Nvidia ($NVDA) made some thrilling announcements at Computex, showcasing their cutting-edge advancements. However, the semiconductor value chain is feeling the pressure post-event. Here’s a snapshot of what Nvidia unveiled and what’s got investors buzzing.
Highlights from Nvidia at Computex:
Processors:
- Blackwell Ultra GPU expected in 2025.
- Rubin GPU & Vera CPU, successors to the current Grace CPU, set for 2026.
Networking:
- Spectrum Ultra X800 Ethernet switch in 2025.
- Spectrum X1600 Ethernet switch in 2026.
- Major performance leap: 20K TFLOPS up from 4K TFLOPS.
Energy Efficiency Focus:
Nvidia’s new Blackwell chips are significantly more power-efficient. For context, training Chat GPT-4 with Pascal would have taken one thousand GWh—equivalent to a month with a 1GW DC or a year with a 100MW DC. Blackwell reduces this to just 3GWh!
This emphasis on efficiency caught some power-generation investors off guard, raising questions about projected data center (DC) power demand and the future of power efficiency. Stocks like $VST and $NRG are feeling the ripple effects.
New Configurations for Blackwell:
- Air-cooled version: 8 GPUs fit into existing infrastructure (H100) compatible with 15KW of power.
- Liquid-cooled version: 4 GPUs with modular potential.
The market didn’t anticipate an air-cooled version, leading to some surprise among investors in companies like $VRT and $NVT.
DGX Blackwell vs. DGX Hopper Comparison:
- NVLink Domains up 9X (72 vs. 8)😲
- AI FLOPS up 45X! (1,440 vs. 32)😀
- Power consumption up 10X from 10KW to 100KW 🤔
Market Insights:
The new Blackwell DGX with 9X more GPUs (72 vs. 8) hints at the potential for massive cluster data centers (500K to 1 million GPUs), compared to today’s 10K-40K clusters. While these are more power-efficient, the overall power demand will soar.
Large data centers with 40K+ GPUs already consume power equivalent to a small city. Imagine the consumption of a 1 million GPU data center!
Weak hands are getting shaken out across the DC value chain (power generation, liquid cooling, etc.). Despite higher power efficiency, the exponential growth in task size and number means absolute power demand will increase.
Considering all these factors, I see $NVDA as a Strong Buy. Stay tuned for my next update on what this means for $AMD.
Untill next time, Santiago
