When Acrab unveiled its Agent Box in late July, the two-year-old Singapore chip startup didn't hold back. The pitch was bold: a personal, always-on, edge AI system for running "agentic" workloads locally—built on custom silicon, designed to run large language models on-device. Think of it as a direct shot at NVIDIA's DGX Spark, the $4,699 "personal AI supercomputer" that began shipping this summer through the likes of Dell, HP, and Lenovo.
But here's the thing. More than a week past launch, Acrab still hasn't told anyone what Agent Box actually costs. Or when it ships. And the company has not disclosed wattage, TDP, or basic PSU requirements.
For hardware engineers trying to place their next edge compute bet, that's not a minor oversight.
What We Know (and What We Don't)
The July 22 announcement positioned Agent Box as something Acrab calls a "high-performance personal edge AI center"—built to handle local LLM inference, vision workloads, and what the company terms "agentic" tasks. Translation: persistent memory across sessions, multimodal interaction, the kind of always-on intelligence that theoretically makes cloud APIs look expensive and clunky by comparison. The appeal is obvious, especially to developers who've watched their OpenAI invoices creep upward every quarter.
Under the hood sits GΞLIX 1, Acrab's debut system-on-chip. According to the company's July 23 press release, it's a 5-nanometer design integrating a 20-core Arm CPU, a multicore neural processing unit, and GPU cores—all sharing a unified memory architecture. Acrab claims 273 GB/s of memory bandwidth and says the chip can handle "up to 100B-parameter class" models running entirely on-device.
That's credible on paper. What's less clear is how it performs in the real world.
A Benchmark Built to Avoid NVIDIA
Acrab's headlining performance figure compares Agent Box to Apple, not its ostensible rival. In internal testing, the system clocked 1,416.8 tokens per second during the prefill phase of Gemma 26B—a 7.5x improvement over a Mac mini M4 Pro running the same workload, with a 40,000-token KV cache and 10,000-token input. It's an impressive number, and specific enough to suggest Acrab isn't just making things up.
But it sidesteps the obvious question: how does this thing stack up against DGX Spark?
NVIDIA's system—now available through OEM partners at $4,699—offers 128 GB of unified memory and up to 1 petaFLOP of FP4 compute. The company claims it can handle local inference for models up to 200 billion parameters. The Grace-Blackwell GB10 chip inside runs at a 140-watt TDP, powered by an external 240-watt supply. Those are published, verifiable specs.
As of early August, not a single independent reviewer has put Agent Box and DGX Spark side-by-side on common LLM or agent benchmarks. Acrab tested against Mac hardware. A choice, certainly. But one that leaves serious buyers in an uncomfortable gray zone.
The TOPS Puzzle
Third-party coverage introduces another wrinkle, and not a reassuring one. IT Home and AIBASE, both publishing July 23, cite 650 TOPS for the GΞLIX 1 chip, along with 8 MB of L1 cache and 256-bit LPDDR5X memory running at 8533 MT/s. A Sohu repost mentions ">700 TOPS." Acrab's own press materials? No TOPS figure at all.
Without an official datasheet, the TOPS figure remains unverified.
Then there's the pricing rumor. Multiple sources have floated "$1,000" as Agent Box's target price, framing it as roughly one-fifth the cost of NVIDIA's offering. Compelling, if true. But as Acrab's own launch materials make clear, the company has disclosed no retail pricing, SKU configurations, or even a preorder window.
When a well-funded hardware startup declines to put a number on its flagship product at launch, that silence speaks volumes. And not necessarily good ones.
Power consumption? Also undisclosed. Acrab has not published wattage, TDP, or even basic PSU requirements—leaving potential buyers to guess at operational costs and thermal constraints.
Serious Money, Uncertain Timeline
Acrab isn't operating out of someone's garage. Six weeks before the Agent Box reveal, the company disclosed over $350 million in cumulative financing, backed by Vertex Ventures SEA & India and Vertex Growth. That kind of capital suggests serious ambition and, presumably, serious production plans.
But without availability details, it's impossible to know whether Agent Box is weeks away from shipping or still quarters out. NVIDIA's DGX Spark units, meanwhile, are listed on their marketplace with established shipping partners. Developers know what they're getting.
For CTOs evaluating edge AI infrastructure, that's not a trivial advantage.
Promise Versus Product
The Agent Box announcement lands at a curious moment. Two days ago, Tom's Hardware reviewed the challenges of running local AI agents on commodity mini-PCs—the setup friction, the utility tradeoffs, the yawning gap between hype and hardware reality. The piece underscores something the industry has been slow to admit: local edge AI boxes sound great in principle but remain unforgiving in practice.
Developers want performance, yes. But they also want clarity. Transparent specs. Verifiable benchmarks. A price tag. A ship date.
Agent Box may deliver on all of that, eventually. The GΞLIX 1 architecture looks credible. The 7.5x prefill claim against Mac hardware is specific enough to be testable. Acrab's funding suggests they're serious. But until the company discloses pricing, power draw, availability, and—most crucially—how its system performs head-to-head against the incumbent it's explicitly targeting, Agent Box remains something closer to a concept than a competitor.
Hardware engineers have seen enough vaporware announcements to know the difference. Whether Acrab does remains, for now, an open question.
