All In Interview - Michael Intrator, Aravind Srinivas, Arthur Mensch, Daniel Roberts - Recap and Notes - March 22 2026

March 23 2026

(from NotebookLM)

Interview 1: Michael Intrader (CEO of CoreWeave)

• Company Background: CoreWeave didn’t start as an AI cloud company. They started in 2017 as an algorithmic natural gas hedge fund, pivoted to crypto mining (Ethereum) using GPUs, and then moved up the complexity stack to CGI rendering and medical research.
• The Pivot to AI: In 2020/2021, they bought A100 GPUs and donated them to a volunteer open-source project called EleutherAI. When those researchers went back to their day jobs, they demanded CoreWeave's infrastructure, which launched their AI training business. Their first large commercial client was Inflection.
• Where They Sit: CoreWeave operates "above the Nvidia GPUs but below the models," focusing solely on the infrastructure and software integration needed for purpose-built AI clouds.
• Thinks of Inference as monetization of the training of the model
• The "Box" Financing Model (Important Concept): Intrader explained how they raised $35 billion in 18 months. They use a special purpose vehicle he calls "the box".
  • They secure a 5-year client contract (e.g., with Microsoft), put it in "the box", and add the purchased GPUs and data center contracts.
  • The client pays "the box" directly.
  • The cash flow pays the data center, power, and debt first; the rest flows to CoreWeave.
  • Debt is paid off in 2.5 years on a 5-year contract, making it super safe for lenders. This model dropped their cost of capital by 600 basis points.
• GPU Depreciation Myth: Intrader says the idea that GPUs are obsolete in 16 months is "nonsense" peddled by short-sellers. Contracts are for 5 years. In fact, older Ampere (A100) prices actually appreciated this year because smaller startups locked out of newer H100 chips are buying them up for different use cases.
• Hardware Bottlenecks: Memory is currently a massive throttle due to the boom/bust cycles of capital-intensive fab building.

Interview 2: Aravind Srinivas (CEO of Perplexity)

• Product Evolution: They evolved from a simple AI search (focusing on accuracy to build trust), to the Comet browser (giving AI browser access to automate tasks), to "Computer" (giving AI full access to execute workflows as an orchestrator).
• The "Switzerland" Strategy: Unlike Google or OpenAI, Perplexity uses multi-model orchestration. They don't care if GPT, Claude, or Llama wins. They have a feature called "Model Council" that runs prompts through multiple models and explicitly tells the user where the models agree, disagree, and have nuances.
• Business Model: They have positive gross margins on all revenue because they are highly efficient at token usage and routing through their orchestration/RAG system, meaning they don't blow up the context window unnecessarily. Their fastest-growing segment is Enterprise (Enterprise Max is $400/month per user).
• Models are specializing - not commoditizing - even Dario admitted this recently
• Local vs. Server Processing: They announced Perplexity Personal Computer, which will sync your local machine (like a Mac Mini) with your phone.
  • Highly private data (taxes, personal notes) will run locally.
  • Server-side data (Gmail, Google Calendar) will run via connectors on the server side.
• API access could be a revenue stream for the provider - would make the subscription more sticky
• The Future Vision: Srinivas believes AI will replace massive amounts of back-office manual work (like HR and recruiting) and eventually allow people to run highly profitable 1-person businesses autonomously.

Interview 3: Arthur Mensch (CEO of Mistral AI)

• Nvidia Partnership: Mistral announced they are training their next generation of frontier models in partnership with Nvidia to produce the best open-source models available.
• Enterprise Customization: They are heavily focused on helping European and US companies customize models for specific verticals (like physics, engineering, and finance) using a product called Forge.
• Data Security Strategy: Enterprises don't want to send their data to a model provider. Mistral solves this using a portable platform. They deploy their training tools directly onto the customer's IT infrastructure so the data never flows back to Mistral.
• Synthetic vs. Human Data: They use synthetic data (generated by larger models) to "warm up" and compress smaller models initially, but Mensch noted that human domain experts are absolutely required at the end to provide the final signal.
• Why Enterprises Can't Just Use Open-Source Hacker Tools: While tools like "Open Claw" are great for individual hackers, enterprises need deterministic gates, observability, and role-based access control (RBAC). Mistral builds "context engines" with metadata tagging so that, for example, an AI agent doesn't accidentally leak HR compensation data to the engineering department.

Interview 4: Daniel Roberts (co-CEO of IREN)

• The Pivot: IREN started 8 years ago in Sydney, building massive data centers (like a 750-megawatt site in West Texas) to mine Bitcoin, but have recently pivoted to swapping out Bitcoin chips for AI GPUs.
• The Real Bottleneck: While power is an issue for the rest of the industry, IREN already secured 4.5 gigawatts of power. Their actual constraint is "time to compute"—the physical manual labor of building data centers in the desert (foundations, water cooling, HVAC). They are paying massive salaries to local tradespeople to get this done.
• Energy Strategy: They run on 100% renewable energy (hydro in British Columbia, wind/solar in Texas). They locate their data centers directly at the source of the power generation because transmission lines are the actual bottleneck. The local utility handles the variable weather by guaranteeing 24/7 power.
• Future of Power (Nuclear): Roberts believes Small Modular Reactors (SMRs) are a decade away but will be crucial for distributed compute in the future.
• The Latency Myth: People used to think data centers had to be near major cities for latency reasons. Roberts pointed out that West Texas to Dallas is only a 6-millisecond roundtrip because there is already massive amounts of dark fiber buried underground.