While every other AI chip company is bleeding from memory price hikes, Cerebras CEO Andrew Feldman just called it a tail…
Milk Road AI Twitter · Milk Road AI (@MilkRoadAI) · 2026-06-28
Cerebras CEO Andrew Feldman argues that soaring HBM memory prices, CoWoS packaging shortages, and 3nm node congestion are structural tailwinds for Cerebras because its Wafer Scale Engine architecture uses none of those constrained resources.
Appears in
Extraction
Topics: cerebrasai-hardwarehbm-memorysupply-chainnvidia-competition
Claims
- Samsung, Micron, and SK Hynix raised HBM3E prices by roughly 20% for 2026 orders, with SK Hynix's HBM capacity sold out for the year.
- Server DRAM prices surged 60–70% in early 2026, significantly raising the cost of Nvidia-based AI servers.
- Cerebras Wafer Scale Engine uses on-chip SRAM instead of HBM, insulating it from HBM price increases.
- Cerebras does not require CoWoS advanced packaging, avoiding TSMC packaging lead times that now exceed 12 months.
- Cerebras manufactures at 5nm rather than the oversubscribed 3nm node, giving it better supply access than Nvidia-based competitors.
Key quotes
Cerebras CEO Andrew Feldman just called it a tailwind.
At the exact moment the rest of the AI chip industry is getting squeezed from three directions simultaneously—memory pricing, packaging availability, and advanced node access—Cerebras is structurally insulated from all three.
HBM is one of the single most expensive components in any Nvidia-based AI server... and that cost is rising every quarter.