Skip to content

Product1 publisher3 min readPublished

Asus asks $1,350 for a 16GB mini PC built to run agentic AI locally

Lenovo's ThinkCentre X Ultra starts at $3,700 in November and the cheapest cluster-capable Mac mini is $1,700. The desk-side inference box is now priced like a workstation. Memory decides which one you need.

The Product Desk · Product desk

Photograph accompanying Asus asks $1,350 for a 16GB mini PC built to run agentic AI locally
Photo: gizmodo.com

What happened

  • Asus will sell its Ascent QN10 mini PC for $1,350 with 16GB of RAM and $1,700 with 32GB, and it puts an 18-core Snapdragon X2 Elite with an 80 TOPS NPU inside both.
  • Gizmodo puts the M6's memory bandwidth at 170GB/s and says the machine may still be held to 7 billion or 8 billion parameter models on its own, even with 32GB of unified memory.
  • Acer told Gizmodo that its SFF RTX Spark, built on a 20-core Grace CPU and a 6,144-core Blackwell GPU, is a second wave machine that may not arrive until early 2027.

Compiled by The Product DeskSomething wrong?How this is made

Why it matters

  • cost Whoever signs for on-desk inference is now approving workstation capex out of a budget line that used to buy a $600 desktop. The entry ticket is $1,350 before any memory upgrade.
  • constraint If a single box tops out around 8 billion parameters, a pilot that needs a bigger model takes two units at every price in this lineup. The plan has to include a second machine.
  • decision Apple forces the cluster question at purchase, because pooling memory requires the version that costs nearly twice what the cheap M6 does.
  • exposure Teams comparing an 80 TOPS rating against Apple's time-to-first-token claim are picking between two vendor-chosen measures. Whoever picked owns the result at the next refresh.

The buyers who created this category thought they were buying a desktop. They cleared out M4 Mac mini stock because it was the cheapest way to attempt local AI, and the shortages ran until Apple's June price hikes [9][10]. That machine launched at $600 [1].

The M6 that replaces it starts at $900, half again as much as the M4 cost at launch two years ago [11][2]. Asus wants $1,350 for a 16GB Ascent QN10 and $1,700 for 32GB, so the memory step alone is $350 [4][3]. Lenovo's ThinkCentre X Ultra, built on the AMD Ryzen AI Max+ PRO 495 and offered with up to 128GB of unified memory, starts at $3,700 when it launches in November [5]. That base price is about 2.7 times the cheapest Asus [4].

Asus calls the QN10 the "world's first" mini PC built for agentic AI, and the claim rests on the 80 TOPS NPU inside its 18-core Snapdragon X2 Elite [2][3]. Apple's claim for the M6 is upwards of five times better AI processing than the previous chip, measured by time to first token [14]. Gizmodo, which reviewed machines in this class, wrote that attempts to benchmark AI by either method are "very vibes-based" [16].

What decides how large a model will fit is memory and bandwidth. Gizmodo puts the M6 at 170GB/s and says that even with 32GB of unified memory it may still be limited to 7 billion or 8 billion parameter models on its own, and that tethering may be the only way to get true on-device performance out of any of these small boxes [15][17]. Pooling processing and memory across several Mac minis requires the M5 Pro version, the same version you need for Thunderbolt 5 ports, and that starts at $1,700 [12][13].

Waiting for cleaner numbers has its own cost. Acer showed an SFF RTX Spark at IFA with a 20-core Grace CPU co-developed with MediaTek and a 6,144-core Blackwell GPU [6]. Nvidia has kept performance details for that chip under wraps [7], and Acer told Gizmodo the machine is a "second wave" PC that may not come until very late this year or, more likely, early 2027 [8].

So the question to settle before the purchase order is which model you intend to run and how much memory it needs. If it fits comfortably in 16GB, the $1,350 Asus is the top of what you have to spend [4]. If it needs more than 32GB, you are choosing between Lenovo's workstation, where $3,700 buys the base model before you get anywhere near 128GB [5], and two clusterable Mac minis at $1,700 each, which comes to $3,400, or $300 under Lenovo's base [5].

What to watch

  • Whether Acer's SFF RTX Spark arrives this year or slips to early 2027, the window Acer gave Gizmodo.
  • A price for Lenovo's 128GB ThinkCentre X Ultra configuration, beyond the $3,700 figure for the base model.
  • Independent time-to-first-token and memory-bandwidth figures for the Asus and Lenovo boxes.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories