Keep pulling the thread on Brian Leuler.
The company that produces the most tokens is predicted to become the most valuable company in the world.
Etched has developed a new power delivery mechanism called "low voltage inference" that allows its chips to run at much lower voltages than GPUs.
Etched's custom interconnect can reduce chip-to-chip latency by more than 5x compared to existing solutions like NVIDIA's Blackwell.
Etched's first-generation chip, using its low-voltage inference technology, runs at less than half the voltage of any other AI chip.
Rob predicts that inference clusters will rapidly scale from current 8-chip or 72-chip (NVL72) domains to thousands or tens of thousands of chips.
Rob asserts that for software companies leveraging AI, the Cost of Goods Sold (COGS) for an incremental user will no longer be near-zero and will instead be a significant cost driven by inference.
Rob predicts that over the next decade, the inference market will become the largest market in the world.
Etched's strategy is to vertically integrate its entire product, and it is the only startup currently building both its own chips and its own rack.
Etched hired Brian Leuler, who previously built all of NVIDIA's HGX and DGX systems.
NVIDIA's HGX and DGX systems account for approximately 80% of the company's revenue.
As part of its "prefetching" strategy, Etched shipped complete racks without its custom chips to customer data centers to pre-configure networking, CPUs, storage, and software.
Before receiving its silicon, Etched emulated its full chip on a cluster of over 700 FPGAs and ran a dozen different models using its complete inference software stack.