Dylan Patel on the AI Chip Race — NVIDIA, Intel & the US Government vs. China

Podcast: a16z Podcast
Hosts: Erik Torenberg, Sarah Wang, Guido Appenzeller
Guest: Dylan Patel — Chief Analyst at SemiAnalysis
Duration: ~1 hour 38 minutes
Listen: Apple Podcasts | YouTube

Dylan Patel, one of the most respected analysts in semiconductors, covers the geopolitics and economics of the AI chip race.

Nvidia’s $5 Billion Bet on Intel

Nvidia invested $5 billion into Intel. Two historic rivals. The deal is already worth roughly $1 billion more than when it was made. AMD is the biggest loser — when two arch enemies team up, it is the worst possible news for the third player. The alliance signals Intel’s willingness to reset its internal GPU efforts and Nvidia’s need for more manufacturing capacity.

China Is Playing Chess, Not Checkers

Western analysts often underestimate China’s semiconductor progress. Huawei was the first company to bring 7nm AI chips to market in 2020 — before Apple. US sanctions forced Huawei to work with SMIC, China’s domestic manufacturer. Before the sanctions were fully enforced, Huawei acquired roughly 2.9 million chips from TSMC through shell companies, resulting in a billion-dollar fine for TSMC.

The HBM Bottleneck

High Bandwidth Memory is the hardest problem in the AI chip race. China is making progress on logic manufacturing (replacing TSMC functionality) but memory production (replacing Samsung, SK Hynix, Micron) is much harder. China is importing more etching equipment critical for HBM, but yield improvement will take years.

How GPUs Are Bought

“Buying GPUs is like buying cocaine. You call a couple people, you text a couple people, you ask — how much you got? What’s the price?” The market operates in cycles of scarcity and abundance. Small quantities are easier now. Large capacity remains brutally hard, especially with the transition to Blackwell architecture.

Jensen Huang’s Bet-the-Company Style

Huang orders production before securing customer commitments. He has said he does not look at spreadsheets. Nvidia consistently delivers first-generation silicon without requiring revisions — an almost unheard-of execution capability in the semiconductor industry. This speed, combined with a software ecosystem that keeps pace with hardware changes, is Nvidia’s real moat.

Oracle Is Winning by Being Flexible

Oracle emerged as a major AI infrastructure player by not being dogmatic about hardware and by making large, willing bets. Their commitment to OpenAI (reportedly $300 billion) signals a willingness that more established hyperscalers have been slow to match.

Specialized AI Workloads Are Coming

The industry is separating inference into prefill and decode operations — each with different computational requirements. Dedicated chips optimized for each will replace general-purpose AI accelerators, improving efficiency and user experience.

Crepi il lupo! 🐺