Dave Ditzel co-designed the SPARC processor at Sun, cofounded Transmeta (the company that employed Linus Torvalds), and is now back with Esperanto — betting that RISC-V can beat Nvidia at AI inference at a fraction of the power. Their ET-SoC-1 chip has 1,092 RISC-V cores on a single die. They are not competing with the H100 on raw training performance. They are going after inference efficiency, edge AI, and total cost of ownership. That is a smarter fight than the one everyone else is having.
Founded
2014
HQ
Mountain View, California
Total Raised
$133 Million
Founder
Dave Ditzel
Status
Private
Website
www.esperanto.aiTHE ORIGIN STORY
Dave Ditzel had spent 40 years in processor design — Sun, Intel, and Transmeta, which he co-founded in 1995. When the AI wave hit in the mid-2010s, he saw that the entire industry was running on GPU architectures designed for graphics rendering, not neural network inference.
RISC-V, the open-source instruction set architecture developed at UC Berkeley, seemed like the right foundation for a purpose-built AI chip that could be fully customized without ARM licensing fees or x86 compatibility constraints. Ditzel founded Esperanto in 2014 with a team of experienced chip architects to build a massively parallel RISC-V design targeting inference workloads.
WHAT THEY ACTUALLY DO
Chip sales and IP licensing. Esperanto sells its ET-SoC-1 silicon to enterprise customers and data center operators.
Revenue also comes from development partnerships with organizations evaluating RISC-V for inference deployment. SoftBank is both an investor and a strategic partner exploring deployment of Esperanto chips across its portfolio companies.
THE PRODUCTS
ET-SoC-1 — a 1,092-core RISC-V AI inference chip manufactured on TSMC's 7nm process. Designed for natural language processing, recommendation systems, and image classification workloads.
The chip delivers competitive inference throughput at significantly lower power draw than comparable GPU solutions.
HOW THEY GREW
Target data center inference workloads where Nvidia's CUDA moat is less relevant — recommendation systems, natural language inference, image classification — and win on performance-per-watt and total cost of ownership. Leverage the RISC-V open ecosystem to attract software developers.
Partner with hyperscalers and cloud providers who are actively looking to reduce GPU dependency costs.
THE HARD PART
Nvidia's moat is not just the hardware — it is CUDA and the decade of developer tooling built around it. Getting engineering teams to optimize models for a RISC-V architecture when CUDA is the path of least resistance is the real sales challenge.
Esperanto needs compelling inference benchmarks that justify the software migration cost.
MONEY TRAIL
Series A
2020 · Led by Playground Global
$75M raised
Series B
2021 · Led by SoftBank
$58M raised
WHO BACKED THEM
SoftBank led the Series B and is a strategic partner. Earlier investors include Playground Global and other semiconductor-focused venture firms.
Related Profiles
Companies
Anthropic
Anthropic's inference demands are exactly the use case Esperanto targets — high-volume, cost-sensitive inference workloads where power efficiency and cost per token matter more than raw training speed.
d-Matrix
d-Matrix is a direct competitor in the AI inference chip space, also targeting the workloads where GPU efficiency falls short.
OpenAI
OpenAI's massive inference load — serving ChatGPT to hundreds of millions of users — represents the exact market Esperanto is positioning its ET-SoC-1 to address at scale.
Head-to-Head
Compare Esperanto Technologies vs another company.