For the first time, a single desktop chip can run state-of-the-art AI models locally without sending data to the cloud, marking a fundamental shift in how AI systems will operate. Acrab, a Singapore-based startup, unveiled GΞLIX 1, a custom system-on-chip (SoC) built on a 5-nanometer process designed to support models in the 100 billion parameter class, along with Agent Box, a personal edge AI system that demonstrates what local-first AI looks like in practice.

The implications ripple across privacy, cost, and reliability. For years, running large language models (LLMs) required cloud infrastructure because the computational demands were simply too high for consumer hardware. GΞLIX 1 changes that equation by integrating CPU, GPU, and NPU (neural processing unit) resources with 273 gigabytes per second of unified memory bandwidth, enough to handle models that previously demanded data center-scale resources.

Why Does Running AI Locally Matter More Than Raw Speed?

The real story behind edge AI (inference happening on devices rather than in the cloud) is not about raw computing power. Capable edge silicon has existed for years; NVIDIA Jetson modules, Hailo accelerators, and Google Coral were all shipping well before edge AI projects reached production at scale. The actual bottleneck was something less visible: the engineering path from a trained model to a deployable artifact that works reliably on constrained hardware.

Consider John Deere's See & Spray technology, which uses computer vision to distinguish weeds from crops and fires individual herbicide nozzles in real time. The system runs on embedded NVIDIA Jetson hardware and cut herbicide use by an average of 59 percent versus broadcast spraying. By 2025, the technology covered 5 million acres and saved 31 million gallons of herbicide. That system could not work in the cloud because the inference window is measured in tens of milliseconds, connectivity in fields is unreliable, and every machine in the fleet needs independent decision-making capability. The latency, autonomy, and fleet lifecycle management are what edge AI is really about.

Three engineering realities define why local inference matters:

• Latency Requirements: Cloud inference involves round-trip network delays measured in tens to hundreds of milliseconds, while edge inference is bounded by local compute in single-digit to low tens of milliseconds, critical for real-time applications like autonomous systems and medical devices.
• Privacy and Data Residency: Cloud inference requires raw data to cross network boundaries and be governed in transit and at rest, whereas edge inference keeps raw data local and only sends derived signals upstream, eliminating data exfiltration risks.
• Cost Structure: Cloud inference scales linearly with request volume, meaning a successful product becomes more expensive to run in direct proportion to its success, while edge inference amortizes costs against hardware already deployed, creating a one-time investment model instead of recurring fees.

Acrab's Agent Box directly addresses the cost problem. By replacing cloud AI's recurring token fees per use, Agent Box functions as a one-time investment with long-term value, allowing AI to move from an occasional tool into an always-available assistant woven into everyday work and life.

What Makes GΞLIX 1 Different From Existing Edge Hardware?

The chip's performance metrics reveal why this matters. In company testing, GΞLIX 1 achieved a prefill rate of 1,416.8 tokens per second under a Gemma 26B A4B configuration with a 40,000 token KV cache and a 10,000 token input, compared with 188.9 tokens per second on a Mac Mini M4 Pro, representing up to 7.5 times faster prefill performance. Prefill speed determines how quickly a model begins responding to a prompt, especially with long context windows, which is why Acrab made it a central design goal.

The chip integrates a 20-core Arm CPU, multicore NPU acceleration, and unified memory architecture engineered specifically for large AI models and agentic workloads. Rather than relying on separate compute components that must communicate across slower interfaces, the unified design reduces bottlenecks that typically plague edge systems.

Agent Box itself represents a philosophical shift in personal computing. For decades, personal computers advanced in predictable steps: faster processors, larger screens, more storage. Agent Box is designed not to run programs, but to host intelligence. It 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, while the system grows more capable and customized as context deepens and memories accumulate.

"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."

Dr. Ken Phua, CEO at Acrab

How to Deploy Edge AI Successfully in Production Environments

Moving from prototype to production edge AI requires solving several engineering problems that go beyond raw compute:

• Model Compression Strategy: Quantization reduces weight and activation precision from FP32 to int8 or int4, cutting memory footprint and increasing throughput; distillation trains a smaller student model against a larger teacher and usually gives the best capability per byte; pruning removes redundant weights or structural units, with structured pruning producing real speedups because it yields dense smaller tensors.
• Hardware Constraints: The accelerator dictates the operator set, meaning a single unsupported operator can erase the entire performance advantage; memory ceilings cap model size before compute does; and power and thermal envelopes bound sustained throughput, since benchmarks measure short bursts while devices in enclosures under continuous load throttle.
• Fleet Lifecycle Management: Updating a model across thousands of devices means staged rollouts, version skew between model and application code, rollback paths that work offline, and updates that survive a power loss mid-write, which is why immutable container-native operating systems paired with lightweight Kubernetes distributions such as K3s, MicroShift, and KubeEdge have become the common substrate.
• Observability Without Data Exfiltration: Inference often moves to the device because the data cannot be centralized, which sits in direct tension with knowing how the model performs in the field; detecting drift requires signals, and deciding which signals can leave the device without recreating the original privacy problem is a design-time decision that is expensive to retrofit.

The sequencing that tends to work in practice is: distill first to set the architecture, quantize to fit the memory envelope, and treat pruning as a targeted optimization rather than a default. This approach balances capability, efficiency, and deployment complexity.

Acrab's broader vision extends beyond Agent Box. The company plans to work with device manufacturers and developers to bring its computing platform into products including AI NAS systems, AI PCs, smart vehicles, and industrial and service robots. The company received over $350 million in cumulative financing from global venture capital firms and strategic industry investors, including early backers Vertex Ventures Southeast Asia & India, Vertex Growth, and K3.

What Does This Mean for the Future of Cloud AI?

This is not a story about cloud AI becoming obsolete. Processing a substantial share of AI workloads locally can reduce dependence on metered cloud inference, lower recurring processing and data transfer costs, and avoid the delay involved in sending every interaction to a remote service. Cloud resources can still be used when a task requires them, allowing developers to choose the right balance between local and cloud execution. The shift is toward hybrid architectures where edge devices handle latency-sensitive, privacy-critical, or high-volume tasks, while cloud systems provide occasional heavy lifting, model updates, and fleet-wide learning.

Healthcare devices already demonstrate this pattern. GE HealthCare received FDA 510(k) clearance in December 2023 for Critical Care Suite 2.1, which runs pneumothorax detection on the mobile X-ray system itself, citing a 57 percent reduction in reporting times for clinically actionable cases and a 17.7 percent increase in clinician detection of small pneumotharaces. Triage happens without patient images leaving the device, addressing both latency and privacy in a single design choice.

The convergence of purpose-built silicon, improved deployment tooling, and a clear economic case for local inference is reshaping how AI systems will be built and deployed. GΞLIX 1 and Agent Box represent the first mainstream expression of this shift, but the underlying pattern is already visible across agriculture, healthcare, retail, and industrial IoT. The question is no longer whether edge AI is possible, but how quickly the ecosystem can move from prototype to production at scale.