Ryzen AI Max+ PRO 495 mini PCs with 192GB RAM emerge — Acemagic, GMKtek, Framework debut ultra-expensive 'supercomputers' that can handle 300B-parameter

What happened

Three PC makers — Acemagic, GMKtek and Framework — are separately bringing mini PCs to market built on AMD's Ryzen AI Max+ PRO 495 processor, configured with up to 192GB of unified memory. The specification is significant because it allows these compact desktops to run AI models with as many as 300 billion parameters entirely on-device, without relying on cloud infrastructure.

The three vendors are pursuing the same underlying chip and memory ceiling but packaging it into their own small-form-factor systems, positioning these machines as a new category of desktop hardware aimed squarely at local, large-scale AI inference rather than general productivity use.

Why it matters

Running a 300-billion-parameter model locally has, until now, been the preserve of data-centre-grade GPU clusters. Unified memory at this scale changes the calculus: it lets a single desktop-class machine hold and process models that previously required distributed cloud compute or multiple discrete GPUs. For technology leaders, this signals a maturing path toward on-premises, privacy-preserving AI deployment — relevant for regulated sectors, sensitive data workloads, or organisations wary of recurring cloud inference costs.

It also reframes what "edge AI" can mean in practice. Rather than trimming models down to fit constrained hardware, these systems suggest the constrained hardware is catching up to the models — narrowing the gap between what's technically deployable in a data centre and what's deployable on a desk.

By the numbers

• 192GB of unified memory available in the top configuration across these mini PCs.
• 300 billion parameters — the scale of AI model these systems are reported to be capable of running locally.
• 3 vendors — Acemagic, GMKtek and Framework — have each announced systems built around the same AMD Ryzen AI Max+ PRO 495 chip.

The Renascence take

The headline here is silicon, but the real story is about who gets to control the AI experience layer. When inference can happen on a desk rather than in someone else's cloud, the constraints that have shaped AI product design — latency, data residency, per-query cost — start to loosen for the organisations that can afford the hardware.

Most coverage of this launch will focus on specs and price tags, but the more interesting question for service leaders is what changes when AI inference is no longer metered by a cloud bill. Local, high-capacity inference removes one of the biggest behavioural frictions in AI-assisted service — the hesitation to run a model "just to check," because it costs nothing marginal to do so. That shifts AI from an occasional decision-support tool into something closer to ambient infrastructure. Operators experimenting with AI-heavy customer or employee workflows should treat this as an early signal to pilot on-premises inference for their most sensitive or highest-volume use cases now, rather than waiting for cloud pricing or data-governance pressure to force the decision later.