TickerTain — Make Analysis Fun Again

▶ 26:40 · Semiconductors · Vertical integration from wafer to rack is necessary to hit gigawatt-scale production velocity

episode shorts · 1

Building a Company for the Obsessed

now playing · Semiconductors

Semiconductorstailwindscore 9/10gavin baker

Low-voltage inference breaks thermal wall: half GPU voltage enables 4x power reduction and sustained peak FLOPs

GPUs thermal throttle because voltage scales quadratically with power. Etched developed a new power delivery mechanism to run at under half the voltage of any AI chip, allowing 2x+ FLOP den…

Memory & Storagetailwindscore 9/10gavin baker

Cluster-scale memory via custom interconnect (5x lower latency than Blackwell) unlocks near-linear scaling for decode

Decode is memory-bound. Current GPUs waste cluster memory bandwidth due to 4,000ns chip-to-chip latency (Blackwell). Etched built a full custom interconnect stack above L2 Ethernet, cutting…

AI Infrastructuretailwindscore 9/10rob

Inference becoming majority of global GDP; token production is the new oil

Inference will become the largest market in the world within a decade. Whoever produces the most tokens at lowest cost per watt becomes the most valuable company. We are at 1/1000th of glob…

Semiconductorstailwindscore 8/10rob

Vertical integration from wafer to rack is necessary to hit gigawatt-scale production velocity

Etched builds chip, board, cold plate, interconnect, rack, and production line in-house. This eliminates vendor dependencies, enables 24/7 parallel development (40 days from silicon to infe…

Grid & Power Infrastructuretailwindscore 8/10rob

Power availability is the hard constraint: tokens per megawatt becomes the key metric for national competitiveness

Data center power scales non-linearly: 100MW to 1GW is easier than 1GW to 10GW. Etched's low-voltage + cluster-memory approach maximizes tokens/megawatt. Nations will compete on inference e…

Semiconductorsmixedscore 7/10rob

Supply chain non-zero-sum: Etched uses 4nm/alternative HBM vs Rubin's 3nm, adding capacity rather than competing for same wafers

Leading-node capacity is scarce but not zero-sum if designs target different process nodes and memory types. Etched chose 4nm and different HPM than Nvidia's Rubin (3nm) to avoid direct waf…

AI Infrastructuretailwindscore 8/10rob

Agent fleets with long time horizons require massive cluster-scale memory and FLOPs; 2027 crossover to majority-agent workforce

Inference-time compute scaling (e.g., 6-month tasks compressed to days) plus massive agent concurrency (millions per task) demands cluster-scale memory for low latency and high FLOPs for th…

←