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

This GTC episode brings together CoreWeave’s Michael Intrator, Perplexity’s Aravind Srinivas, Mistral’s Arthur Mensch, and IREN’s Daniel Roberts for a concrete look at where AI value is accumulating.

Key points

• CoreWeave’s advantage came from repeatedly repurposing the same underlying asset. It moved GPUs from crypto to rendering, medical research and eventually AI, even donating early A100 capacity to researchers because learning how large-scale neural-network infrastructure worked was worth the cost.

• CoreWeave turned financing itself into a competitive advantage. Long-term customer contracts, GPUs and data-center commitments are bundled into independent financing “boxes”; the company says this structure helped it raise $35 billion in 18 months and lower its cost of capital by 600 basis points.

• Perplexity is expanding from answering questions to doing work. Srinivas describes a progression from accurate internet-connected answers, to browser automation, to “Computer” — an orchestration layer using many specialized models and agents to execute entire tasks.

• Perplexity’s enterprise business may become more important than consumer growth. Enterprise revenue is growing faster than consumer revenue, with $40 and $400 monthly tiers, usage-based charges beyond credits, and Srinivas saying every dollar of revenue currently carries positive gross margin.

• Mistral sees open models as infrastructure for customization, not simply ideology. Enterprises can combine proprietary data, decades of internal IP and specialist knowledge with models they can modify, deploy privately, run on their own hardware and tailor to specific industries.

• Enterprise agents require much more control than consumer agents. Mensch argues that mission-critical automation needs deterministic gates, observability, sandboxes and role-based access controls. Connecting agents directly to company data also changes information flows enough to potentially reshape management itself.

• IREN’s early bet was really on scarce physical infrastructure. Bitcoin initially generated cash flow while it accumulated land, grid connections and data centers; today it controls 4.5 GW of power capacity, while a $9.7 billion Microsoft contract represents only about 5% of capacity.

• AI’s digital growth is colliding with physical bottlenecks. IREN’s 750 MW Texas site illustrates the new constraint: power, construction, cooling, memory, fiber and skilled tradespeople. Roberts describes the job as permanent “whack-a-mole” against real-world limits while demand compounds exponentially.

The central idea

The AI stack is becoming a competition over much more than who has the best model.

The companies gaining leverage are controlling scarce infrastructure, capital, distribution, proprietary data and automated workflows — turning those advantages into systems that become harder to replicate as AI demand scales.