Singapore-based Acrab, the agentic AI compute startup that has quietly raised more than US$350 million, on July 23 pulled the wraps off its first custom silicon: the 5nm GΞLIX 1 edge AI SoC, paired with an "Agent Box" reference system designed to run 100-billion-parameter models directly on a desk.
Silicon built for local agents
GΞLIX 1 is a heterogeneous SoC combining a 20-core Arm CPU, multicore NPU acceleration and 273 GB/s of unified memory bandwidth. Rather than treating CPU, GPU and NPU as separate compute islands, Acrab said the chip coordinates execution across all three with a memory architecture explicitly engineered to keep large models resident locally and responsive under long prompts and 40K+ context windows.
In company testing, GΞLIX 1 hit a prefill rate of 1,416.8 tokens per second on a Gemma 26B A4B configuration with a 40K KV cache and 10K token input — up to 7.5x faster, Acrab claimed, than a Mac mini M4 Pro running the same workload.
Agent Box: an always-on personal AI center
Agent Box is the reference product Acrab built around GΞLIX 1: a compact personal AI system that runs local large-model inference, persistent memory, multimodal interaction and agent orchestration all on-device. Acrab pitched it as a one-time capex alternative to metered cloud tokens, aimed at freeing users from "token anxiety" and keeping personal context out of remote data centers.
"Generative AI helped people find answers. Agentic AI will help them get things done," CEO Dr. Ken Phua said in the launch. "Running models in the 100 billion parameter class on a system small enough to sit on a desk presents a significant computing challenge."
A full-stack bet on edge agents
Acrab is also shipping the software layer around the chip: an optimized runtime, developer toolchain, agent OS primitives, and reference designs for what it calls "agent-native" devices. The company plans to license the platform to makers of AI NAS boxes, AI PCs, smart vehicles, and industrial and service robots.
Timing and competitors
The launch lands as the edge AI silicon race intensifies. NVIDIA's $1.5B Amkor packaging deal underscores its grip on the AI stack, but startups from EdgeCortix to Etched are pushing dedicated inference silicon. Acrab's differentiator is running very large open-source models locally without a data center — a niche that could resonate with device makers looking to avoid ongoing cloud inference bills as agents become always-on. The chip debut also follows Etched's $300M Series C.
Reporting based on Acrab's July 23, 2026 PR Newswire release and coverage from Engineering.com and TechPowerUp.