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

CoreWeave (Michael Moritz)

• Business Evolution: CoreWeave originated in 2017 as an algorithmic hedge fund focused on natural gas, pivoted to cryptocurrency mining (Ethereum), then migrated to CGI rendering, medical research, and finally high-performance AI infrastructure for hyperscalers.
• Market Positioning: The company positions itself "above NVIDIA GPUs but below the models," offering a purpose-built cloud that handles software integration, observability, and operations specifically for training and inference.
• Client Base: Key commercial clients include Inflection (Mustafa Suleyman), Microsoft, OpenAI, and various hyperscalers; the business model relies on long-term contracts averaging five years.
• Hardware Lifecycle: Moritz disputes the industry narrative of 16–18 month GPU obsolescence, stating CoreWeave uses a six-year depreciation schedule based on client contracts and the fact that older GPUs (like A100s) are appreciating in value for secondary markets and inference tasks.
• Financing Model ("The Box"): CoreWeave utilizes a securitized financing structure where client contracts (e.g., with Microsoft) are placed in a special-purpose vehicle ("the box") to secure debt; this vehicle governs cash flow, paying data center costs, interest, and principal first, with returns flowing to CoreWeave.
• Capital Efficiency: The financing model has reduced CoreWeave's cost of capital by 600 basis points over the last two years, allowing them to raise $35 billion in 18 months to scale infrastructure.
• Hardware Supply Chain: Access to NVIDIA hardware (H100, H200, GB200) is allocated by order date without favoritism; wait times are long due to massive global demand, not just chip scarcity.
• Constraints: The primary bottlenecks beyond GPUs are electricity, fiber optics, memory, and storage; memory supply is currently cyclical, while energy availability is driving data center location strategies.
• Inference Trends: Moritz identifies inference as the monetization phase of AI investment, noting a massive shift from research-focused compute to commercial-scale deployment where models are actively used to solve real-world problems.

Perplexity AI (Arvind Srinivasan)

• Product Philosophy: Perplexity is evolving from a search engine into "Perplexity Computer," an AI operating system that aggregates multiple specialized models (an "orchestra") to perform complex, multi-step tasks autonomously.
• Enterprise Growth: The enterprise segment is the company's fastest-growing revenue stream, outpacing consumer growth, with "Enterprise Max" plans priced at $400/month; the company reports positive gross margins on all revenue but has not yet reached overall profitability.
• Strategic Differentiation: As a neutral platform, Perplexity avoids dependency on a single model provider (unlike Microsoft/OpenAI or Google), allowing it to route queries to the best-performing model for each specific task (e.g., Claude for writing, Code Llama for coding).
• Autonomy & "Computer" Feature: The "Computer" agent can natively control browsers (via Comet), access Google Workspace/Slack, and execute complex workflows like generating CRMs or analyzing podcasts, effectively abstracting the need for manual human input.
• Local-Cloud Hybrid: To address privacy and security, Perplexity is launching "Perplexity Personal Computer," a solution that synchronizes with local hardware (e.g., Mac Mini) to run sensitive data processing locally while delegating complex tasks to server-side agents.
• Revenue & Pricing: The company charges based on token usage and subscriptions; Pro is $20/month, Enterprise Pro is $40/month, and Enterprise Max is $400/month, with no bundled subscriptions for third-party AI models.
• Job Market Impact: Srinivasan views AI as a tool for "rugged individualism" that lowers barriers to entry for solo entrepreneurs, potentially displacing some roles but creating opportunities for high-value, low-cost one-person businesses.
• Developer Tools: The release of "Comet" and "Computer" allows non-engineers to build applications via natural language, significantly accelerating iteration cycles and enabling the creation of bespoke software without traditional coding.

Mistral AI (Arthur Mensch)

• NVIDIA Partnership: Mistral announced a partnership with NVIDIA to train the next generation of frontier open-source models, leveraging NVIDIA's hardware to produce specialized models for enterprise sectors like finance, engineering, and physics.
• Open Source Strategy: The company focuses on open models to allow enterprises to customize, deploy on-premise, or edge-deploy models, ensuring data sovereignty and the ability to integrate proprietary industrial intellectual property (IP) that cannot be used with closed-source APIs.
• Data Securitization: To serve high-security industries (e.g., ASML, banking), Mistral deploys its training platform directly onto customer infrastructure, ensuring data never leaves the client's environment while their engineers provide domain expertise to refine the models.
• Synthetic vs. Human Data: While Mistral uses synthetic data for model warm-up and compression, Mensch asserts that high-quality human signals from domain experts are ultimately required for training models in mission-critical verticals.
• Enterprise Constraints: The industry shift toward autonomous agents requires new "context engines" and "sandbox" architectures to handle data segregation (e.g., preventing HR data from leaking to engineering teams) and ensure observable, deterministic governance for compliance.
• Market Outlook: Mensch believes that while general models will handle orchestration, specialized vertical models built on open infrastructure will win in the enterprise space due to the need for deep customization and IP protection.

Iron Mountain / Iron (Daniel Roberts)

• Energy-First Strategy: Iron built its business model around securing massive amounts of renewable energy (hydro, wind, solar) in remote locations (e.g., West Texas) before securing AI customers, flipping the traditional data center model of following load centers.
• Scale & Capacity: The company has secured 4.5 gigawatts of power capacity, enough to power the entire Bay Area, with flagship sites like a 750-megawatt Texas facility; Microsoft has signed a $9.7 billion contract for 5% of their capacity.
• Power Constraints: While NVIDIA GPUs are the most visible constraint, the primary bottleneck for Iron is "time to compute"—securing grid connections and skilled labor to build out infrastructure quickly.
• Workforce Dynamics: Iron is driving a resurgence in trade jobs, with local electricians and construction workers earning salaries between $150,000 and $300,000, often requiring a 3-month "tour of duty" in remote towns.
• Grid & Sustainability: Iron leverages existing utility grids to handle the variability of renewable energy (the "duck curve"), ensuring 24/7 reliability; they do not need to build massive battery farms themselves as the utility underwrites the consistency.
• Jevons Paradox: Roberts argues that as compute efficiency improves and costs drop, demand will not saturate but will instead expand exponentially (e.g., faster image generation leading to far more content creation).
• Custom Silicon: While Google, Amazon, and Meta are developing custom chips, Iron believes NVIDIA remains the safest path for rapid scaling due to its established ecosystem, though custom silicon will play a growing role over time.
• Infrastructure Architecture: The industry is shifting toward "The Data Center is the New Computer," where the focus is on the entire building's fabric, including high-speed cabling (InfiniBand/Ethernet) and low-latency connectivity to maximize cluster performance.
• Nuclear Energy: Roberts is actively tracking Small Modular Reactors (SMRs) as a future solution for unlocking clean, dense power near data centers, noting that while deployment takes a decade, the strategic conversation must begin now.