An Interview with CoreWeave Executives: AI Demand Seems to 'Intensify' Every Day
Key Context spoke with CoreWeave co-founder Brannin McBee and VP of Corporate Development Nick Robbins to discuss the state of AI demand.
CoreWeave is regarded as the innovative early market leader of the neocloud sector.
It is the only vendor to earn a top Platinum-tier rating from AI research firm SemiAnalysis. Founded in 2017, the company provides large-scale GPU capacity to startups and enterprises.
Key Context recently spoke with CoreWeave co-founder and Chief Development Officer Brannin McBee and VP of Corporate Development and Investor Relations Nick Robbins to discuss the state of AI demand and the neocloud market.
Here are lightly edited highlights from our conversation:
Tae: When did the agentic AI demand wave take off?
Brannin: We saw the real beginnings of that in Q4 as we were having the engineering conversations with our clients of what they were expecting to bring to market in Q1.
That’s a really important lens that we consistently have with our clients. It’s a deep interconnected engineering relationship. That allows us to be ahead of where everything is going instead of reactive. I would say from a product perspective for the AI market, Q1 was just this massive inflection in inference and AI consumption that only continues to accelerate.
Tae: What is the state of AI demand right now? Would you say there is no letup in the last few weeks versus months ago?
Nick: It finds new ways to intensify seemingly every day.
Tae: Tell us about the rising trend of CPUs versus GPUs in the agentic AI wave. Will you install rows of Vera CPU racks next to Nvidia GPU servers?
Brannin: CoreWeave has been running CPUs since 2023. We’ve had that full cloud product consistently. It’s not, are we beginning to add CPU? It’s more what is the client asking for? And is it increasing on a relative basis? Absolutely, yes. Storage has increased on a relative basis to prior generations as well, as agentic and reasoning have really taken off within the models. I see that trend continuing.
Nick: The answer to your question is yes. You’re absolutely going to see a bunch of Vera CPUs crammed next to a bunch of Vera Rubin servers. Last year, we effectively fundamentally re-architected our base case data center design to leave room for a lot more storage and a lot more CPUs to live next to the GPUs.
The reason we did that is informed by this unique place we sit in this ecosystem. We’re the only independent cloud that serves all the most sophisticated users of this technology. Like there’s no other independent AI cloud that can say, Anthropic and OpenAI and Meta and Google and Microsoft and Nvidia and so on are customers. That’s created this helpful flywheel for our business or positive reinforcement loop where we understand where our customers are taking the technology and we plan accordingly.
Tae: Are you primarily going to use Nvidia Vera CPUs going forward?
Nick: It depends on the workload. We’re customer led in what we do. We definitely expect to be an early and meaningful adopter of Vera CPUs. We’ve already disclosed that. Today, the fleet is actually mostly AMD at the moment, but that probably shifts over time based on customer demand and there’s a ton of interest for Vera CPU.
Brannin: That’s a good reminder just on how our contracts work. As you know, 98 plus percent of our revenue is contractually driven. We aren’t guessing at what clients want for the infrastructure. They’re telling us exactly what builds they’re looking for. It is all client led. They are defining what it is that we are building.
Tae: Talk about the competitive landscape. How do you go to market versus neoclouds like SpaceX, Nebius, Oracle versus hyperscalers like Azure, AWS and Google?
Brannin: The perspective I like to take on differentiation sits with the third-party verification that’s out there. Nine out of the ten top AI labs, excluding China, use our platform. SemiAnalysis consistently ranks us singular in our performance. I don’t think it’s out of our friendship with Jensen that we get allocated GPUs the way we do. That’s a supplier who has deep confidence in our execution track record and engineering capabilities to be the best representation of his product on the planet.
Nick: We win with the hyperscaler because we’re really good at execution. We can stand these things up incredibly quickly and they run very well. We’re winning with the research lab because we deliver the most performant version of the technology and the efficiency on a per token basis. We’re winning with the enterprise because of, yes, the infrastructure works well and we built a really nice orchestration layer that’s best in class and that’s where platinum ranking and all that stuff comes in. But increasingly importantly, it’s because we’ve built, certainly amongst the AI clouds, the most sophisticated top of the layer across inference and development tools that empowers enterprises to productionize AI. What that means is we are building and delivering products that ultimately help enterprises who are less sophisticated take data and turn it into models, turn it into agents that they can then run internally and we can cross-sell CoreWeave cloud into.
Tae: What are the current bottlenecks? Data center powered shells? GPUs? Electricians?
Brannin: Powered shells. It’s the components within powered shell. You’re absolutely correct to highlight electricians. It is a complicated space, but I think what’s important there is we have 49 of these sites online and running. We’re not just putting hopes and prayers into one or two sites. We’ve done this 49 times. That’s an immense amount of execution track record. That’s an immense amount of knowledge that we have built on how to deal with supply chain problems and who are the right vendors and the wrong vendors to be working with within that supply chain.
(Editor’s note: Powered shells are the data center buildings and do not include the actual computing server hardware.)
Tae: What can you tell us about HBM memory costs and shortages? How are you’re dealing with that? Do customers have to pay for the price hikes?
Nick: The answer to that is yes. The way we designed our business model is to lock in what we’re charging for the GPUs at the same time that we’re signing purchase orders for what we will be charged for the GPUs. Or servers more broadly, which obviously includes an HBM component. And that is our way of insulating ourselves from the day-to-day price moves.
If price goes higher on our component price on the next deal, we make that into what we’re comfortable charging the customer, and therefore we protect our margin. We’re very well protected from it in our ability to pass through those costs onto the customer. It’s something that we monitor very closely.
Access to components is not the biggest bottleneck today. Powered shell is. But there may be points in time in the future in which that answer flip flops.
Tae: How do you expect the Vera Rubin ramp to go? How is it going to go in the second half of this year?
Nick: We obviously were the first in the world to bring up and fully validate a VR [Vera Rubin] cabinet. As we were with GB200, GB300 last year, I expect VRs to start to show up later this year. I expect that heavy, heavy ramping to be throughout 2027 in the same way that GB started showing up in 2025, but the heavy, heavy ramp was really throughout 26. Like you had plenty stood up towards the end of last year, but this year has been really ramping GB at scale. I expect there to be a pretty similar pattern for VR over the course of the next 12 to 18 months.
Tae: On the Nvidia vs. ASICs debate. What are the customers saying now? In the past I heard from startups who say they are sticking with Nvidia versus Amazon because Nvidia is the most optimized, most reliable, battle tested. Is it still the case?
