Acrab Bets $350M on Local 100B-Parameter AI — the Numbers Are Impressive, the Proof Is Thin
Singapore’s Acrab unveiled its GΞLIX 1 chip and Agent Box system on July 23, claiming 7.5x faster prefill than M4 Pro and local 100B-parameter inference — backed by $350M in funding and zero independent benchmarks.
On July 23, 2026, Singapore-based Acrab 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 computing platform. The pitch is straightforward: stop renting intelligence by the token from cloud providers and run serious AI locally, on a box that sits on your desk. Acrab emerged from stealth with over $350 million in cumulative financing to build hardware and software infrastructure for agentic AI. Not bad for a company that didn’t exist before 2024.
The founding team isn’t learning on the job, either. Acrab is led by Dr. Ken Phua, whose career in silicon includes heading Asia Applications Engineering at Arm and later serving as co-CEO of Arm China. When someone with that resume bets their next act on edge AI silicon, it’s at least worth paying attention.
What GΞLIX 1 Actually Is
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. That bandwidth figure matters a lot for inference: memory-bound workloads, which large language models absolutely are, live or die by how fast you can move weights from memory to compute. For years, models in the 100-billion-parameter class have required cloud infrastructure, and GΞLIX 1 is designed to bring the latest AI models at this scale into locally operated edge systems.
The headline performance claim comes from Acrab’s own benchmarks. 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. Prefill speed matters most when processing long documents or large context windows — the kind of heavy lifting that agentic workflows demand. If those numbers hold up outside of Acrab’s own lab, that’s genuinely significant.
The word “if” is doing a lot of work in that sentence. Every figure above comes from company-internal testing. No MLPerf submissions, no independent verification, no third-party lab has touched this chip. That’s not unusual for a pre-production announcement, but it does mean the 7.5x multiplier currently exists only on a press release.
Agent Box: The Cloud-Killer Pitch
Powered by GΞLIX, Acrab’s Agent Box is a high-performance personal edge AI center designed for local large model inference, persistent memory, multimodal interactions, and agent orchestration. By replacing cloud AI’s recurring per-token fees, Agent Box is a one-time investment designed to relieve what the company calls “token anxiety.” That framing will resonate with anyone who’s watched their OpenAI bill creep up month after month.
In Dr. Phua’s own words:
Pro tip ✅
“Generative AI helped people find answers. Agentic AI will help them get things done. 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 of Acrab (prnewswire.com)
The economic logic is sound in theory. Cloud API bills compound. Hardware depreciates. But the math only works once you know what Agent Box actually costs — and Acrab hasn’t said. No price, no ship date, no foundry partner disclosed. It’s a very well-funded promise right now.
The Fine Print on “100 Billion Parameters”
There’s a catch worth flagging: the benchmark configuration Acrab uses — Gemma 26B A4B — is a mixture-of-experts model that activates roughly 4 billion parameters per forward pass, not 100 billion. The chip is designed for 100B-class models, but the number Acrab puts in front of journalists uses a configuration running a fraction of that capacity. That’s not fraud, but it is the kind of detail that gets quietly omitted from headlines. Beyond the SoC, Acrab says it has built the software and system layers needed to turn local model inference into working agentic products, including an optimized runtime, developer toolchain, and agent operating system capabilities that help devices understand context, retain memory, and coordinate real-world action.
Acrab has already validated the GΞLIX platform in some demanding real-world deployment environments, and is working toward its first industry adoption and mass production. “In process” is doing similar heavy lifting to that earlier “if.” Qualcomm and Intel already ship neural processors in AI PCs today. Acrab’s differentiation is the parameter ceiling — but until Agent Box ships and ships at a sane price, Acrab is competing in a category it has yet to actually enter.
The Money Behind the Mission
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. Acrab says it will use the capital to accelerate platform development, deepen research and development in next-generation agentic compute systems, expand collaborations with global technology partners, and strengthen its presence in key international markets. Temasek-linked backing gives the company serious institutional weight in Southeast Asia — and meaningful access to manufacturing relationships in the region.
What’s Next
Acrab has the pedigree, the funding, and a chip spec that reads well on paper. The 5nm SoC with 273 GB/s bandwidth and a 20-core Arm CPU is a credible architecture for the problem it’s solving. Agentic systems are starting to appear in personal AI PCs, home hubs, vehicles, industrial operations, and robotics, and in each of these settings, AI that is private-by-design and able to act on its own depends on compute that can run locally and respond in real time. That’s the right market to be building for in 2026. But a startup in the chip business — one of the most capital-intensive and timeline-hostile industries that exists — needs more than impressive press release benchmarks to be taken seriously. The next milestone that matters is simple: ship Agent Box, publish the price, and let someone independent run the benchmarks. Until then, GΞLIX 1 is a very expensive prototype with excellent PR.
