Mar 23, 2026 · 1h 37m · allin
Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN
Jason Calacanis · 33m spoken Michael Intrator · 23m spoken Aravind Srinivas · 16m spoken Arthur Mensch · 7m spoken Daniel Roberts · 7m spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Recorded live at NVIDIA GTC 2024, host Jason Calacanis interviews the CEOs of CoreWeave, Perplexity AI, Mistral AI, and IREN to examine the rapidly evolving ecosystem of AI cloud infrastructure, enterprise security, autonomous software agents, and data center energy demands.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 38.9% of the talking time here. How this is scored →
The hosts as informed peer 4.8 Guest teaching 2.7 Guest disagreement 1.2 The hosts pushing back 0.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 10:36 Calling GPU short-seller depreciation arguments nonsense
Michael Intrator forcefully rejects short-seller claims that GPUs become obsolete in 16 months, dismissing the narrative as baseless trader rhetoric contradicted by CoreWeave's 5-year customer contracts.
Hardest push from the hosts ▶ 47:55 Challenging Perplexity's independent survival against tech giantsJason directly challenges Aravind on how a 400-person startup can survive independently against Sam Altman, Elon Musk, and Google spending tens of billions on CapEx.
Biggest teaching moment ▶ 18:44 CoreWeave 'box' debt financing vehicle breakdownMichael Intrator provides an detailed explanation of CoreWeave's SPV 'box' structure, demonstrating how collateralized client contracts and cash waterfalls dropped their cost of capital by 600 basis points.
The host holds their own ▶ 1:30:05 Applying Jevons paradox to AI compute demandJason invokes Jevons paradox and induced demand theory to explain why software efficiency gains and cheaper tokens will increase, rather than decrease, total GPU infrastructure consumption.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| CoreWeave Origin Story: From Crypto Mining to AI Compute | 3 | 2 | 1 | 0 | Jason asks Michael Intrator about CoreWeave's pivot from crypto mining to AI infrastructure. He references basic industry terms like quants and researchers, while Intrator explains their progression through CGI rendering and EleutherAI open-source donation. | |
| Decommoditizing Compute at Scale vs. Traditional Clouds | 4 | 2 | 1 | 1 | Intrator describes compute decommoditizing at scale and CoreWeave's focus on living between NVIDIA GPUs and foundation models. Jason prompts him on hyperscaler competition (AWS) and inference expansion. | |
| GPU Depreciation Debate and Chip Lifespans | 5 | 1 | 3 | 1 | Jason introduces the GPU depreciation debate and quotes short-sellers like Michael Burry. Intrator forcefully calls the 16-month obsolescence claim 'nonsense' pushed by short sellers, while Jason reinforces this using an iPhone global trade-in analogy. | |
| Data Center Power Constraints & Long-Term Infrastructure Value | 4 | 2 | 1 | 1 | Jason notes that energy-to-compute ratios dictate GPU replacement timing. Intrator explains that data center power availability determines when old hardware is repurposed rather than physical chip decay. | |
| GPU Allocation Dynamics & CoreWeave's 'Box' Debt Financing | 5 | 5 | 1 | 1 | Jason asks detailed questions about debt interest rates and corporate paper versus venture capital. Intrator delivers an in-depth breakdown of CoreWeave's 'box' SPV financing structure and 600 bps reduction in cost of capital. | |
| AI Demand Relentlessness & Semiconductor Supply Chains | 5 | 3 | 1 | 1 | Jason draws analogies to New York Knicks waitlists and telecommunication fiber boom-bust cycles. Intrator details supply chain throttles like memory fab capital intensity and power shells. | |
| Declining Token Costs and the Future of AI Infrastructure | 5 | 2 | 0 | 0 | Jason shares early YouTube founder stories about bandwidth and storage cost drops, as well as Andrej Karpathy's recursive prompting. Intrator quotes Sarah Friar on token prices dropping from $32 to $0.09. | |
| Perplexity Evolution: Search, Browsing, and Computer | 5 | 2 | 1 | 1 | Jason outlines his user journey through Perplexity, Comet browser, and Perplexity Computer. Aravind Srinivas outlines Perplexity's positioning as an orchestra conductor coordinating specialized models. | |
| Local Hardware, Privacy, and Smart Workstation Trends | 5 | 2 | 1 | 1 | Jason mentions running local models on Mac Studio and the new Dell/NVIDIA 750GB RAM workstation. Aravind explains Perplexity Personal Computer running on a local Mac Mini as a private orchestration server. | |
| AI as the Operating System & Enterprise Integration | 4 | 3 | 1 | 0 | Aravind describes AI becoming the operating system where users input high-level objectives rather than programmatic instructions. Jason highlights Linux's potential as the ideal headless backend. | |
| Corporate Growth, Subscription Business Model & Unit Economics | 5 | 3 | 2 | 2 | Jason presses Aravind on how a 400-person company can stay independent against hyperscalers with $100B CapEx budgets. Aravind pushes back that Perplexity's neutrality as a multi-model orchestrator is an advantage giants cannot replicate. | |
| Harnesses, Auto-Routing, and the Model Council Feature | 5 | 3 | 1 | 1 | Jason shares generating a custom CRM application through Claude and inquires about auto-routing queries. Aravind explains Perplexity's 'Model Council' feature that synthesizes agreement across multiple LLMs. | |
| Context Window Breakthroughs and Autonomous Business Workflows | 5 | 3 | 1 | 1 | Jason describes using mega-prompts to research guest interview histories back 10+ years. Aravind shares how Perplexity Computer autonomously generated board memos, press briefs, and historical podcast analysis. | |
| AI Job Displacement and the Rise of Solo Entrepreneurs | 6 | 2 | 1 | 2 | Jason proposes a paid user API model for sites like Reddit and LinkedIn to allow authenticated agent access. Aravind discusses AI-enabled solo entrepreneurship, framing job displacement as a shift toward individual agency. | |
| Mistral AI's Open-Source Strategy and Enterprise Customization | 4 | 3 | 1 | 1 | Arthur Mensch announces Mistral's NVIDIA partnership and open-source strategy. He explains how enterprise customers use Forge post-training to adapt general open models into specialized vertical domain models. | |
| Enterprise Data Privacy, Synthetic Data, and AI Governance | 6 | 4 | 2 | 1 | Jason shares giving OpenClaw root access to his firm's G Suite and Slack, realizing privacy hazards like employee PIPs and comp data leaks. Mensch highlights why enterprises require portable platforms, RBAC, and context engines. | |
| IREN's Transition from Bitcoin Mining to Renewable AI Compute | 5 | 3 | 1 | 1 | Jason asks Daniel Roberts about trade union salaries ($150k-$300k) and setting up data center towns in remote West Texas. Roberts explains IREN's 100% renewable energy strategy using stranded hydro, wind, and solar. | |
| AI Infrastructure Scaling, Power Grids, and Future Energy | 6 | 3 | 1 | 1 | Jason cites Jevons paradox and induced demand to argue lower token costs will explode compute demand rather than shrink it. Roberts dispels data center latency myths with 6ms Texas roundtrip statistics. |
Statements from this episode (41)
CoreWeave bought and donated NVIDIA A100s to EleutherAI to learn AI compute
“We actually went out and bought a bunch of A-one hundreds and donated them to a group that was working on Luther AI.”
Intrator: Compute decommoditizes at scale as clusters grow for frontier AI models
“Computing decommoditizes its scale, right? Like, when, you know, anybody can run a GPU, but can you run a cluster that's large enough to train a model that can change the world?”
Intrator: CoreWeave's software and infrastructure sit directly above NVIDIA GPUs and below AI models
“When you're thinking about what we do, we kind of live above the NVIDIA GPUs, but below the models.”
Intrator: CoreWeave's first major commercial client was Inflection AI
“Our first like large commercial was inflection.”
Michael Intrator: AI inference represents the monetization of AI investment
“I always think of inference as the monetization of the investment in artificial intelligence.”
Intrator: CoreWeave's average client AI compute contract is five years
“Our clients come into us and they buy compute for five years, for six years. Our average contract is five years.”
CoreWeave uses six-year GPU depreciation schedule and expects longer lifespans
“We use a six year depreciation. We believe that the GPUs will last in excess of six years, but we felt like that was a fair and reasonable approach to a technology cycle that's moving at this velocity.”
Intrator: Prices for older Nvidia A100 GPUs appreciated over the last year
“The A-Hundreds, the Amperes, this year, the price has appreciated through the year.”
Intrator: Data center power opportunity defines GPU hardware obsolescence
“My expectation is, is obsolescence will be defined by the moment in time where the power in the data center For me, will be able to be repurposed for a higher margin than the existing infrastructure provides.”
Intrator: CoreWeave created the first GPU-collateralized loan structures
“CoreWeave has really been the innovator around a lot of the financing engines that have come to bear on this. We did the first GPU based loans.”
Intrator: CoreWeave raised $35B in 18 months for infrastructure
“CoreWeave, which is a company that many people haven't ever heard of, was able to go out and raise thirty five billion dollars in 18 months to build infrastructure at scale.”
Intrator: CoreWeave lowered its cost of capital by 600 bps in two years
“When you think about our cost of capital over the last two years, we have dropped our cost of capital by 600 basis points.”
Intrator: Global AI compute demand has exceeded world capacity for four years
“For years now, for four years, the depth of the demand for the service we provide has been relentless and overwhelms the global capacity of the world to deliver enough compute to enable all of the demand for artificial intelligence to be stated.”
Intrator: A 2023 fab investment cycle would have met current AI demand
“There was probably an investment cycle that needed to happen back in 2023. That would have brought on the necessary fab capacity to be able to serve.”
Intrator: OpenAI token costs dropped from $32 to nine cents
“She was talking about the cost of a million tokens when ChatGP three came out, and it was 32 dollars in change, and now a million tokens cost nine cents.”
Srinivas: Perplexity is developing local AI integration for private data orchestration
“So we know something called personal computer perplexity, personal computer. That's essentially going to take all the trust and reliability and the server side execution of perplexity computer, but synchronize it with your local computer so that you can use it…”
Calacanis: Local open-source models deliver about 80% of frontier performance for free
“I have started running Kimmy 2.5 on a Mac studio. It's not as good as Claude or Gemini or Grok, but you can probably do about 80% there for free.”
Srinivas: Local AI execution will start off as sub-agents for private data
“My prediction is that initially start off as a sub agent. So whatever you need to go like your tax returns, your personal photos, your emails, your calendar, all that stuff, those local apps, your personal notes, very personal notes. You can make sure that the…”
Srinivas: AI is replacing traditional operating systems with objective-driven computing
“AI is the operating system. Like earlier in the traditional operating system, you execute programmatically. Now you start with objectives, not specific instructions.”
Perplexity CEO: Perplexity Computer is available as an enterprise Slack bot
“Computer exists as a Slack bot right now. That you can add to your Slack workspace on the enterprise plan.”
Srinivas: Perplexity staff message AI agent on Slack more than colleagues
“And our entire company works like that. People are talking more to computer on Slack than other people.”
Srinivas: Enterprise is Perplexity's fastest-growing revenue segment
“It's the fastest growing business for us. It's growing faster than the consumer and revenue”
Srinivas: Perplexity has positive gross margins on all revenue
“Our one thing that perplexity has is every revenue we make, unlike certain other wrapper companies, every revenue perplexity makes has positive gross margins.”
Srinivas: Big Tech AI labs cannot aggregate rival models without admitting failure
“It makes no sense for them. It would be an admission that all the data centers in CapEx they've built out still couldn't produce them the best model.”
Srinivas: Value accrues to orchestration harnesses as AI models specialize
“So the, when models are kind of specializing, the, there's a bigger value in the one who knows how to build a great harness. That can take the best in each model.”
Srinivas: Perplexity Model Council synthesizes agreement and nuances across models
“So the Model Council is a feature we built where it will not just give you the answers of each model, but it will tell you exactly where they agree, where they disagree, and where the nuances are.”
Srinivas: Non-engineers at Perplexity AI ship code using Slack AI bots
“Even non-engineers are shipping code here by just pinging a slack bot and asking it to fix bugs.”
Srinivas: Anthropic Opus was an inflection point in AI model orchestration
“That was an inflection point when models were, started being amazingly good at orchestration and reasoning and tool calls.”
Srinivas: Automating web workflows requires AI that natively controls browsers
“Until the whole world is organized around CLIs and tools. There's still a lot of tasks we have to do manually on the web, on the browser, open tabs, fill up forms, click on things, upload stuff. All that stuff, if you want to automate, you need a browser. You …”
Mensch: Mistral AI will train next-gen frontier models with NVIDIA
“Well, we are announcing that we are going to be training the next generation of Frontier models with NVIDIA.”
Mensch: Mistral AI has 25% of business and researchers in US
“First, we have 25% of our business in the US and 25% of our researchers are actually here.”
Mensch: Open AI models allow parameter addition impossible with closed models
“If you have open models, you can actually add new parameters. You can make a lot of deeper things that you cannot do with closed models.”
Mensch: Mistral deploys model training tools directly on client infrastructure
“Overall the data segregation is super important and the way we have solved that is through a portable platform. So our technology is a set of services, a set of training tools, A set of data processing tools that I can take and that I can put on the infrastruc…”
Mensch: AI synthetic data is efficient but cannot replace human training signal
“It's mostly an efficient way of training models to have bigger models that are used as teachers for smaller models, but it's not enough. And so you also need human signal.”
Mensch: Open-source AI agent frameworks lack enterprise governance and control
“And to automate the full process as an enterprise, well, you can use OpenClo, but it's going to be it's actually not really enough, because you have data problems, you have governance problems, you can't observe the process that is running, and you can't contr…”
Roberts: IREN's $9.7B Microsoft deal uses 5% of its capacity
“We signed a 9.7 billion dollar contract with them late last year, but as I was explaining to you before the show, that's five percent of our capacity, so.”
Roberts: IREN controls 4.5 gigawatts, rivaling Bay Area's annual power usage
“So we've got four and a half gigawatts. For context, that's almost as much power annually as the Bay Area uses in its entirety.”
Roberts: IREN has operated on 100% renewable energy since inception
“We have used 100% renewable energy since inception.”
Roberts: There are currently no idle GPUs sitting in data centers globally
“There are no idle GPUs in the world sitting in a data center.”