CoreWeave interview: The demand for AI seems to grow every day

Original title: Ai Demand Seems to 'Intensify' Every Day

Photo by Tae Kim

Original by Peggy, Block Beats

THE EDITOR PRESSES: THIS INTERVIEW PROVIDES A WINDOW TO OBSERVE THE AI CALCULATOR CYCLE: DEMAND IS NOT COOLED BY THE LAST GPU BUYOUT, BUT IS BEING PUSHED UP BY SMARTS, REASONING AND ENTERPRISE-LEVEL AI APPLICATIONS。

This paper interviewed Brannin McBee, co-founder and chief development officer of CoreWeave, and Nick Robbins, Vice-President of Enterprise Development and Investor Relations, to discuss the current situation in the AI market. CoreWeave’s core statement is straightforward – AI seems to be increasing in a new way every day, and the real bottleneck is shifting from “the existence of a GPU” to more complex infrastructure issues: data centres’ electricity casings, CPUs, storage, electricians, supply chain implementation capabilities, and the willingness of clients to pay the high cost of a new generation of algorithms。

CoreWeave is unique in that it is in the middle of the AI infrastructure chain: it serves front-line clients such as Openai, Anthropic, Meta, Google, Microsoft, Nvidia, as well as directly senses changes in the needs of research laboratories, business clients and super-large cloud manufacturers. So what it sees is not just a "gPU shortage," but an AI load itself is changing structurally. With the rise of the agentic AI and resoning models, the need for computing is no longer confined to GPU, and the importance of CPU and storage is increasing, and the design of a new generation data centre must leave space for Vera CPU, Vera Rubin servers and more storage。

This explains why AI infrastructure competition is shifting from simple chip purchases to a more comprehensive project delivery capability. Those who have faster access to the power data centre, deploy servers, access the supply chain and optimize each token cost are closer to the core of the AI capital expenditure cycle. CoreWeave has repeatedly emphasized "customer-driven" behind it is a much larger judgement: AI cloud manufacturer is no longer just a salesman, but is rebuilding the next generation of AI factories ahead of schedule, based on the front-line customer road map。

For investors and industry observers, the most interesting part of this interview was not a single-point number, but rather the direction of AI’s infrastructure needs: GPU remains important, but bottlenecks are spreading; Nvidia is still at the core, but CPU, HBM, storage and data centre power is becoming a new variable; and AI’s demand is still growing, but future success and failure may depend on who delivers complex infrastructure on a sustained, stable and scale。

The following is the original text:

CoreWeave is seen as an innovative early market leader in the area of neocloud。

It is the only cloud servicer who has received the highest-level "Platinum Rating" of the IAI research institute, SemiAnalysis. CoreWeave was established in 2017 to provide large-scale GPU computing for start-ups and large-scale enterprises。

Key Context recently interviewed Brannin McBee, Co-founder and Chief Development Officer of CoreWeave, and Nick Robbins, Vice-President of Enterprise Development and Investor Relations, on the current situation of the AI demand and the neocloud market。

The following are the editorial highlights of the dialogue:

AI DEMAND CONTINUES TO INCREASE

Tae: When did the AI demand wave start

Brannin: We saw the real beginning last quarter. At that time, we were communicating at the engineering level with our clients about the products they expected to market in the first quarter of this year。

This has been a very important perspective when we look at client needs. There is a deep, interconnected engineering relationship between us and our clients. It is this relationship that allows us to see trends in advance, rather than to react passively when changes occur。

IF FROM THE POINT OF VIEW OF THE PRODUCTS OF THE AI MARKET, I WOULD SAY THAT THE FIRST QUARTER WAS A TIME WHEN REASONING AND AI CONSUMPTION WERE TURNING HUGE, AND THIS ACCELERATION CONTINUES TO THIS DAY。

Tae: What is the current status of AI needs? Have there been no signs of a slowdown in recent weeks compared to a few months ago

Nick: It seems to be growing in new ways every day。

Tae: Please describe the upward trend in the demand for CPUs relative to GPUs in the AI wave of smarts. Will you deploy a platoon of Vera CPU racks next to the Nvidia GPU server

Brannin: CoreWeave has been running CPU since 2023. We've always had complete cloud products. So the question is not whether we're just starting to add CPUs, but what does the client really need? Is this demand in relative terms on the rise? The answer is very clear, indeed。

As intelligent bodies and reasoning are actually emerging in models, storage needs are rising in comparison with previous generations. I think this trend will continue。

Nick: The answer to your question is yes. You'll definitely see a lot of Vera CPU deployed next to a lot of Vera Rubin servers. Last year, we actually fundamentally redesigned the basic data centre programme to allow for more storage and more CPUs to be deployed next to the GPU。

We do so because we are in a very unique position across the ecosystem. We are the only independent cloud service provider serving all the most advanced technology users. No other independent AI cloud service provider can say that Anthropic, OpenAI, Meta, Google, Microsoft, Nvidia and others are their own customers。

This has created a useful wheel for our business, or a feedback loop: we can understand where clients are taking technology and plan accordingly。

THE BOTTLENECKS ARE NO LONGER JUST GPU

Tae: Will you use Nvidia Vera CPU in the future

Nick: This depends on the specific workload. We do things on the basis of client demand. We do expect to be an early and important adopter of Vera CPU, as we have already revealed. At present, our fleet is actually predominantly AMD, but over time, this may change according to client needs. Client interest in Vera CPU is very strong。

Brannin: It's also a good reminder of how our contract works. As you know, 98% of our income is contract-driven. We're not guessing what kind of infrastructure clients want. Clients will tell us very clearly what configuration they need. Everything is client driven. It's the customer that defines what we're building。

Tae: Talk about competition patterns. How did you get on the market and compete with a huge cloud service provider like SpaceX, Nebius, Oracle, and Azure, AWS, Google

Brannin: I prefer to look at third-party verification in terms of differentiation. Apart from China, nine of the top 10 AI laboratories in the world are using our platform. SemiAnalysis has always singled out us as the highest level of performance. I don't think that the reason we get this GPU assignment is because we have a personal relationship with Jensen。

This demonstrates the strong confidence of suppliers in our implementation record and engineering capacity, which we can best demonstrate globally。

Nick: We can win a huge cloud service provider because we're very good at it. We can build these systems very quickly, and they work very well. We can win research lab clients because we provide the best technical versions and perform the best in every token。

We have been able to win corporate clients because the infrastructure is indeed working well, and we have built a very good, comparable and optimal hierarchy, which is a recognized source such as platinum ratings。

BUT INCREASINGLY, AMONG AI CLOUD SERVICE PROVIDERS, WE HAVE BUILT THE MOST SOPHISTICATED LAYER OF CAPACITY, COVERING REASONING AND DEVELOPING TOOLS TO HELP BUSINESSES ACTUALLY PUT AI INTO PRODUCTION。

This means that we are building and delivering products that will ultimately help enterprises that are relatively less technologically mature to transform data into models and into in-house smarts, and that we can also cross-market CoreWeave cloud services in the process。

Tae: What are the current bottlenecks? Is it a data centre shell that already has electricity? GPU

Brannin: It's powered shell, a data centre shell with electricity. More precisely, it is a component inside the shell. You specifically mentioned the electrician, which is perfectly correct. This is a complex area。

But the important thing is that we already have 49 such sites on line and running. We are not placing hope on one or two stations. We've done it 49 times。

This is a very deep implementation record。

This also means that we have accumulated a great deal of knowledge about how to deal with supply chain issues, about which suppliers in this supply chain are fit to cooperate and which suppliers are not。

editor: powered shells refer to the data centre building itself, excluding actual computing server hardware。

Tae: What can you tell us about the costs and shortages of HBM memory? How do you respond? Do clients need to bear the cost of price increases

Nick: The answer is yes. Our business model is designed to lock down the GPU prices we charge our customers while signing GPU purchase orders and determining how much we pay. More broadly, the price of the server, which clearly includes HBM costs。

That is how we isolate the effects of daily price fluctuations。

If the cost of our components increases in the next deal, we will reflect it in the price that we think can be charged to the client, thereby protecting our profitability. We are well protected in channelling these costs to our clients. This is something that we are following very closely。

currently, access to components is not the biggest bottleneck. the biggest bottleneck is powered shell. but at some point in the future, the answer may change。

Tae: How do you expect Vera Rubin's deployment to climb the slope? What will happen in the second half of this year

Nick: We are obviously the first company in the world to start and fully validate VR, or Vera Rubin's cabinet. Last year we did the same on GB200, GB300. I expect VR to start showing up later this year。

I EXPECT A TRULY LARGE, VERY ROBUST DEPLOYMENT TO CLIMB THE SLOPES THROUGHOUT 2027. THIS RHYTHM IS SIMILAR TO THE GB: THE GB BEGAN TO APPEAR IN 2025, BUT THE REAL LARGE-SCALE RAMP ACTUALLY RUNS THROUGH 2026. THAT IS TO SAY, THERE WERE A LOT OF DEPLOYMENTS AT THE END OF LAST YEAR, BUT THIS IS THE YEAR OF A TRULY MASSIVE DEPLOYMENT OF GB。

I EXPECT IN THE NEXT 12 TO 18 MONTHS, THE VR WILL HAVE A VERY SIMILAR RHYTHM。