Edited by | Yan Zheng

In the beginning of 2025, the market competition of AI chips has once again become a competitive topic on the other side of the ocean. Recently, the revenue growth rate of dominant Nvidia began to slow down, compared with the past 30 days of Broadcom and other companies' stock prices continued to rise, many people in the industry are guessing that the first to break the monopoly of Nvidia's second "Nvidia" will be who? How will the cloud rental market change?

On Dec. 25, the latest episode of the BG2 podcast saw hosts Bill Gurley and Brad Gerstner sharing the stage again, and featured SemiAnalysis founder and principal analyst Dylan Patel, Let's discuss the ongoing trends in AI chips and semiconductors and the latest trends in AI supergiants.

In the podcast, the three discuss Nvidia's many rivals in the chip race: AMD, Intel, Google, Amazon and even some new internal developments including OpenAI, which has threatened to build its own chips, as well as the positive and negative benefits of the storage market and future forecasts, such as they predict that the growth rate of the memory market will exceed that of GPU (leading silicon), and the free cash flow of Microsoft and Meta will be close to 0 in 2026. This is the result of an arms race spearheaded by Musk.

Before starting to discuss this topic, the three also explored the wall of AI pre-training, synthetic data generation, and the paradox of building larger clusters since they are all saying "pre-training is obsolete," and computing when reasoning.

Without further ado, this gives you a wonderful view.

• Microsoft earns 50 to 70 percent gross margins on OpenAI models. This is the share of the profits they get from AI or the share they give up.
• There are 15 different companies out there offering API inference for different models like Llama and Alibaba and DeepSeek and Mistral.
• If we just deploy lalama 7B-quality models, we're already overbuilt, and that's not even a joke.
• For the lab, you have to have the best model or you're nothing.
• Nvidia's highest cost of sales is not TSMC, which is what people don't realize, it's actually mainly HBM memory.
• Samsung is not a big hit, because the low-end is suffering, and the high-end can not break through.
• The only reason AMD's Gpus are considered better: there is more HBM memory on the chip.
• Google does engineering for some things that Nvidia might not be as focused on. So actually, the interconnect between chips may be competitive or even better than Nvidia in some ways, and slightly less so in others because they've been working with Broadcom.
• Broadcom is well suited to make competing products for NV switches, which many would argue is one of Nvidia's biggest competitive advantages over others in terms of hardware.