I just read a very interesting interview with Jensen Huang from NVIDIA, and there are some points that are really worth discussing. The guy was very straightforward on several issues that everyone has been debating.

First, about NVIDIA's competitive advantage – it’s not just about having better chips. The real reason is that they built a huge ecosystem around CUDA. Their supply chain is so large that suppliers are willing to invest heavily because they know NVIDIA can absorb and sell everything. It’s like a virtuous circle: the more people use NVIDIA, the more incentive there is to develop on top of them, which attracts more customers. Simple as that.

Now, regarding the TPU vs GPU issue – Jensen made it clear that GPUs are much more flexible. TPU is optimized for matrix multiplication, but AI involves much more than that. When you want to invent new attention mechanisms or hybrid architectures, you need programmability. And that’s where NVIDIA has the edge. The guy said they achieve 10x to 100x improvements in efficiency not just through Moore’s Law, but through algorithmic innovation. From Hopper to Blackwell, energy efficiency increased by 30 to 50 times. That’s insane.

But the point that caught the most attention was the discussion about China. Jensen was very blunt: abandoning the Chinese market is a loser’s mentality. He argues that China already has enough computing capacity – they have abundant energy, chip factories, and the world’s best AI researchers. If you don’t sell to them, they will develop their own technology. More than that: forcing them out accelerates their chip industry instead of slowing it down.

Jensen also touched on something few mention: AI is a five-layer cake – application, model, infrastructure, chip, energy. The US needs to win in ALL layers, not just one. If you give up the chip market, you lose an entire layer of technological competition.

Regarding TPU vs NVIDIA, he was clear: Anthropic is an exception, not a trend. Most TPU growth comes only from them. Other companies continue choosing NVIDIA because the ecosystem is better, more mature, more reliable.

A very interesting insight: the guy explained that NVIDIA manages to double production year after year because they anticipate bottlenecks years ahead. They work with suppliers to expand capacity for CoWoS, HBM, photonics – everything in parallel. Meanwhile, each chip generation achieves 10 to 20 times more efficiency just through architectural innovation.

About why NVIDIA doesn’t become a cloud provider: their philosophy is to do the maximum of what’s necessary and the minimum of what’s unnecessary. They believe that if they didn’t build the computing platform, no one else would. But cloud services? Plenty of companies can do that. That’s why they invest in CoreWeave, Crusoe, and others – they want the ecosystem to flourish, not monopolize everything.

I also found it interesting when he said they’ve never used demand-based pricing. They keep prices fixed regardless of how much people want to buy. It’s like saying: you can count on us. This year Vera Rubin, next year Vera Rubin Ultra, then Feynman. Every year, like clockwork.

The final point that stood out: if you look at history, NVIDIA could have done everything wrong in the beginning – 60 companies making 3D graphics, and they were the least likely to survive. But because they stayed focused, kept innovating, and maintained an open ecosystem, they managed to dominate. Now with AI, it’s the same – it’s not about winning today, it’s about building to win in the next 20 years.

It’s a dense conversation, but basically: NVIDIA’s moat isn’t just because they have better chips. The moat is the entire ecosystem – developers, frameworks, reliability, constant innovation. And meanwhile, the geopolitical discussion about chips and China remains complex – it’s not as simple as it seems.