FULL INTERVIEW: Dylan Patel Says We’re Still Underestimating AI - TBPN Recap

Podcast: TBPN

Published: 2026-02-03

Duration: 44 minutes

Guests: Dylan Patel

Summary

Dylan Patel argues that the potential of AI and space-based data centers is still underappreciated. The limitations of current AI hardware, geopolitical influences on semiconductor supply, and the future role of robotics are key components of this underestimation.

What Happened

Dylan Patel discusses the future of space-based data centers, predicting that the cost of Starship launches will decrease substantially by the end of the decade, making space-based operations more feasible. However, challenges such as chip reliability and heat dissipation remain significant barriers. Patel sees Tesla's self-driving chip redundancy as a potential model for improving satellite chip reliability.

Patel highlights the high initial failure rate of NVIDIA GPUs, with 10-15% failing when first turned on. This failure rate drops over time, but it underscores the challenges in current AI hardware. NVIDIA's CPX chip is noted for its video and image generation capabilities, while Cerebrus is valued for speeding up inference streams.

The conversation touches on the substantial OpenAI deal with Cerebrus, involving 750 megawatts, indicating a strong demand for faster prompt processing. Patel is skeptical of the bet that 1% of global data center capacity will be in space by 2028, citing the need for a gigawatt in space, which he deems unlikely.

Geopolitical dynamics play a crucial role in shaping AI infrastructure, with TSMC and memory manufacturers facing constraints in expanding fab capacity. Patel discusses the strategic positioning of Google's data centers, optimized for cross-data center operations, and the increasing feasibility of multi-data center training due to reduced data transfer needs.

The episode delves into geopolitical concerns, noting the risk of pushing China too far and potentially provoking aggressive actions. Patel emphasizes the importance of strategic decisions like Ben Thompson's suggestion to sell chips to China but ban lithography tools, aiming to create dependency.

Robotics emerges as a major focus, with Patel optimistic about their potential in everyday tasks such as laundry and dishwashing, though not yet for more complex roles like those in TSMC's Arizona fabs. The episode also highlights Meta's significant revenue from AI, despite criticisms of their product offerings, and their strategic investments in AI and wearables.

Key Insights

• Space-based data centers are becoming more feasible with the anticipated reduction in Starship launch costs by the end of the decade, though challenges like chip reliability and heat dissipation remain significant.
• NVIDIA GPUs have a high initial failure rate of 10-15% upon first activation, highlighting ongoing challenges in AI hardware reliability.
• OpenAI's substantial deal with Cerebrus involves 750 megawatts, reflecting a strong demand for faster prompt processing capabilities.
• Geopolitical dynamics impact AI infrastructure, with constraints on TSMC and memory manufacturers expanding fab capacity, and strategic positioning of Google's data centers optimizing for cross-data center operations.