Dive into the cutting-edge worlds of technology and science with Dwarkesh Patel's meticulously researched interviews that bring together brilliant minds to unpack revolutionary ideas in AI, biotech, and beyond. Each episode delivers profound, thought-provoking conversations that challenge conventional wisdom and illuminate the paths shaping humanity's future.

Dwarkesh Podcast

Contrary to expectations of rapid depreciation due to technological advancements, an H100 GPU is worth more today than three years ago because it can now run more advanced AI models that generate significantly higher value, such as GPT-5.4 compared to GPT-4. This dynamic highlights how compute value increases with model improvements in a capacity-constrained environment, benefiting early committers like OpenAI over more conservative players like Anthropic.

By 2030, the primary bottleneck for scaling AI compute will shift to ASML's EUV lithography tools, with production limits potentially capping cumulative AI chip capacity at around 200 gigawatts, challenging ambitious goals like Sam Altman's vision of deploying a gigawatt per week amid the ongoing AI boom.

Despite ambitious demands for massive AI compute scaling, such as Elon Musk's 100 gigawatts per year by 2028, the semiconductor supply chain lags far behind due to insufficient AGI-driven production ramps and architectural inefficiencies, making reliance on older 7nm processes inadequate for bridging the gap. This highlights a critical bottleneck in AI advancement, where performance differences between chip generations can reach 20x due to networking and data movement constraints, not just raw FLOPS.

Why is TSMC allocating the majority of its 3nm process capacity to Nvidia rather than diversifying among competitors like Google or Amazon to fracture the market and maintain leverage?

What prevents ASML from significantly increasing its production of EUV tools beyond 100 per year by 2030, despite the explosive demand for AI chips?

Given the compounding performance advantages of newer chip architectures, what fraction of the demanded AI compute could realistically be achieved by repurposing older 7nm fabs, and under what economic conditions would this become viable?

Get Beren Millidge on Dwarkesh Podcast

The Case for This Conversation

What People Are Saying

Key Catalysts

Updates

Shape What Gets Discussed

Supporters

Featured Videos

Be the first to add your voice to this conversation

More voices on Dwarkesh Podcast

Vaclav Smil

Karl Friston

Casey Handmer

Where else should Beren Millidge appear?

You Might Also Enjoy