Tomasz Tunguz Says AI Infrastructure Bottlenecks Are Shifting From GPUs to Storage

What happened

Venture capitalist Tomasz Tunguz said AI infrastructure bottlenecks are shifting from GPUs to memory, CPUs, and storage. He noted the GPU shock in early 2023 pushed H100 rental rates above $9 per hour and cut server shipments by 22%, then manufacturers shifted capacity to HBM, causing enterprise SSD prices to rise 80% in one quarter and DRAM prices to climb over 60%.

Confirmed

Global impact / market context

Higher storage and memory prices can raise AI companies' spending, reducing cash available for other projects. Server makers may sell fewer units, while chip makers could earn more. Investors might see tech earnings pressured by these cost increases.

Analyst inference

AI buildout relies on many components, not just GPUs, so any bottleneck can slow growth and raise costs. Companies buying servers may delay their capital spending plans, shifting investor attention toward memory and storage suppliers.

Analyst inference

What to watch

1. Tunguz described a relay race pattern, with each constraint freezing supply chain parts for years and locking in a higher cost base, meaning prices could stay elevated long-term. Confirmed
2. Watch whether SSD and DRAM prices keep rising, as past increases suggest ongoing pressure on server costs and possibly lower profit per sale for hardware firms. Proposed
3. If cost constraints persist, server shipments might fall, hurting hardware revenue, while AI developers could seek cheaper storage or computing options to offset expenses. Analyst inference

Evidence

• binance.com