Acrab, a Singapore-based AI computing infrastructure company founded in 2024, has completed a $130 million Series B funding round, bringing its total funding to over $480 million. The round was led by Vertex Growth and other investors. The new capital will be used to expand product capacity, grow the technology ecosystem, and accelerate development of next-generation AI computing platforms.
Just under a month ago, Acrab unveiled its first-generation edge AI chip, GΞLIX 1, along with a personal AI center device called Agent Box. The first-generation products have already begun customer onboarding and are entering mass production preparation.
This pace is unusual for a chip startup. Over the past two years, the most certain and crowded story in AI infrastructure has been in the cloud. Large models keep growing, GPUs get more expensive, and tech giants like Microsoft, Google, Meta, and Amazon continue to expand data centers, with Nvidia at the core of this compute expansion. But AI won't stay in data centers forever.
As models become more compressed and inference efficiency improves, PCs, cars, robots, and other smart devices can handle more AI tasks. The next wave of compute demand is shifting from training larger models to bringing intelligence into every device. Acrab is betting on this growing segment of compute demand.
Cloud is crowded; next AI compute moves to the edge
The first infrastructure war of the large model era was a typical heavy-capital game. Training and running larger models requires more GPUs, higher bandwidth, and more complex data centers. As model scale rises, capital, supply chain, and engineering capabilities raise the bar, making cloud compute a competition only a few giants can sustain long-term.
In contrast, edge AI has historically handled lighter tasks. Fixed functions like photography, noise reduction, and speech recognition can be covered by a single NPU, while complex large model inference still happens in the cloud. But this boundary has been shifting over the past two years.
On one hand, model quantization, compression, and inference optimization continue to improve, allowing the same compute to run more powerful models. On the other hand, PCs, cars, and robots are seeing continuous improvements in memory, bandwidth, and AI compute. Tasks that previously had to be sent to the cloud because models were too large and devices lacked compute are now increasingly able to stay local.
Agents amplify this demand. Personal files, long-term memory, schedules, device status, and real-time interactions originate on the device. A long-running agent that has to go back and forth to the cloud for every file read, tool call, or status update would accumulate latency, inflate token bills, and require continuous upload of personal and device data. As agents shift from occasional Q&A to always-on operation, the per-token cloud model creates unsustainable cost pressure for users.
This doesn't mean edge will replace cloud. A new division of labor is more likely: the most complex and heavy inference stays in data centers, while high-frequency, privacy-sensitive, low-latency, and stateful tasks gradually move to the edge. This directly raises the demand for edge AI compute. The more capable the model, the more tasks can stay local; the more complex local tasks become, the higher the requirements for memory, bandwidth, and compute performance.
Hardware vendors are already preparing for this change. AMD has proposed Agent Computer, and Qualcomm and Nvidia are integrating more complete local AI capabilities into PCs and edge computing platforms. The product forms differ, but the direction is similar: the next AI infrastructure competition will happen not only in ever-larger data centers but also in increasingly powerful terminals.
Acrab has chosen this position. It hasn't entered the crowded cloud GPU battlefield. Instead, leveraging experience from large-scale server chip systems, it designs computing platforms around edge large models and agents, targeting the high-end edge AI server market and betting on long-term upgrades in edge compute as model capabilities continue to sink.
This technical barrier is much higher than traditional consumer electronics. Acrab brings a team with large-scale server chip and system software experience into a technical deep-water zone that traditional 3C teams find hard to reach.
After agents, terminals need more than just an NPU
While edge model capabilities advance, the way AI consumes compute is also changing. In the past, large model applications were mostly one-question-one-answer: the user gives input, the model completes inference and generates output, and the computation ends. When agents truly start working, a single request stretches into a continuous task chain.
Microsoft Azure Research's recent paper, "Architectural Implications of Agentic AI Workflows," observed this change in production environments. The study found that agent tasks constantly switch between model inference, tool calls, and task orchestration, repeatedly crossing CPU-GPU boundaries. Meanwhile, compute demand is no longer as stable as traditional inference; it shows noticeable waiting, recovery, and short-term peaks.
This means edge chips for the agent era can't just boast a TOPS number. CPU, NPU, memory, and software environment need to work together. Whether a device can sustain an agent over long periods is increasingly a system-level question.
This aligns with the view of Acrab's CEO, Ken Phua. He previously led Asia-Pacific application engineering at Arm UK and participated in global IP strategy, then served as co-CEO of Arm China. In his view, as AI enters the heterogeneous computing stage, execution efficiency depends not only on the NPU but also on smooth CPU-NPU collaboration.
Therefore, Acrab's recently released AI chip GΞLIX 1 doesn't just emphasize 700 TOPS peak compute. Unified memory, CPU-NPU collaboration, prefill (input processing), and agent orchestration are also built into the product design.
This product philosophy is also reflected in the company name. Acrab is an acronym for Agentic Compute, Reasoning, Action, and Brain, and also comes from the star Beta Scorpii, a multiple-star system symbolizing different computing units working together.
Acrab also published a long-context test. Running the Gemma 26B A4B model with a 40K token context cache and 10K token input, the GΞLIX 1 platform achieved a prefill speed of 1416 tokens per second, claimed to be 7 times faster than leading edge AI computing platforms. Notably, these results are from the company's own tests; actual performance may vary with model loading, software scheduling, and specific tasks, and third-party verification is still needed.
But Acrab's focus on prefill, unified memory, and heterogeneous collaboration indicates it's targeting not just whether a model can run, but the overall compute efficiency required when long-context and agent tasks truly enter terminals. In other words, Acrab isn't betting on "the edge needs a faster NPU" but rather that when local AI moves from one-shot inference to long-running work, terminals will need a new computing foundation.
From GΞLIX 1 to Agent Box, Acrab moves toward mass production
The name GΞLIX is derived from "Helix," borrowing the imagery of DNA evolution and adaptation to express the platform's goal of supporting continuous AI evolution, environmental adaptation, and long-term planning.
Based on GΞLIX 1, Acrab also built Agent Box. Defined as a Personal AI Center, it's a small edge AI device that sits on a desk. Powered by GΞLIX 1, it can run models with hundreds of billions of parameters locally, integrating long-term memory, multimodal interaction, tool calling, and agent orchestration.
In the launch demo, a user asked for help creating a space-themed gift for a child. The agent provided ideas, helped choose a plan, and then called relevant tools to advance the creation. Each step is simple; the challenge is whether a single natural language instruction can be decomposed into continuous actions and whether task state can be passed between models and tools.
This is what distinguishes Agent Box from an ordinary "local large model computer." According to Acrab, its software platform covers local large models, optimized runtime, development tools, and agent orchestration. Agent Box serves both as a complete device and as a reference system: they put GΞLIX 1, models, and software stack into one machine to verify that complete agent tasks can run locally, then replicate this capability to AI PCs, AI NAS, cars, industrial equipment, robots, and other terminals.
For Acrab, the real commercial opportunity isn't just Agent Box itself, but whether GΞLIX 1 and its software platform can be adopted by more device manufacturers. The platform is now moving toward mass production validation.
According to Acrab, the GΞLIX chip was taped out in November 2025, followed by power-on and system verification. The company claims the chip passed 12 key verifications within 6 hours of power-on and booted Linux and Android systems within 24 hours. In July of this year, GΞLIX 1 and Agent Box were officially launched. The first-generation computing platform has begun customer onboarding and is advancing toward mass production.
The latest $130 million Series B comes at this milestone: the first chip is made, the reference system is unveiled, and the next steps are expanding product capacity, growing the ecosystem, and continuing R&D on next-generation AI computing platforms.
Vertex Growth has invested since Acrab's early stages and has continued to invest as the company hits different technical milestones. Investors previously disclosed that GΞLIX has been validated in some demanding real-world deployments and is progressing toward first industry adoption and mass production.
Acrab, founded less than two years ago, has gone from chip tape-out to product launch and customer onboarding, while raising over $480 million in total funding. It now stands at the threshold of commercial validation. The key questions ahead are which mass-produced devices GΞLIX 1 will enter, when the first customers will appear, and whether these projects can generate sustained shipments.
The $480 million gives Acrab the chips to continue betting on edge compute. A true computing platform for agentic AI is becoming a reality.