers a window into the AI computing cycle: demand has not cooled due to the previous wave of GPU purchases; instead, it is being further driven by agents, inference, and enterprise-grade AI applications.

This article features an interview with Brannin McBee, Co-founder and Chief Development Officer of CoreWeave, and Nick Robbins, Vice President of Corporate Development and Investor Relations, discussing the state of AI demand and the neocloud market. The core message from CoreWeave executives is straightforward — AI demand seems to intensify in new ways every day, and the real bottleneck is shifting from "whether GPUs are available" to more complex infrastructure issues: data center powered shells, CPUs, storage, electricians, supply chain execution capabilities, and how much customers are willing to pay for next-generation computing power.

CoreWeave's uniqueness lies in its position at the center of the AI infrastructure chain: it serves top-tier clients like OpenAI, Anthropic, Meta, Google, Microsoft, and Nvidia, while also directly sensing demand shifts from research labs, enterprise customers, and hyperscale cloud providers. Therefore, it sees not just whether GPUs are in short supply, but a structural transformation in AI workloads themselves. With the rise of agentic AI and reasoning models, computing demand is no longer centered solely on GPUs; the importance of CPUs and storage is also growing. Next-generation data center designs must reserve space for Vera CPUs, Vera Rubin servers, and additional storage.

This also explains why the competition in AI infrastructure is shifting from simple chip procurement to more comprehensive engineering delivery capabilities. Whoever can secure powered data centers faster, deploy servers, streamline supply chains, and optimize cost per token is closer to the core of this AI capital expenditure cycle. CoreWeave repeatedly emphasizes being "customer-driven," which actually reflects a larger judgment: AI cloud providers are no longer just selling computing power; they are preemptively restructuring the next generation of AI factories based on the roadmaps of the most cutting-edge customers.

For investors and industry observers, the most noteworthy aspect of this interview is not a single data point, but the direction of change in AI infrastructure demand: GPUs remain important, but bottlenecks are spreading; Nvidia remains core, but CPUs, HBM, storage, and data center power capacity are becoming new variables; AI demand continues to grow, but future victories may depend on who can deliver complex infrastructure consistently, stably, and at scale.

The following is the original text:

CoreWeave is regarded as an innovative early market leader in the neocloud space.

It is the only cloud provider to receive the highest tier "Platinum Rating" from the AI research firm SemiAnalysis. Founded in 2017, CoreWeave provides large-scale GPU computing power for startups and large enterprises.

Key Context recently interviewed Brannin McBee, Co-founder and Chief Development Officer of CoreWeave, and Nick Robbins, Vice President of Corporate Development and Investor Relations, to discuss the state of AI demand and the neocloud market.

Below are edited highlights from the conversation:

AI Demand Continues to Intensify

Tae: When did the wave of demand for agentic AI begin to erupt?

Brannin: We saw the real beginning in the fourth quarter of last year. At that time, we were having engineering-level discussions with customers about products they expected to bring to market in the first quarter of this year.

This has always been a very important perspective in how we view customer demand. We have a deeply interconnected engineering relationship with our customers. It's this relationship that allows us to see trends in advance, rather than reacting passively after changes occur.

If I look at it from the perspective of AI market products, I would say that Q1 was a moment of a massive inflection point for inference and AI consumption, and this acceleration continues even now.

Tae: What is the current state of AI demand? Compared to a few months ago, is there absolutely no sign of a slowdown in recent weeks?

Nick: It seems to intensify in new ways every day.

Tae: Can you talk about the rising demand for CPUs relative to GPUs amid the agentic AI wave? Will you deploy rows of Vera CPU racks next to Nvidia GPU servers?

Brannin: CoreWeave has been running CPUs since 2023. We have always had a full cloud product line. So, the question isn't whether we are just starting to add CPUs, but rather what the customer actually needs. Is this demand rising in a relative sense? The answer is yes, very clearly.

As agents and reasoning capabilities truly emerge in models, storage demand is also rising compared to previous generations. I believe this trend will continue.

Nick: The answer to your question is yes. You will absolutely see a large number of Vera CPUs deployed alongside a large number of Vera Rubin servers. Last year, we fundamentally redesigned our base data center design to make room for more storage and more CPUs to be deployed next to GPUs.

We did this because we are in a very unique position within the entire ecosystem. We are the only independent cloud provider serving all the most advanced technology users. No other independent AI cloud provider can say that Anthropic, OpenAI, Meta, Google, Microsoft, and Nvidia are all their customers.

This creates a beneficial flywheel, or positive feedback loop, for our business: we can understand where customers are taking the technology and plan accordingly.

The Bottleneck is No Longer Just GPUs

Tae: Will you primarily use Nvidia Vera CPUs in the future?

Nick: It depends on the specific workload. We are driven by customer demand. We do expect to be an early and significant adopter of Vera CPUs, and we have disclosed this. Currently, our fleet is still primarily AMD, but this could change over time based on customer needs. Customer interest in Vera CPUs is very strong.

Brannin: This is also a good reminder to talk about how our contracts work. As you know, over 98% of our revenue is contract-driven. We are not guessing what kind of infrastructure customers want. Customers tell us very clearly what configurations they need. Everything is customer-driven. The customers define what we build.

Tae: Talk about the competitive landscape. How do you go to market and compete against neoclouds like SpaceX, Nebius, Oracle, and hyperscalers like Azure, AWS, and Google?

Brannin: In terms of differentiation, I prefer to look at it from a third-party verification perspective. Nine out of the top ten AI labs globally, excluding China, use our platform. SemiAnalysis has consistently rated us alone at the highest level in terms of performance. I don't believe we get our GPU allocation because of personal friendships with Jensen.

This shows that suppliers have deep confidence in our execution track record and engineering capabilities, trusting us to best represent their product capabilities on a global scale.

Nick: We were able to win hyperscaler customers because we are very good at execution. We can build these systems very quickly, and they run very well. We were able to win research lab customers because we provide the most powerful version of the technology with the best performance per token.

We were able to win enterprise customers because our infrastructure works well, and we have built an excellent, best-in-class orchestration layer, which is also a source of recognition like the Platinum Rating.

However, what is increasingly important is that, among AI cloud providers, we have built the most mature layer of capabilities covering inference and development tools, helping enterprises put AI into production.

This means we are building and delivering products that ultimately help enterprises with lower technical maturity turn data into models, and models into agents that can run internally, while we can cross-sell CoreWeave cloud services in the process.

Tae: What are the current bottlenecks? Is it powered data center shells? GPUs? Or electricians?

Brannin: It's powered shells. More precisely, it's the components inside these shells. You specifically mentioned electricians, and you are absolutely right. It's a complex field.

But importantly, we already have 49 such sites up and running. We are not pinning our hopes on just one or two sites. We have done it 49 times.

This is a very deep execution track record.

It also means we have accumulated a tremendous amount of knowledge on how to handle supply chain issues, which suppliers in this supply chain are good to work with, and which are not.

Editor's note: Powered shells refer to the data center building itself, excluding the actual computing server hardware.

Tae: Can you share anything about the cost and shortage of HBM memory? How are you dealing with it? Do customers have to bear the cost increases?

Nick: The answer is yes. Our business model is designed so that when we sign a GPU purchase order and determine how much we are going to pay, we simultaneously lock in the GPU price we charge our customers. More broadly, that's the server price, which obviously includes the HBM cost.

This is how we insulate ourselves from the impact of daily price fluctuations.

If our component costs rise in the next transaction, we reflect that cost in the price we believe we can charge our customers, thereby protecting our margins. We are well-protected in passing these costs on to customers. This is something we monitor very closely.

Currently, getting components isn't the biggest bottleneck. The biggest bottleneck is powered shells. But at some point in the future, this answer could change back and forth.

Tae: How do you expect the Vera Rubin deployment ramp to unfold? What will the situation be like in the second half of this year?

Nick: We were clearly the first company globally to start and fully validate the VR, or Vera Rubin, cabinet. We did the same with GB200 and GB300 last year. I expect VR to start appearing later this year.

I expect the truly large-scale, very strong deployment ramp to span the entire year of 2027. This cadence is similar to GB: GB started appearing in 2025, but the real large-scale ramp actually spanned the entire year of 2026. That is, a significant amount was deployed late last year, but this year is truly the year of large-scale GB deployment.

I expect VR to follow a very similar cadence over the next 12 to 18 months.