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

Episode Summary

Executive Summary: At NVIDIA GTC, the transcript spotlights three AI infrastructure and application leaders: CoreWeave, Perplexity, Mistral, and IREN. The discussion centers on how AI demand is reshaping compute, financing, product design, and data-center location strategy. Speakers argue that scale, specialization, and orchestration—not just raw model quality—will determine winners, while power, memory, and capital remain the main bottlenecks.

Main Topics: CoreWeave’s evolution from crypto mining to AI infrastructure (Priority: 5/5): Michael Intrator explains how CoreWeave began as a hedge-fund-adjacent crypto mining operation, then expanded into CGI rendering, batch computing, research, and finally large-scale GPU infrastructure for AI training and inference. GPU depreciation, lifespan, and demand durability (Priority: 5/5): Intrator argues that claims of rapid GPU obsolescence are overstated because customers sign multi-year contracts and older chips continue to have value for inference, experiments, and less bleeding-edge workloads. Perplexity’s product evolution into AI orchestration and computer control (Priority: 5/5): Aravind Srinivas describes Perplexity’s progression from search/answering to browser automation and then to 'computer'—a system that orchestrates multiple models and can execute tasks across browser, desktop, and local hardware. Open source, vertical specialization, and enterprise AI at Mistral (Priority: 4/5): Arthur Mensch outlines Mistral’s strategy of training frontier open models with NVIDIA, then specializing them for enterprise use cases through on-prem deployment, data segregation, and forward-deployed engineering. Data-center buildout, power, and location strategy at IREN (Priority: 5/5): Daniel Roberts explains how IREN moved from Bitcoin mining to AI data centers, emphasizing that power availability, grid connections, and proximity to renewable energy sources are now the key constraints. AI’s impact on labor, productivity, and business formation (Priority: 4/5): Across the interviews, the speakers argue that AI will automate repetitive work, reduce the need for some hires, and enable smaller teams and solo founders to build and run businesses more autonomously.

Key Arguments: CoreWeave’s early crypto mining work was effectively 'tuition' that taught the company how to manage risk, capital, and GPU infrastructure before AI demand exploded. GPU depreciation is not collapsing as skeptics claim; if customers are willing to pay for five-year contracts, the hardware still has economic value. Inference is the monetization layer of AI: training creates the model, but inference is where real-world value and revenue are realized. Perplexity’s advantage is multi-model orchestration: it can route tasks to the best model for each job rather than betting on a single frontier model. Enterprise AI requires control planes, access controls, and context engines so sensitive data like compensation or HR information does not leak across the organization. Open models are attractive for enterprises because they can be customized, deployed on-prem or at the edge, and trained with customer-specific data without sending data back to the vendor. AI data centers are increasingly constrained by power, memory, networking, and storage—not just GPUs—and the best sites are where excess energy already exists. AI will likely displace some jobs, but it will also lower barriers to entrepreneurship and allow small teams to run more of a business autonomously. The market for AI compute is still demand-constrained; if more capacity appeared, usage would expand rather than remain flat, consistent with Jevons-style demand expansion.

Data Points: CoreWeave average contract length: 5 years - Intrator says clients typically buy compute on five-year contracts, supporting his argument that GPUs retain value longer than critics claim. CoreWeave depreciation schedule: 6 years - The company uses a six-year depreciation assumption for GPUs. CoreWeave capital raised: $35 billion in 18 months - Intrator cites this as evidence that the financing structure around GPU-backed infrastructure is working. Cost of capital reduction: 600 basis points - CoreWeave says it has lowered its cost of capital over the last two years through its financing structure. CoreWeave contract payback period: 2.5 years - Intrator says a five-year deal pays for the infrastructure within about two and a half years. Perplexity consumer usage: Several tens of millions of users monthly - Srinivas describes the scale of Perplexity’s consumer product. Perplexity team size: About 400 employees - Srinivas says the company remains relatively small despite rapid growth. Perplexity enterprise pricing: $40/month and $400/month - Enterprise Pro is $40/month; Enterprise Max is $400/month. Perplexity enterprise savings: More than $100 million saved - Srinivas says top enterprise customers have saved over $100 million using Computer. Mistral US business share: 25% - Mensch says a quarter of Mistral’s business is in the US. Mistral researcher share in US: 25% - He also says a quarter of the company’s researchers are in the US. IREN flagship Texas site: 750 megawatts - Roberts says the company’s flagship Texas data-center site is 750 MW. IREN total power capacity: 4.5 gigawatts - Roberts says IREN has secured 4.5 GW of power. IREN Microsoft contract: $9.7 billion - Roberts says IREN signed a large contract with Microsoft late last year. IREN capacity used by Microsoft deal: 5% - Roberts says the Microsoft contract represents only 5% of IREN’s capacity. West Texas renewable generation: 45–50 gigawatts - Roberts says the region has substantial wind and solar generation. West Texas transmission capacity: 12 gigawatts - He notes the transmission line capacity to Dallas/Houston is much smaller than generation. Perplexity token cost reduction: $32+ to $0.09 per million tokens - Srinivas cites a dramatic drop in token costs since early ChatGPT-era pricing. Perplexity enterprise tiers: $40/month and $400/month - Reiterated pricing for enterprise customers. IREN job postings: 129 - Roberts says the company currently has 129 job ads open.

Pivotal Quotes: "“Inference is the monetization of the investment in artificial intelligence.”" — Michael Intrator: CoreWeave discussion on why deployed AI workloads matter more than just model training. "“The AI itself is the computer now.”" — Aravind Srinivas: Perplexity’s vision for Computer as an orchestration layer across models, browser, and desktop tasks. "“The reality is, it’s gangbusters. We cannot meet demand.”" — Daniel Roberts: IREN’s view of AI infrastructure demand and the ongoing shortage of compute capacity.

Implications: AI winners will be those who control infrastructure, orchestration, and specialized deployment. Expect continued demand for power, GPUs, memory, and enterprise-safe automation, plus more opportunities for small teams to build real businesses with AI.