Singapore’s Acrab Wants to Kill Your Cloud AI Bill With a Chip
Singapore’s Acrab unveiled GΞLIX 1, a 5nm SoC backed by $350M, designed to run 100B-parameter AI models locally on a desktop without cloud routing.
On July 23, Acrab, a Singapore-based company building agentic AI compute infrastructure, unveiled GΞLIX 1, its first-generation edge AI system-on-chip, alongside Agent Box, a personal edge AI system powered by the company’s full-stack platform. Two years after founding, the company is making a direct argument: stop renting intelligence from the cloud and buy it once, in a box that sits on your desk.
For years, models in the 100-billion-parameter class have required cloud infrastructure. GΞLIX 1 is designed to bring the latest AI models at this scale into locally operated edge systems. That is a bold claim. Whether the silicon fully delivers on it is a separate conversation — and one worth having before anyone gets too excited.
What Is Actually Inside the Chip
GΞLIX 1 is built on a 5-nanometer process and is Acrab’s first SoC designed specifically for edge AI. The SoC features a 20-core Arm CPU, multicore NPU acceleration, and 273 GB/s of unified memory bandwidth. It packs a 3 TFLOPS GPU, a neural processing unit, and multimedia processing engines in its heterogeneous architecture, delivering up to 650 TOPS of AI performance. That is a lot of numbers from one press release, and they all come from Acrab’s own testing — no independent MLPerf results have been published.
In company testing, GΞLIX 1 achieved a prefill rate of 1,416.8 tokens per second under a Gemma 26B A4B configuration with a 40K KV cache and a 10K token input, compared with 188.9 tokens per second on Mac Mini M4 Pro, representing up to 7.5x faster prefill performance. That headline number sounds dramatic, but read the fine print: the test used a mixture-of-experts configuration that keeps only around 4 billion parameters active at once — well below the 100-billion ceiling the chip is marketed around. Decode speed, which determines how fast text actually appears on screen, was not reported. So the 7.5x claim is real, but narrow.
The Business Model Is the Real Story
Powered by GΞLIX, Agent Box is a high-performance personal edge AI center with local large model inference, persistent memory, multimodal interactions and agent orchestration capabilities. By replacing cloud AI’s recurring token fees, Agent Box is a one-time investment designed to eliminate what Acrab calls “token anxiety” and let AI move from an occasional tool into an always-available assistant woven into everyday work and life. That pitch lands differently depending on how heavily you actually use cloud AI — if you are spending serious money on API calls every month, a one-time hardware purchase starts looking rational pretty fast.
Agent Box brings together local language and vision model inference, multimodal interaction, persistent memory, and an orchestration layer that can understand goals, break tasks into steps, and coordinate action across agents, systems, and connected devices. Users’ data and memories remain private and stored locally on the device, while the system grows more capable and customized as context deepens and memories accumulate. That persistent memory angle is genuinely interesting — cloud AI resets with every session unless you are paying for extended context or building your own memory layer.
One catch: Acrab has not disclosed the actual price of Agent Box, a firm ship date, or which foundry manufactures the 5-nanometer silicon. Without a price, the total-cost-of-ownership argument against cloud APIs cannot be fully evaluated. It is a real gap in an otherwise well-packaged announcement.
The People and the Money Behind It
CEO Dr. Ken Phua brings a global technology perspective shaped by his Singapore roots, UK doctoral training, and years of leadership in the silicon industry, including a distinguished career at Arm UK, where he spearheaded Asia Applications Engineering, co-led global IP strategies, and later served as co-CEO of Arm China. That pedigree matters: building a custom SoC from scratch is not something a startup does without deep hardware DNA, and Phua has it.
In June 2026, Acrab announced it had received over $350 million in cumulative financing from global venture capital firms and strategic industry investors, including early backers Vertex Ventures Southeast Asia and India, Vertex Growth, and K3. Vertex, the global venture platform backed by Temasek, was among the earliest investors in Acrab through its Vertex Ventures SEA and India and Vertex Growth funds. Temasek-linked money coming in early is a reasonable signal that this is not just a slide deck with ambitions.
Pro tip ✅
“Generative AI helped people find answers. Agentic AI will help them get things done,” said Dr. Ken Phua, CEO of Acrab. “Running models in the 100-billion-parameter class on a system small enough to sit on a desk presents a significant computing challenge. GΞLIX 1 is designed to deliver the performance, memory bandwidth and responsive local inference required, while Agent Box shows how that capability can become a complete user experience.” — PRNewswire
Bigger Ambitions Than One Box
Beyond the SoC, Acrab has built the software and system layers needed to turn local model inference into working agentic products, including an optimized runtime and developer toolchain, agent operating system capabilities, and reference designs that help devices understand context, retain memory, and coordinate real-world action. The plan is not to sell desktop boxes forever. Founded in 2024, Acrab develops a full-stack compute architecture spanning AI silicon, local LLM inference, operating systems, multi-modality human-machine interfaces, and agent orchestration technologies. The licensing play — targeting AI NAS systems, AI PCs, smart vehicles, and industrial robots — is where the real scale comes in, assuming the platform gains traction.
Why It Matters, With Caveats
The strategic direction here is sound. As AI moves from generating answers to completing tasks, agents increasingly need to understand context, remember preferences, and coordinate tools and devices in real time. Running these capabilities locally can produce faster responses, keep sensitive information under the user’s control, and maintain core functions when cloud connectivity is limited. Acrab is not the only company thinking this way, but it may be the first to ship dedicated silicon at this parameter-class target from outside the established chip incumbents.
What the company still owes the world: independent benchmark results, a price tag, a shipping date, and a demonstration of the full 100-billion-parameter ceiling under real-world decode conditions. The 7.5x prefill claim is a carefully chosen metric on a carefully chosen test. It is not fabricated — it is just incomplete. If GΞLIX 1 can back up the headline with MLPerf numbers and a competitive price, Acrab will have built something genuinely disruptive. Until then, it is a very well-funded, very credible bet on the edge — and one worth watching closely.
