In the procurement of AI chips, the priority has shifted from performance or power efficiency to simply "when can I get them." Against the backdrop of a high-bandwidth memory (HBM) supply crunch driven by the AI data center boom, buyers' decision criteria are changing dramatically.
AI Chip Supercycle · Supply Crunch
"Just Get Me the Chips" — Availability Now Beats Performance
A structural shortage of AI GPUs and high-bandwidth memory has flipped buyer priorities: customers now say "if you can get me those chips today, I'll buy whatever that chip is." The squeeze is set to outlast 2027.
36–52
weeks lead time for H100 units via resellers
23%
of DRAM wafer capacity now consumed by HBM (up from 19%)
2030
year through which the HBM supply gap is warned to persist
Per-chip memory demand keeps swelling
Blackwell B200 carries 192GB of HBM3E — a 140% jump over the H100's 80GB.
192GB
Blackwell B200 · +140%
The bottleneck: CoWoS packaging capacity
TSMC's advanced 2.5D packaging — essential to fuse GPUs with HBM — is racing to grow, yet still expected to fall short of demand.
→
120–130k
planned end-2026
Wafer starts per month — yet still short of demand, delaying Blackwell & Rubin ramps.
Prices climbing, ripple reaching consumers
+60% in 2025
DRAM prices — a further +30–40% expected in 2026.
8–10×
memory per AI server vs prior levels; even GeForce RTX 50 gaming supply stays tight through 2026.
The takeaway
Advanced memory and HBM are nearly sold out across SK hynix, Micron, Samsung and TSMC for all of 2026 — and the squeeze is expected to outlast 2027.
Supply-side gains
Memory & packaging stocks rose substantially year-to-date.
Buyer-side limited
Chip buyers saw only limited gains; some compromise on specs just to secure supply.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…