Rankings/Etched
Etched logo

Etched

NativePrivateUpdated November 3, 2026

Inference chips built for transformers: the Sohu ASIC runs one model architecture far faster than general-purpose GPUs.

CEO
Gavin Uberti
Founded
2022
HQ
San Jose, United States
Total Funding
$1.6B
Employees
201-500
Flag a data issue
AI Infrastructure
44.8Bluescore
Momentum—
Market MapLocked
Rank change (7d)1

Products & Offerings

Frontier Inference Cluster

Liquid-cooled rack-scale inference system combining Etched chips, memory, networking and software, which Etched says runs dense, MoE and state-space models.

InfrastructureEnterpriseFlagship2026

Sohu

Etched's first-generation inference chip on TSMC N4P, originally designed to run only transformer models, with 144GB of HBM3E.

InfrastructureEnterpriseCore Platform2024

Low Voltage Inference (LVI) prefill chip

Compute chip that runs at low voltage to fit more transistors without overheating, speeding up the prompt-processing (prefill) stage.

InfrastructureEnterpriseEmerging2026

Cluster Scale Memory (CSM)

Proprietary memory and interconnect that lets many chips share one fast, low-latency memory pool for the token-generation (decode) stage.

InfrastructureEnterpriseEmerging2026

Executive Summary

What they do

  • Designs complete rack-scale AI inference systems it calls frontier inference clusters
  • Builds custom silicon on TSMC N4P, including the original Sohu chip
  • Splits inference into a low-voltage prefill chip and Cluster Scale Memory decode

Why it matters

  • A credible alternative to Nvidia GPUs for high-volume AI model serving
  • Reports over $1B in contracts and a first rack running at Jane Street
  • Raised about $1B in summer 2026 at a $21B valuation; reportedly fielding offers above $40B

Best fit

  • AI labs and model providers whose costs are dominated by inference
  • Cloud and neocloud providers looking for more compute supply beyond Nvidia GPUs
  • Latency-sensitive operators like trading firms running large models at scale

Watch-outs

  • No independent public benchmarks yet; speed claims come from Etched
  • Only one production customer publicly confirmed so far, Jane Street
  • Inference only; no support for training or fine-tuning workloads

Unlock Full Intelligence

Pro members get the complete briefing

  • Complete analyst verdict
  • Risk assessment
  • All 6 Bluescore dimension scores
  • Traction signals
  • Bluescore history & momentum tracking
  • Unlimited company access
Upgrade · $49/mo
Analyst VerdictMonitor

Etched has working silicon, a first production rack running at Jane Street, more than $1B in reported contracts and very large funding, so its technology is real. Its performance claims have not been independently verified, volume manufacturing is just starting, and it has only one public customer. Most buyers should watch for third-party benchmarks and more customer references. Large-scale inference operators that can run liquid-cooled racks could start early technical talks now.

Signal Provenance

Latest observation date and source for each signal category

Signal CategoryLast UpdatedSource
Funding Signals
Last funding round, Funding stage
Oct 29, 2026Montclare research agents
Hiring Velocity
Headcount, Headcount growth
Nov 3, 2026Montclare research agents
Technical Output
Research papers, GitHub activity
Nov 3, 2026Montclare research agents
News Sentiment
News mentions, News sentiment
Nov 3, 2026Montclare research agents
Partnership Signals
Enterprise partnerships, Distribution partners
Nov 3, 2026Montclare research agents
Revenue Indicators
ARR estimate, Revenue growth
Nov 3, 2026Montclare research agents
Leadership Signals
CEO tenure, Founder-led
Nov 3, 2026Montclare research agents
Competitive Position
Category rank, Market share
Nov 3, 2026Montclare research agents

Why this Bluescore?

6 dimensions, each scored 0–10 against all — companies. Each adds its weight × score to the Bluescore — see methodology

D1 · Business Model & Revenue (21%)
4.3
+9.0 of 21 pts · 43th pct
D2 · Product Depth (21%)
4.9
+10.4 of 21 pts · 56th pct
D3 · Data & Model Moat (19%)
3.4
+6.4 of 19 pts · 2th pct
D4 · Talent & Research (16%)
4.3
+6.9 of 16 pts · 57th pct
D5 · Customer Adoption (13%)
5.3
+6.9 of 13 pts · 50th pct
D6 · Investment Momentum (10%)
5.2
+5.2 of 10 pts · 59th pct

Risks

Performance claims are vendor-reported, with no independent benchmarks published so far
High
Ramping volume production of first-generation silicon and racks risks delivery delays
High
Only Jane Street is a confirmed production customer, so few references
Medium
Proprietary software stack may make it costly to move workloads from CUDA
Medium
Faces strong competition from Nvidia, Google TPUs, Cerebras and MatX
Medium
Feedback