TLDR
At the CASPA summit, Bernstein's Stacy Rasgon and SemiAnalysis's Dylan Patel discuss how financing constraints, supply chain bottlenecks, and pricing power are reshaping AI infrastructure. They reveal that semiconductor industry revenue doubled largely through price hikes, not volume growth, and that custom silicon from hyperscalers could deliver 5x compute per dollar by cutting out vendor margins. The conversation centers on a critical iron law: as long as performance gains outpace price increases, high pricing is sustainable; if innovation slows, the commercial loop collapses.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
The semiconductor industry was a net value destroyer for decades until pandemic-era pricing power allowed it to cover cost of capital for the first time.
Memory companies are now raising prices aggressively against former pressure from buyers like Apple, reversing a decades-long dynamic.
Nvidia's valuation multiples are at historic lows despite record earnings, and have compressed further in the last nine months.
Custom silicon economics are so compelling that hyperscalers can get 5x compute per dollar by designing their own chips and sourcing wafers directly from TSMC.
Power is not the top bottleneck — substrates are — and distributed generation (turbines, fuel cells) can bridge much of the gap.
Key takeaways
Financing has become the fourth major constraint alongside power, supply, and equipment, slowing capacity expansion; hyperscalers like Meta and Google are now doing equity raises, while emerging cloud providers without anchor contracts from top AI labs cannot secure funding at viable rates.
Substrate materials (PCB, etc.) are the #1 supply chain bottleneck for the next two years, affecting all components including CPUs and AI accelerators, due to conservative Japanese supplier investment; memory and optical components follow, while power is less binding than consensus expects thanks to distributed generation options.
Cleanroom space is fully utilized until 2027, and wafer fab equipment spending has surged from $120B to $155B expected this year, with $200B+ projected for next year — but Rasgon doubts this is feasible due to insufficient fab space to install the tools.
HBM manufacturing consumes 3-4x more wafer area per gigabyte than DDR5, driving memory to 70-80% of silicon area in AI racks; memory makers now enjoy 70-90% gross margins and single-digit P/E ratios, classic cycle-peak signals, but Rasgon asks: what if prices don't crash in 2028, implying a structural shift away from cyclicality.
The semiconductor industry historically destroyed economic value (McKinsey study: -$5B including Intel, -$50B excluding it over 1996-2004); pandemic-era pricing power reversed this — unit shipments flat since 2019 but ASP up 50%, with 70-80% of revenue growth from memory ASP gains and the rest from non-memory price hikes.
Nvidia's pricing is justified by performance-per-dollar: Ampere to Hopper doubled ASP but delivered 10x+ performance; Hopper to Blackwell raised ASP 50% for 3-30x performance; Rubin targets 2x rack cost for 10x performance. The iron law: customers accept price hikes only while performance gains far exceed them.
Hyperscalers are aggressively pursuing custom silicon: OpenAI's first-gen chip reportedly beats Blackwell; Google sells TPUs, Amazon's Trainium works with PyTorch, Meta and Microsoft have programs. Custom chips eliminate vendor margins — $60B/GB build cost drops to ~$20B without Nvidia/Broadcom/Marvell profits, enabling 5x compute for same spend.
Enterprise AI consumption will overwhelmingly shift to buying tokens via API rather than self-hosting models, mirroring the cloud transition; batching efficiency and model scale economics make direct model-service purchase the dominant model for pharma, hedge funds, and others.
Notable quotes
“半导体行业在历史上是一个价值毁灭者”
▸ Notable quotes · 5:54“看2023年或者2024年的行业单位出货量其实跟2019年基本持平,但是行业混合平均售价,也就是ASP高出了50%”
▸ Notable quotes · 6:57“只要性能提升的幅度持续远超价格涨幅,高昂的成本就完全在客户承受范围之内”
▸ Notable quotes · 9:01“现在建设1吉瓦数据中心的成本是600亿美元,但这里面剔除掉芯片厂商比如英伟达、博通、Marvell这些公司的利润之后,实际成本只有200亿美元左右”
▸ Notable quotes · 10:14“企业不会在硬件端进行大规模采购然后自己运行模型,批处理的效率和模型的规模经济效应决定了直接购买Token然后与自有数据深度融合才是绝大多数人消费AI的终极形态”
▸ Notable quotes · 12:34“只要这个铁律不破,整个行业的高定价就是可持续的,但如果哪一天性能迭代放缓了,客户就会开始反击,整个商业闭环瞬间就会崩塌”
▸ Notable quotes · 13:19Chapters
0:40 — Financing Constraints: Financing joins power, supply, and equipment as a fourth constraint slowing AI infrastructure buildout; hyperscalers tap equity markets while emerging clouds without anchor contracts face 15%+ debt costs or no funding at all.
3:02 — Supply Chain Bottlenecks: Substrate materials are the #1 bottleneck for 2+ years due to conservative Japanese investment; memory and optics follow; power is less binding than thought with 50GW global buildout underway and distributed generation options; cleanrooms full until 2027, equipment spending surging but limited by fab space.
5:44 — Memory Economics & HBM: HBM consumes 3-4x wafer area per GB vs DDR5; memory makers at 70-90% gross margins with single-digit P/E — classic cycle peak — but Rasgon questions whether 2028 supply growth will actually crash prices, signaling a possible structural shift.
5:44 — Pricing Power & Inflation: Industry historically destroyed value; pandemic gave pricing power — units flat since 2019 but ASP up 50%, 70-80% of revenue growth from memory ASP, rest from non-memory hikes; no company saw margin compression from cost increases.
8:28 — Nvidia Performance vs Price: Nvidia's pricing justified by performance-per-dollar: 10x+ perf per 2x price (Ampere→Hopper), 3-30x per 1.5x (Hopper→Blackwell), 10x per 2x (Blackwell→Rubin). Iron law: customers pay only while performance gains far exceed price hikes.
9:34 — Custom Silicon & Hyperscaler Economics: OpenAI chip beats Blackwell; Google, Amazon, Meta, Microsoft all building custom chips. Removing vendor margins cuts $60B/GB build cost to ~$20B, enabling 5x compute per dollar. Nvidia's moat (CUDA, scale, balance sheet) is real but share is declining as rivals capture HBM/CoWoS capacity.
11:31 — Enterprise AI Consumption Model: Enterprises will buy tokens via API, not self-host — batching efficiency and model scale economics mirror cloud transition. Coding agents already delivering 10x+ productivity gains in financial modeling.
Speakers
Stacy Rasgon — Senior Managing Director, Bernstein Research
Dylan Patel — Founder, SemiAnalysis
Host (Da Fei) — Host, Best Partner (最佳拍档)