Acrab Agent Box Mini PC: CATL-Backed Startup Challenges Nvidia DGX Spark

<p><em>Originally published at <a href="https://tekmag.thsite.top/acrab-agent-box-mini-pc-catl-backed-startup-challenges-nvidia-dgx-spark/" rel="noopener noreferrer">https://tekmag.thsite.top/acrab-agent-box-mini-pc-catl-backed-startup-challenges-nvidia-dgx-spark/</a></em></p> <p><strong>Singapore-based startup Acrab has unveiled the Agent Box, a mini PC powered by its custom G≡LIX 1 chip that claims to run 100-billion-parameter AI models entirely on-device at one-fifth the cost of Nvidia's DGX Spark. Backed by CATL, the world's largest EV battery maker, the device delivers 700 TOPS of processing power in a compact form factor.</strong></p> <p>Acrab announced the Agent Box on July 23, 2026, according to PC Watch, which first reported the news. The company positioned the device as a direct competitor to Nvidia's $5,000 DGX Spark, claiming similar performance at roughly one-fifth the price and half the power consumption. TechRadar Pro and Robotics Media covered the announcement in early August 2026.</p> <p>The startup has reportedly raised over $350 million, per Robotics Media, though no official funding announcement was located. The company's website at acrab.ai confirms the product focus but does not list detailed specifications on its homepage. CATL's involvement appears to be inferred from copyright text visible in the announcement video rather than a formal press statement from the battery manufacturer.</p> <p><strong>Key Takeaways</strong></p> <ul> <li>Singapore-based Acrab unveiled the Agent Box mini PC on July 23, 2026</li> <li>Custom G≡LIX 1 chip built on 5nm process with 20-core Arm CPU and 3 TFLOPS GPU</li> <li>Claims 700 TOPS processing power, comparable to Nvidia DGX Spark at one-fifth the cost</li> <li>Runs 100B parameter models entirely on-device without cloud reliance</li> <li>Internal benchmark shows 1,416.8 tokens/sec prefill—7.5x faster than Mac mini M4 Pro</li> <li>Backed by CATL, the world's largest EV battery manufacturer</li> </ul> <h2> The Agent Box hardware </h2> <p>The Agent Box centers on Acrab's proprietary G≡LIX 1 system-on-chip, fabricated on a 5nm process node. The chip integrates a 20-core Arm CPU, a GPU delivering 3 TFLOPS of compute, and a multicore neural processing unit optimized for <a href="https://tekmag.thsite.top/openai-jalapeno-broadcom-custom-llm-inference-chip-challenges-nvidia/" rel="noopener noreferrer">large language model</a> inference with VAE vector acceleration.</p> <p>Memory architecture includes 273 GB/s unified memory bandwidth, 8 MB of L1 cache, and 768 GB/s of L2 cache bandwidth. Acrab claims the total processing power reaches 700 TOPS, making it suitable for running large language models entirely on-device without relying on cloud infrastructure.</p> <h2> Performance claims against Nvidia and Apple </h2> <p>Acrab's primary competitive claim targets Nvidia's <a href="https://tekmag.thsite.top/nvidia-rtx-spark-superchip-computex-2026/" rel="noopener noreferrer">DGX Spark</a>, a $5,000 AI-focused mini computer. According to the company's internal testing, the Agent Box achieves performance close to the DGX Spark while consuming roughly half the power and costing approximately one-fifth of the price.</p> <p>In specific benchmarks, Acrab reported a prefill rate of 1,416.8 tokens per second using the Gemma 26B A4B model with a 40,000 token KV cache and 10,000 token input. The company states this is 7.5 times faster than an Apple Mac mini M4 Pro, which achieved 188.9 tokens per second on the same workload. No independent benchmark validation has been found for these claims.</p> <h2> Software stack and use cases </h2> <p>The Agent Box runs a software stack that includes optimized AI runtimes, developer toolchains, OS primitives, and orchestration frameworks designed for autonomous agent use cases. Acrab demonstrated voice-command-driven 3D modeling and printing, robotic vacuum control, and smart home integrations including lights, air condition