Rising hardware costs and delays make patience key for next-gen AI rigs
DDR5 at 9200MHz and 12-channel Genoa could beat RTX 5090 in token generation by 2027
An AI hardware enthusiast advises patience: save for 2–3 years as component costs surge. New Chinese DRAM makers mass-produce 64GB RDIMMs, and Micron's 256GB DDR5 modules hit 9200 MT/s. Combined with a 12-channel Genoa system, such speeds could eventually crush an Nvidia 5090 in token generation, using a cheap GPU just for prefill. Meanwhile, Apple's M5 Studio Ultra has been pushed from Q1 to Q2 and now Q3 due to shortages, and RTX Blackwell Pro 6000 prices keep climbing. Yet with demand high, supply will follow—or surplus if demand falls. Either way, the outlook 3 years out is very optimistic.
- Genoa build costs surged from $6,000 to $30,000 due to AI demand; Apple M5 Ultra delayed to Q3 2026.
- New Chinese DRAM makers are mass-producing 64GB DDR5 RDIMMs; Micron's 256GB modules reach 9,200 MT/s.
- Expected 16-channel systems with 12,000 MT/s memory in 2–3 years will crush RTX 5090 in token generation.
Why It Matters
Patience could save thousands and yield a vastly superior inference system within three years.