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

What It Covers

Four AI infrastructure and software CEOs — CoreWeave's Michael Intrator, Perplexity's Aravind Srinivas, Mistral's Arthur Mensch, and IREN's Daniel Roberts — speak at NVIDIA's GTC conference about GPU financing structures, agentic computing, enterprise AI deployment, open-source model specialization, and the physical infrastructure constraints shaping the next decade of AI development.

Key Questions Answered

• •GPU Depreciation Reality: CoreWeave uses a six-year depreciation schedule for GPUs, and average customer contracts run five years. The claim that GPUs become obsolete in 16–18 months is driven by short-sellers, not operational data. Older Ampere A100s have actually appreciated in price through the year as new companies with smaller models enter the market and absorb previously unavailable capacity at lower cost points.
• •Project Finance "Box" Structure: CoreWeave finances GPU infrastructure by isolating each client deal into a discrete special-purpose vehicle containing the customer contract, GPU assets, and data center lease. Cash flows in a waterfall — paying power, interest, and principal first — with surplus returning to CoreWeave. Within 2.5 years of a 5-year deal, all principal and interest is repaid. This structure enabled $35B raised in 18 months and reduced cost of capital by 600 basis points.
• •AI Orchestration as Competitive Moat: Perplexity's strategy centers on model-agnostic orchestration — routing queries across GPT, Claude, Gemini, Kimi, Qwen, and others based on task specialization. A "Model Council" feature runs the same prompt across multiple models simultaneously and surfaces where they agree, disagree, and diverge in nuance. This positions Perplexity as infrastructure-layer neutral, capturing value regardless of which frontier model wins.
• •Hybrid Local-Server AI Architecture: Perplexity's "Personal Computer" product synchronizes server-side agent orchestration with a local Mac Mini, running privacy-sensitive tasks — tax data, personal notes, local files — on-device while delegating complex long-running tasks to server infrastructure. Users install one executable with no API key management. This hybrid model addresses enterprise and consumer privacy concerns while maintaining access to frontier model capabilities on demand.
• •Enterprise AI Requires Data Governance Primitives: Deploying AI agents across enterprise data requires a "context engine" — a semantic map of data sources tagged with access permissions — to prevent compensation data, HR records, or confidential IP from flowing to unauthorized employees or external systems. Mistral deploys forward engineers on-site at client facilities, keeping all training data within the customer's own infrastructure with zero data flowing back to Mistral's servers.
• •Power Arbitrage as Data Center Strategy: IREN secures renewable energy by co-locating data centers at the generation source — wind and solar in West Texas, hydro in British Columbia — rather than near population centers. West Texas has 45–50 gigawatts of wind and solar but only 12 gigawatts of transmission capacity to load centers. Data centers monetize stranded renewable energy locally and export compute output at the speed of light, eliminating transmission infrastructure costs entirely.
• •Jevons Paradox Drives Compute Demand: Faster, cheaper inference does not reduce total compute consumption — it expands it. As image generation drops from two minutes to five seconds with 10x more available compute, users generate exponentially more images. Every software efficiency gain lowers the cost per token, which induces new use cases, new users, and new applications that consume more aggregate compute than existed before, creating a self-reinforcing demand cycle with no visible ceiling.