Inference chip startup emerging from stealth (with Ed Ludlow)

Etched is emerging from stealth with $800M funding to build rack-scale inference systems using a novel low-voltage chip design. Their technology allows one chip to use memory from other chips in the system, running at under half the voltage of typical Nvidia GPUs. This enables more users per chip and lower cost per token, positioning them as a competitor in the inference market as the industry shifts from training to running AI models.

Yields

implicit

RUT

Oil

Metals

USD

Etched 2.5
Other
Gavin Uberti 7.0

Inference chip startup emerging from stealth (with Ed Ludlow)

ndx

The shift from training to inference represents a massive market opportunity. Etched's technology and $800M funding from sophisticated investors like Jane Street and TSMC-backed venture capital indicates strong belief in continued AI infrastructure buildout and demand for inference compute.

Ed Ludlow

Etched is emerging from stealth with $800M funding, taking on chip heavyweights like Nvidia as the industry shifts from training AI models to running them (inference).
Gavin Uberti

Etched is building rack-scale inference systems with cluster-scale memory technology that allows very low latency communication between chips, enabling one chip to use HBM and SRAM memory of other chips.
Gavin Uberti

We run at under half the voltage of typical Nvidia GPUs, resulting in very large power savings, which means more users on the same chip and more mileage out of the same HBM and SRAM balance.
Gavin Uberti

We think about this as economies of scale - more infrastructure allows much lower cost per user, and as models get bigger, our cluster-scale memory technology gives more and more advantage.
Gavin Uberti

We are seeing best-in-world performance on benchmarks and have had customers test the system. We have a world-class platform team, including our VP who ran Nvidia's DGX and HGX team.

SignalTube

markets at a glance