Adi Gelvan

In the recent Dwarkesh interview, NVIDIA CEO Jensen made a very important point that Dwarkesh didn’t pick up on. AI is more than chips. It is a five layer stack, each of which has to be running well for AI to make an impact. At the bottom layer, we have raw energy. The US has historically been slow to expand this, but without access to massive amounts of new energy, AI will be constrained. It’s the fundamental reason why Elon wants to launch AI datacenters into space, for energy access. The next layer is chip manufacturing, which TSMC and Samsung and others do. Here too, Elon is predicting a bottleneck and it’s why he’s kicking off his Terafab project. Next is chip design. This is where NVIDIA fits in. NVIDIA has over 25 years of chip design expertise, now helped by custom in house AI chip designers. Here Tesla is just learning the ropes with AI4, Dojo and soon AI5 chips. Next is computer science AI algorithm development to expand the frontier of intelligence. While we think we have pretty smart AI models now, we are actually at the very beginning of this layer. There is easily 20 years of continuous improvement ahead of us here. Elon’s xAI is trying to catch up with the likes of Anthropic, OpenAI, Google, etc. The final layer is useful applications for consumers and business alike. ChatGPT and MidJourney were the original modern AI applications, but we will see thousands of specialized AI applications that directly tackle hard subjects in all science, engineering, creative arts, etc. Grok and Tesla FSD are Elon’s current AI applications, but Optimus will soon join them.