Jason Calacanis interviews four AI CEOs at NVIDIA's GTC conference: Michael Intrator (CoreWeave CEO), Arvind Srinivas (Perplexity CEO), Arthur Mensch (Mistral AI CEO), and Daniel Roberts (IREN co-CEO). These leaders represent different aspects of the AI infrastructure stack - from GPU cloud computing and AI search to specialized model training and data center power infrastructure.

The conversations explore the evolution from crypto mining to AI compute, the competitive dynamics between frontier models, the challenges of scaling AI infrastructure, and the massive capital requirements driving the industry. Key themes include GPU depreciation debates, the shift toward specialized vertical models, agentic AI capabilities, and the critical role of power infrastructure in enabling AI scaling.

Each CEO discusses their company's unique positioning in the AI ecosystem, from CoreWeave's financing innovations and long-term contracts to Perplexity's multi-model orchestration approach, Mistral's focus on open-source specialization, and IREN's renewable energy-powered data centers supporting the industry's massive compute demands.

CoreWeave's Evolution from Crypto Mining to AI Infrastructure

CoreWeave began in 2017 as an algorithmic hedge fund focused on natural gas, pivoting to crypto mining with GPUs because "there's some brilliant engineer that built the ASIC, and they're probably going to be better at running it than we are" - Michael

The company weathered crypto winters through disciplined risk management from their hedge fund background, then progressively moved up the complexity stack from crypto to CGI rendering, batch computing, medical research, and finally neural networks by 2020-2021

CoreWeave donated A100s to Eleuther AI as "tuition we paid to learn how to run this business," with volunteers later requesting the same infrastructure at their day jobs, launching the commercial business

The company recognized early that "computing decommoditizes at scale" - while anyone can run a GPU, building clusters large enough to train world-changing models requires different capabilities entirely

GPU Depreciation Debate and Hardware Lifecycle Reality

Michael dismisses GPU depreciation concerns as "nonsense" driven by traders with short positions, noting CoreWeave's average contract length is five years with clients willing to pay for that duration

A100 Ampere GPU prices have actually appreciated through the year as new companies emerge with different use cases and model sizes, similar to how older iPhones retain value in developing markets

CoreWeave uses six-year depreciation schedules, believing GPUs will last beyond that timeframe, with obsolescence ultimately defined by when power costs exceed infrastructure margins rather than technical capability

"If people are willing to pay me for it, it still has value" - Michael's simple approach to GPU valuation, emphasizing market demand over theoretical depreciation timelines

Innovative Financing Through Structured 'Box' Contracts

CoreWeave pioneered GPU-based loans through structured financing vehicles called 'boxes' containing client contracts, GPU purchases, and data center agreements with waterfall cash flows

The box structure ensures clients pay directly into the vehicle, which first covers data center costs, power bills, interest and principal before returning profits to CoreWeave

Within two and a half years of five-year deals, all principal and interest are paid off, giving lenders confidence in the "one rule of lending: give me my money back"

This financing innovation enabled CoreWeave to raise $35 billion in 18 months while dropping their cost of capital by 600 basis points toward hyperscaler borrowing rates

Perplexity's Multi-Model Orchestra Strategy

Perplexity positions itself as "Switzerland" in the AI model wars, with multi-model orchestration as their key advantage since "if GPT wins, Gemini wins, Claude wins, Llama wins, it doesn't matter to us" - Arvind

The platform serves tens of millions of users monthly with thousands of enterprise customers, growing faster on the corporate side than consumer, with Enterprise Max at $400/month per person

Perplexity Computer represents an "orchestra of everything AI can do today" with models as instruments and sub-agents as musicians, orchestrating capabilities across hundreds of specialized models

The company maintains positive gross margins on all revenue through subscription fees and efficient token routing, unlike "wrapper companies" that simply resell tokens at a loss

Agentic AI and the Future of Computer Interaction

Perplexity announced Personal Computer, synchronizing server-side AI orchestration with local Mac Mini hardware to provide hybrid local/cloud agent execution with privacy controls

The vision positions "AI as the operating system" where users operate at the objective level rather than giving specific instructions, with models handling file systems, connectors, and tools automatically

Arvind predicts AI will enable "one person, $1 billion companies" by automating business operations from ad campaigns to customer support, allowing entrepreneurs to "be sipping wine in Napa" while their business runs autonomously

The Model Council feature addresses the common practice of querying multiple AIs by automatically identifying where models agree, disagree, and showing nuanced differences rather than requiring manual comparison

Mistral's Specialized Vertical AI Approach

Mistral AI focuses on open-source models that can be deeply customized for enterprise verticals, working with companies like ASML and HSBC through their Forge platform for specialized training

The company deploys portable training platforms directly on customer infrastructure, ensuring "the flow of data doesn't go, there's no flow coming back to Mistral because everything stays there" - Arthur

Mistral sends PhD-level forward deployment engineers to work with subject matter experts, transferring knowledge about domain-specific tasks like image scanning and defect detection

While general-purpose models handle orchestration, enterprises benefit from specialized models that can leverage "decades of IP that you've been accruing in financial services, in heavy manufacturing"

Power Infrastructure as the New Constraint

IREN operates entirely on renewable energy, using hydro in British Columbia and wind/solar in West Texas where 45-50 gigawatts of renewable capacity exceeds 12 gigawatts of transmission capacity

The company's strategy involves going "to the source of low-cost, excess renewable energy, monetize it into this digital commodity, export it at the speed of light as tokens"

IREN's flagship Texas site operates at 750 megawatts with four gigawatts total capacity - "almost as much power annually as the Bay Area uses in its entirety" - Daniel

The constraint has shifted from GPU availability to "time to compute" - the physical construction and deployment of data centers, requiring thousands of tradespeople in remote locations

Economic Impact and Workforce Transformation

Data center construction is creating high-paying trade jobs with salaries in the $150-300k range, while IREN hires locally and provides over $1 million in community grants for playgrounds and fire departments

The industry locates where "heavy electrical infrastructure" exists, often where "old manufacturing and industry has closed down," leveraging existing infrastructure while retraining local workforces

Token costs have dropped dramatically from "$32 and change" for a million tokens when ChatGPT-3 launched to "nine cents" today, demonstrating the power of capital markets driving engineering competition - Sarah Fryer

The technology is "bringing down this incredible barrier that kept human creativity contained" by enabling anyone with a good idea to overcome previously insurmountable technical obstacles - Michael