The 3 Bottlenecks That Will Decide Who Wins the AI Race决定 AI 竞赛胜负的 3 个瓶颈

The Setup

The big four hyperscalers (Amazon, Meta, Google, Microsoft) are collectively deploying ~$600 billion in CapEx this year. Anthropic and OpenAI are raising tens of billions each. Yet most investors still frame AI infrastructure as “buy NVIDIA.”

Dylan Patel, CEO of SemiAnalysis and one of the most respected voices on semiconductor supply chains, just sat down with Dwarkesh Patel to explain why that framing is already outdated — and where the real bottlenecks are forming.

Key Takeaways

1. Logic chips: NVIDIA is still winning, but the allocation game is over

NVIDIA locked in TSMC capacity years ago. Google is now getting squeezed on leading-edge nodes. An H100 bought today is actually worth more than when it launched three years ago, because inference demand has exploded and the chips are fully utilized. The bet isn’t “will NVIDIA sell more chips” — it’s “who secured allocation.”

2. Memory is the next crisis

The memory crunch is enormous and incoming. HBM (High Bandwidth Memory) is already constrained. As model inference scales, memory bandwidth becomes the true bottleneck — not raw compute. SK Hynix, Micron, and Samsung are the chokepoints here, not NVIDIA.

3. ASML becomes #1 constraint by 2030

ASML’s EUV machines are the only way to manufacture leading-edge chips. There’s exactly one company making these machines, and they can only produce a limited number per year. By 2030, ASML’s output — not foundry capacity, not GPU design — will be the hard ceiling on AI compute scaling globally.

The hedge fund paradox

Why aren’t more institutional funds making the “AGI trade”? According to Patel: career risk. Fund managers can’t hold NVDA through a 50% drawdown even if they believe in the long-term thesis. Retail investors and long-only funds with high conviction have a genuine structural edge here.

Why It Matters

For Rex’s portfolio: this isn’t just interesting theory. The compute stack maps directly to where capital is flowing:

• NVIDIA — still the dominant GPU, but priced for perfection
• HBM memory producers — underappreciated bottleneck, especially SK Hynix
• ASML — the longest-duration “AI tax” with monopoly characteristics
• Hyperscalers — burning cash to secure compute before competitors can; the CapEx isn’t speculative, it’s defensive

Power, by contrast, is not a bottleneck in the US — Patel thinks scaling power is more manageable than media suggests.

What to Watch

• ASML quarterly order books — a leading indicator for AI compute capacity 2-3 years out
• HBM supply announcements from SK Hynix and Micron
• Whether OpenAI/Anthropic revenue growth continues at current trajectory (~$6B/month additions)
• China’s ability to scale semis domestically — Patel gives a timeline, worth tracking