Four AI CEOs—from CoreWeave, Perplexity, Mistral, and IREN—discuss the infrastructure, products, and business models shaping the AI era at Nvidia's GTC conference. The episode covers GPU economics, the shift from training to inference monetization, enterprise automation tools, and how specialized open-source models are challenging closed-source AI incumbents.
Key takeaways
•GPU depreciation is a myth: CoreWeave's CEO explains that A100s and H100s maintain value for 5-6 years through evolving use cases (training → inference → rendering), with enterprise contracts supporting this lifespan and secondary markets creating new demand.
•Inference is where AI monetization happens: As models become commoditized through competition, the real value comes from deploying inference at scale—CoreWeave's infrastructure enables this through purpose-built cloud solutions above Nvidia chips but below model layers.
•Perplexity's multi-model orchestration strategy neutralizes big tech advantages: By routing queries across Claude, ChatGPT, Gemini, and open-source alternatives like Qwen, Perplexity avoids vendor lock-in and offers specialized model strengths without building its own foundation models.
•AI agents need sandbox controls and data governance for enterprise adoption: Open-source tools like OpenClaw enable automation, but enterprises require deterministic gates, observability, and context engines to prevent sensitive data (like compensation) from leaking across departments.
•Specialized verticals and open-source models unlock enterprise customization: Mistral's approach of deploying models on customer infrastructure—with PhDs doing knowledge transfer—lets enterprises fine-tune models on proprietary data while maintaining security, competing against closed-source incumbents.
•Personal computing with local orchestration is the next frontier: Perplexity's Personal Computer using Mac Mini as a local server allows AI agents to access private data (emails, calendars, files) without sending it to cloud servers, balancing autonomy with privacy.
•Token costs have dropped 99.7% in two years, driven by competition and scale: OpenAI's price per million tokens fell from $32 to $0.09, demonstrating how capital markets and competition recursively compress AI costs and unlock new business models.
"We actually went out and bought a bunch of A100s and donated them to a group that was working on Luther AI. They were working on an open-source project."
"We were the first ones to bring the H100s at scale, we were the first ones to bring the H200s at scale, first ones with the GB 200s, and now you've got the GB300s."
"I take my contract with Microsoft and I put it in the box. I go to Jensen and I buy the GPUs, I put it in the box. I take my data center contract, I put it in the box."
"Then you came out with the Comet browser and I was like, 'Holy cow, I can give this a series of instructions. Go to my LinkedIn, find everybody from this company, put them into a Google sheet and b..."
"Just the last couple of weeks I had been Claude-pilled in using Claude but you came out with computer and I started using computer and boy it's good."
"We're going to do this with the Mac Mini where you synchronize your computer with the Mac Mini so that becomes your local server all the agent orchestration that has to do with your local private d..."
"Last week, I took two people in my back office and I said, 'Stop working on OpenClaw. Your job is to do the back office automation at our venture firm only using Perplexity.' And they were like, 'O..."
"And then I realized, wow, I can with my enterprise edition of Gmail essentially, I can just summarize for my entire 21 person investment company every conversation going on in Gmail and then correl..."
