All-In: Four CEOs CoreWeave/Perplexity/Mistral/IREN
Category: Expert Interviews · Duration: 98 min · ▶ Watch
Speakers: Jason Calacanis · Michael Intrator, Aravind Srinivas, Arthur Mensch, Daniel Roberts
Segments (5)
- 00:00 · Introduction
- Jason Calacanis introduces the interviews at Nvidia GTC.
- 00:38 · Michael Intrator, CoreWeave
- Discussion on CoreWeave’s pivot to AI, GPU lifespan, and financing models.
- 03:00 · Aravind Srinivas, Perplexity
- Perplexity’s evolution, local AI computing, and enterprise pricing.
- 07:15 · Arthur Mensch, Mistral AI
- Open-source AI models, enterprise customization, and synthetic data.
- 09:02 · Daniel Roberts, Iris Energy
- Transitioning from Bitcoin mining to AI data centers and solving power constraints.
Specific Prices (3)
| Timestamp | Item | Value | Context |
|---|---|---|---|
| 05:33 | Perplexity Enterprise Pro | $40/month | Subscription tier for enterprise users. |
| 05:38 | Perplexity Enterprise Max | $400/month | Higher tier subscription for enterprise users. |
| 01:03 | Used Smartphone | $50 | Analogy used to explain that older hardware still has value in emerging markets. |
Memory Facts (1)
- [09:28] Dell/Nvidia workstation RAM capacity
- 750 GB
Bottleneck Claims (3)
- [02:28] Memory is a cyclical bottleneck.
- Evidence: Fabs overbuild and underbuild based on demand cycles.
- [02:19] Power is the main constraint for data centers.
- Evidence: It takes years to get grid connections and build power infrastructure.
- [10:20] Finding available power is harder than finding compute.
- Evidence: Iris Energy leverages its existing power infrastructure built for Bitcoin mining.
Predictions (3)
- [01:49, Long-term] GPUs will have a useful life in excess of 6 years.
- [04:14, Near-term] Desktop Linux computers will make a comeback as local AI servers.
- [09:39, Long-term] AI compute demand will continue to grow exponentially.
Key Technologies (4)
- GPUs: Provides the parallel compute power necessary for AI training and inference.
- Local AI Orchestration: Routes tasks between local hardware and cloud models for privacy and efficiency.
- Synthetic Data: Artificially generated data used to pre-train models before human fine-tuning.
- InfiniBand / Ethernet: High-speed networking fabrics connecting GPUs within a data center.
Companies Mentioned (9)
CoreWeave · Perplexity · Mistral AI · Iris Energy (IREN) · Nvidia · Microsoft · OpenAI · Anthropic · Apple
Notable Quotes (3)
The GPU depreciation debate is nonsense. — Michael Intrator @ 01:43
AI is the operating system. — Aravind Srinivas @ 03:54
The data center is the new computer. — Daniel Roberts @ 10:43
Key Topics
AI Infrastructure Financing · GPU Lifespan and Depreciation · Local vs. Cloud AI Computing · Open Source AI Models · Data Center Power Constraints · Renewable Energy for AI
Takeaways
- GPU depreciation fears are overstated; older hardware retains value for inference and emerging markets.
- Innovative financing models are enabling massive infrastructure build-outs.
- The future of AI involves a hybrid approach, orchestrating tasks between local devices and cloud models for privacy and cost efficiency.
- Open-source models are crucial for enterprise adoption, allowing companies to retain control over their data.
- Power availability, not just compute, is the primary bottleneck for scaling AI data centers.
- Data centers for AI training can be located near remote renewable energy sources, as latency is less critical than for traditional web serving.