A 100-billion-parameter AI model has, until this year, needed a warehouse of GPUs to run. A Singapore startup says it can fit one in a box on your desk. This month investors handed it another $130 million to prove the point.
Its GΞLIX 1 chip targets 100-billion-parameter models on local hardware, a class of model that has until now required data-center GPUs.
The bet is that agents, not chat, are the real edge opportunity, and that a one-time device beats recurring cloud token fees.
Acrab is not arguing that the cloud is wrong. It is arguing that the next wave of AI, the agent that acts rather than answers, may belong somewhere else entirely.
TIMELINE: Acrab
────────────────────────────────────────────────────────────
2024 Jun 2026 23 Jul 2026 6 Aug 2026 NOW
🔧 💰 🔬 💰 ◉
Founded $350M+ GΞLIX 1 $130M Commercial-
stealth SoC + Agent Series B ization
emerge Box unveiled closed & revenue
────────────────────────────────────────────────────────────
Source: Acrab releases via PR Newswire; DealStreetAsia; TechNode GlobalAcrab's chronology from founding to its Series B, 2024 to August 2026.
From stealth to a half-billion war chest
Acrab was founded in 2024 and stayed quiet. That changed in June 2026, when it emerged from stealth with more than $350 million in cumulative financing from Vertex Ventures Southeast Asia & India, Vertex Growth, and K3. The money was a vote of confidence in a specific thesis: that AI compute is consolidating in the wrong place.
The company builds a full stack. Purpose-designed silicon sits at the bottom. Edge AI models sit in the middle. Software orchestration ties them together so agents can run, remember, and act without calling a remote data center. Most AI infrastructure bets pile capital into bigger clouds. Acrab is betting the opposite direction.
The GΞLIX 1 bet
On 23 July 2026, Acrab unveiled GΞLIX 1, its first-generation edge AI system-on-chip (SoC). The chip is built on a 5-nanometer process and is designed to run models in the 100-billion-parameter class entirely on local hardware. For context, models at that scale have, until now, been a cloud-only affair.
Largest model class on GΞLIX 1 at the edge
Models in this class previously required cloud GPUs. Source: Acrab, July 2026.
The silicon is concrete. GΞLIX 1 pairs a 20-core Arm CPU with multicore neural processing unit (NPU) acceleration and 273 GB/s of unified memory bandwidth. It integrates compute and memory in one architecture. That choice decides whether a 100-billion-parameter model feels instant or sluggish on a device.
Alongside the chip, Acrab showed Agent Box, a personal edge AI system. It does local large-model inference, keeps a persistent memory, handles multimodal input, and orchestrates agents on the device itself. The pitch is privacy and uptime: sensitive data never leaves the box, and the assistant keeps working when the connection drops.
Our goal is to give device makers and developers the foundation to bring agentic intelligence into many different products and environments.— Dr. Phua, CEO of Acrab
The $130 million vote
On 6 August 2026, Acrab closed a $130 million Series B led by its existing backers, Vertex Ventures Southeast Asia & India and Vertex Growth, with institutional investors across Europe and Southeast Asia joining. The company says it has clear paths to industrial deployments and expects revenue within 2026.
Acrab's latest raise
Brings cumulative financing past $350 million. Source: DealStreetAsia, Aug 2026.
The raise is not a research grant. Acrab describes it as the moment it moves from building foundations to shipping product. The capital funds scaling, ecosystem expansion, and a next-generation compute platform. For a silicon company, that is the expensive part: tape-outs, supply agreements, and the long road from a working chip to a shipped device.
Turning points
Three things separate this from the usual AI-hardware story. First, the target is agents, not chat. Second, the compute is local by design, which keeps data on the device and keeps working when connectivity drops. Third, the economic pitch is a one-time device rather than a recurring cloud token fee, a model Acrab frames as relief from what it calls token anxiety.
Each point is a bet against the current center of gravity in AI. Centralized training and inference have pulled in hundreds of billions of dollars. Acrab is wagering that the action moves to the edge just as the agent era begins. That is a coherent story. It is also the claim every edge-AI chip startup makes, and most never ship at volume.
| Dimension | Cloud inference | Acrab edge (GΞLIX 1) |
|---|---|---|
| Where models run | Remote data-center GPUs | On-device SoC |
| Model scale | Any size | Up to 100B params locally |
| Data residency | Leaves the device | Stays on device |
| Cost model | Recurring token fees | One-time device |
What it means for the edge-versus-cloud compute race
The tension here is real. As we wrote on AMD's agentic-AI data-center wave earlier this month, the centralized play is winning today's headlines and today's revenue. Acrab is making the opposite wager: that the agent people actually use will live next to them, not in a distant facility.
For investors, the question is not whether edge AI is real. It is whether a startup can out-engineer the incumbents who are also racing toward on-device inference. The list is not short: Qualcomm, Apple, and Nvidia all treat the edge as a priority, and they have fabrication relationships and volumes Acrab does not. Acrab's answer is a full-stack bet, silicon plus models plus orchestration, rather than a single layer.
Acrab has the capital and the silicon. What it lacks is a track record of shipping at volume, and proof that 100-billion-parameter models on a desk are something customers will pay for. The $130 million does not settle that. It buys Acrab the time to find out.
