newsroom
$AMD · AMD's inference curve slightly worse than Blackwell but still an $800B company — shifting curve further creates massive value
now playing · $AMD
$ETCHEDbullish· high· posgavin baker
Etched builds full-stack inference racks targeting 10x throughput via low-voltage and cluster-scale memory
Etched vertically integrates chip, board, rack, interconnect, and software to solve inference bottlenecks: thermal throttling via low-voltage design (half GPU voltage) and memory…
$NVDA···bearish· highgavin baker
Nvidia's GPU architecture thermally throttles at high utilization; Etched claims 10x inference advantage via low-voltage design
GPUs are fundamentally limited by thermal constraints: adding more FLOPs doesn't increase performance because chips self-regulate clock speed. Nvidia's general-purpose architectur…
$TSMC···bullish· highgavin baker
TSMC's customer service and yield collaboration model is a structural moat beyond process technology
TSMC's real value is service, not just tech: they run yield experiments on their own dime based on customer recommendations and adopt changes across the line when proven. This par…
$AMD···neutral· lowrob
AMD's inference curve slightly worse than Blackwell but still an $800B company — shifting curve further creates massive value
AMD's interactivity-throughput curve is slightly worse than Blackwell's, yet AMD commands $800B market cap. Etched aims to shift the curve 'much further out,' implying order-of-ma…
$GOOGL···neutral· mediumrob
Google TPU, Meta MTIA, Microsoft Maia, OpenAI Jalapeno are non-existential side projects vs Etched's existential focus
Hyperscaler internal chips lack existential urgency: Google won't fail if TPUs fail, Meta won't fail if MTIA fails. Their FLOP density lags Nvidia's B300 because they only need 's…