
Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN
All-In with Chamath, Jason, Sacks & Friedberg
Hosted by Jason Calacanis
- 1h 37m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
Jason Calacanis interviews four AI infrastructure and application CEOs at Nvidia's GTC conference: Michael Intrater of CoreWeave, Aravind Srinivas of Perplexity, Arthur Mensch of Mistral AI, and Daniel Roberts of IREN. The conversation spans the full stack of the AI buildout, from GPU financing and data center construction to multi-model orchestration and enterprise deployment.
Intrater explains CoreWeave's 'box' financing structure, which bundles client contracts, GPU purchases, and data center leases into self-contained cash flow vehicles that enabled the company to raise $35 billion in 18 months. He dismisses the GPU depreciation debate as nonsense, noting that average contracts run five years and older A100s have appreciated in price through the year. Srinivas positions Perplexity as a model-agnostic orchestrator, comparing the company to Switzerland, and describes the evolution from search to agentic browsing to a full computer running on Linux servers. Mensch outlines Mistral's strategy of deploying open-source models into enterprise environments with forward-deployment engineers, emphasizing data segregation and vertical specialization. Roberts describes IREN's 4.5-gigawatt power portfolio in West Texas and the shift from GPU scarcity to 'time to compute' as the binding constraint, with tradespeople earning $150,000 to $300,000 to build data centers in remote locations.
- CoreWeave uses a six-year depreciation schedule and sells older GPUs at higher prices than a year ago, contradicting the 16-month obsolescence narrative. 11:30
- CoreWeave's 'box' structure bundles a client contract, GPU purchase, and data center lease into a single cash flow vehicle, enabling $35 billion in debt raised over 18 months. 20:16
- Perplexity's model-agnostic orchestration means it benefits regardless of which frontier model wins, a structural advantage no single-model company can replicate. 47:55
- Memory has become the current bottleneck in AI infrastructure due to a confluence of surging AI demand and a missed 2023 fab investment cycle. 26:32
- IREN controls 4.5 gigawatts of power capacity in West Texas, nearly matching the Bay Area's annual electricity consumption, making power rather than GPUs their primary constraint. 1:22:05
Highlights
- 10:42
GPU Depreciation Debate Dismissedinsight
Intrater calls the GPU obsolescence debate nonsense driven by short-position traders, pointing to five-year average contracts and A100s appreciating in price through the year.
- 19:22
The Box Financing Structurestrategy
Intrater walks through how CoreWeave bundles a Microsoft contract, GPU purchase, and data center lease into a self-contained 'box' that governs cash flow, enabling lenders to underwrite each deal independently.
- 26:32
Memory as the New Bottleneckdata
Intrater explains that memory has become the throttle due to a missed 2023 fab investment cycle colliding with surging AI demand, creating a boom-bust dynamic similar to energy.
- 47:55
Perplexity as Switzerlandinsight
Srinivas argues that Perplexity's model-agnostic orchestration is a structural moat: whether GPT, Gemini, or Claude wins, Perplexity routes to the best model for each task, something no single-model company can or would do.
- 55:32
All In Podcast Stock Analysisstory
Srinivas describes asking Perplexity Computer to download every All In podcast, extract all public company mentions, chart them over time, and analyze their impact on stock prices, finding a correlation around Google.
- 1:22:05
IREN's 4.5 Gigawatt Portfoliodata
Roberts reveals IREN controls 4.5 gigawatts of power in West Texas, nearly matching the Bay Area's annual consumption, and explains that the constraint is now time to compute rather than GPU availability.
- 1:29:38
Jevon's Paradox in AI Computeinsight
Roberts argues that increasing compute supply will increase demand rather than reduce it, using image generation latency as an example: if images take 5 seconds instead of 2 minutes, users generate far more of them.
- 1:36:11
Roberts explains that IREN's West Texas site has only 6 milliseconds round-trip latency to Dallas, proving data centers do not need to be near population centers due to existing fiber infrastructure.
About this episode
(0:00) Intro live from Nvidia GTC (0:37) CoreWeave CEO, Michael Intrator (32:58) Perplexity CEO, Aravind Srinivas (1:07:11) Mistral CEO, Arthur Mensch (1:18:57) IREN CEO, Daniel Roberts Our episode is sponsored by the New York Stock Exchange - a modern marketplace and exchange for building the future. It all happens at the NYSE - https://nyse.com Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect
Transcript
17,737 words · 5 speakers
Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.
Topics
- AI Infrastructure
- GPU Financing
- Multi-Model Orchestration
- Data Centers
- Open Source AI
- Enterprise AI
- Energy and Compute
- Agentic AI
Page updated · transcript and summary are machine-generated