Acrab’s Edge Chip Aims to Run 100B AI Models Locally

Acrab, a Singapore startup that emerged from stealth in June 2026 with more than $350 million in backing, unveiled its first silicon on July 23, 2026: GΞLIX 1, a 5-nanometer edge processor, alongside Agent Box, a desktop system built around it. Both are aimed at the same goal — running open-source AI models in the 100-billion-parameter class on hardware small enough to sit on a desk, without routing data through a cloud service.

That is a wager against the prevailing shape of generative AI, where models that large live in data centers and users pay by the token. Models at this scale have generally required cloud infrastructure; Acrab’s case is that the economics and the privacy math now favor pulling them onto local hardware, where responses return faster, data stays on the device, and the system keeps working when the network does not.

Acrab is led by Dr. Ken Phua, who ran Asia applications engineering at Arm and later served as co-CEO of Arm China, and the chip reflects that lineage. Rather than bolting on a discrete accelerator, GΞLIX 1 integrates CPU, GPU and NPU on a single die with one shared pool of memory. “Running models in the 100 billion parameter class on a system small enough to sit on a desk presents a significant computing challenge,” Phua said.

Inside the silicon

GΞLIX 1 pairs a 20-core Arm CPU with a multicore neural processing unit and 273 GB/s of unified memory bandwidth, according to Acrab, and is built to spread one large model across all three compute blocks at once. The company treats memory bandwidth and time-to-first-token (the lag before a model starts responding to a long prompt) as the binding constraints at this scale, and says it tuned the chip to hold power low enough to run all day.

The single performance figure Acrab disclosed is a vendor number, not an independent test. In its own benchmarking, the company reported a prefill rate of 1,416.8 tokens per second against 188.9 on a Mac Mini M4 Pro, or roughly 7.5 times faster, using a Gemma model configuration it labels “26B A4B” with a 10,000-token input. That setup keeps only about 4 billion parameters active at a time, well short of the 100-billion-parameter ceiling the chip is sold on, and the number covers prefill alone. Acrab published no decode-speed figure and no MLPerf result, so the claim rests on the company’s own measurement.

A box you buy instead of a bill you pay

Agent Box is the first product built on the platform, and it carries Acrab’s sharpest economic argument. The company casts it as a one-time purchase that removes the recurring per-token fees of cloud AI: it runs language and vision models locally, holds a persistent memory of the user’s data on the device, and coordinates tasks across connected systems. It is aimed at homes and personal workspaces, and Acrab says it grows more tailored as its store of context accumulates. The pitch is capital spending in place of an operating bill — buy the hardware once rather than metering every request.

What Acrab has not said is what Agent Box costs, when it ships, or which foundry makes the 5-nanometer chip. Those are the numbers a buyer would need to actually run the comparison. Until there is a price and a delivery date, the trade between a fixed hardware cost and an open-ended cloud subscription is a claim rather than a calculation.

A wider bet on the edge

The pitch enters a contested field. Qualcomm (QCOM ) and Intel (INTC ) already ship neural processors in AI PCs, and a class of edge AI boxes runs smaller models on-device today; what Acrab claims to change is the parameter ceiling, moving 100-billion-class models onto a single local system rather than the far smaller models that run comfortably on the silicon shipping in phones and laptops now. Acrab says GΞLIX 1 and its software stack are meant to serve as a horizontal foundation for other manufacturers, with plans to reach AI network-storage boxes, AI PCs, smart vehicles, and industrial and service robots. The ambition lands as the broader infrastructure conversation shifts from training toward inference, and as component makers push more silicon into edge and on-device form factors.

The company has raised over $350 million cumulatively since its 2024 founding, a total it disclosed in June 2026, with early support from Vertex Ventures SEA & India and Vertex Growth, the Temasek-linked funds, plus an investor it identifies as K3. Acrab says the GΞLIX platform has been validated in real deployments and is moving toward its first industry adoption and mass production. What it has not yet shown is the chip running a 100-billion-parameter model at the speeds its pitch depends on, measured by anyone other than Acrab.