Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN

This summary synthesizes a panel discussion featuring four CEOs from leading AI companies: CoreWeave, Perplexity, Mistral AI, and IREN. The conversation, taking place at Nvidia's GTC conference, delves into the evolving landscape of AI infrastructure, the development of AI models, the business models supporting this growth, and the future trajectory of the industry. It offers insights into how these companies are addressing the immense demand for compute, innovating in financing and deployment, and shaping the accessibility and capabilities of AI for both consumers and enterprises.

CoreWeave: Building the Foundational Infrastructure for AI

Michael Intrator, CEO of CoreWeave, details the company's journey from its origins in algorithmic hedge funds and cryptocurrency mining to becoming a hyperscale provider of GPU compute for AI.

• Origins in Compute Flexibility: CoreWeave began in 2017, initially mining crypto. However, their focus quickly shifted to the underlying compute power of GPUs, recognizing their versatility beyond mining for applications in CGI rendering, medical research, and batch computing. This early understanding of compute's potential allowed them to weather crypto winters and scale their operations.
• Learning Through Experimentation: The company invested in A100 GPUs and donated them to an open-source AI project (Ought AI). This "tuition" for learning how to run large-scale, parallelized computing provided invaluable insights into the infrastructure and software requirements for AI development. The positive feedback from researchers using this infrastructure directly led to the launch of CoreWeave's dedicated AI compute business.
• Scaling for AI Models: CoreWeave recognized that "computing decommoditizes at scale." While individual GPUs are accessible, the ability to train world-changing models requires massive clusters. Their strategy shifted to scaling their compute delivery to larger and more sophisticated clients, focusing on the infrastructure and software layers between GPUs and AI models.
• Inference as Monetization: Intrator views inference—when AI models provide answers to user queries—as the primary monetization of AI investment. The massive scale of inference traffic highlights the consumption and deployment of AI, not just its development.
• Leading Edge GPU Deployment: CoreWeave prides itself on being at the forefront of deploying new Nvidia architectures at scale, being among the first to offer H100s, H200s, GB200s, and GB300s. They observe that cutting-edge GPUs are used for training, then repurposed for experiments, and finally for inference, demonstrating their long-term utility.
• GPU Depreciation Debate: Intrator dismisses the notion that GPUs become obsolete quickly, citing that clients sign 5-year contracts. He argues that depreciation is a non-issue as long as customers are willing to pay for the hardware, which they do for extended periods. The appreciation of older GPUs like A100s in price is attributed to new use cases and companies entering the market, utilizing this capacity.
• Innovative Financing Model: CoreWeave has pioneered financing structures, particularly GPU-based loans. They utilize a "box" system where client contracts, GPUs, and data center agreements are pooled. This structure governs cash flow, prioritizing data center costs, power, and debt repayment before returning profits to CoreWeave. This model has enabled them to raise significant capital and demonstrate strong returns, reducing their cost of capital over time.
• Demand Outstrips Supply: Intrator emphasizes that demand for AI compute has been relentlessly high and consistently overwhelms global capacity. This imbalance is driven by the need to enable all the diverse use cases of AI.
• Constraints Beyond GPUs: While GPUs are critical, other constraints like electricity, power, memory, storage, networking, and optics also limit deployment. Memory, in particular, is currently a bottleneck due to cyclical industry dynamics and high demand.
• AI as the New Operating System: Intrator posits that AI is becoming the new operating system, shifting the paradigm from executing programs to achieving high-level objectives. This fundamentally changes how users interact with technology, moving from specific instructions to desired outcomes.