ASUS ProArt GR1X Could Make Traditional AI Workstations Look Outdated — But There's a Catch

The ASUS ProArt GR1X does make one thing about traditional AI workstations look dated: you no longer need a 600W graphics card in a full tower to load a 70B or 120B-class model on your desk. A 1.1-litre box with 128GB of unified memory can do that at 140W. The catch is speed and growth. Its memory runs at about 300 GB/s, roughly a sixth of a high-end GPU's, so models that fit on a big graphics card generate tokens much faster there. And a tower can take more GPUs, more memory and a different card in two years; the GR1X can only take more storage.

How do the two approaches compare?

Source: ASUS GR1X product page (8 Nov 2026); NVIDIA RTX PRO 6000 and RTX 5090 specifications; GPU memory share from XDA (26 Oct 2026) and PCWorld (29 Oct 2026) on RTX Spark laptops; RTX 5090 pricing from Tom's Hardware and NVIDIA. The GR1X price is our estimate from comparable RTX Spark devices.

The table shows two different philosophies. The GR1X puts a lot of slower memory next to a modest GPU in a tiny, quiet box. A tower puts a fast GPU with its own fast memory in a large, power-hungry box that can change over time.

Where does the GR1X genuinely win?

• Models that don't fit on a consumer GPU: gpt-oss-120b needs about 65GB for its weights, which a 32GB RTX 5090 can't hold. In PCWorld's test, an ASUS ProArt P16 laptop with the same chip finished a gpt-oss-120b research task in 33 minutes 13 seconds, against 52 minutes 51 seconds for an RTX 5090 desktop that ran out of video memory. PCWorld also notes the 5090 is faster in general.
• Power and space: a 140W desktop is easier to run all day in an office than a tower with a 575W or 600W card. Our 575W RTX 5090 post shows what high-power cards ask of a room and a power supply.
• Cost per gigabyte of GPU-usable memory: at our estimate, a 128GB GR1X costs about a third of an RTX PRO 6000 card alone, before the tower around the card.

What is the catch?

The first catch is bandwidth. When a model generates text, every new token means reading the model's active weights from memory. A rough ceiling for speed is memory bandwidth divided by the bytes read per token. At about 300 GB/s, a dense 70B model at 4-bit (about 42.5GB) tops out at about 7 tokens per second for one user. The same model doesn't fit a single 32GB card, but anything that does fit a 96GB RTX PRO 6000 runs against six times the bandwidth. StorageReview measured gpt-oss-120b at 163 tokens per second on an RTX PRO 6000 in LM Studio; a same-chip RTX Spark laptop managed about 63 in Techtesters' run. We explain this further in the 128GB memory trap.

The second catch is growth. LPDDR5X is soldered. There are no PCIe slots for a second GPU, and no ConnectX port, which is what lets two NVIDIA DGX Spark units share a model. When your needs grow, a tower changes a part; a GR1X owner buys another machine.

The third catch is software. The GR1X runs Windows 11 on Arm, where CUDA is a developer preview and vLLM and TensorRT-LLM aren't native. A tower runs the same Linux and x86 stack as most AI servers.

The GR1X makes big models possible in a small box. A GPU tower makes them fast, and keeps them upgradeable. Most buyers need to decide which of those they're paying for.

So are traditional AI workstations outdated?

Not for the jobs they were built for. Training and fine-tuning larger models, serving several users, rendering, simulation and any workload that fits in 32–96GB still favour a discrete GPU with high-bandwidth memory. What has changed is the entry point for loading large models. A developer who only needs to run a 120B-class model for themselves no longer has to buy a 96GB professional card to do it. See why the GR1X might be the wrong choice for serious AI developers for the workloads that still belong on a tower, and GR1X vs RTX PRO 6000 for the direct comparison.

Which should you choose?

• Choose the GR1X if you want one person to run large open models quietly at low power, mostly on Windows, and you accept modest generation speed.
• Choose a GPU tower if your models fit in 32–96GB, you need speed, several users, Linux, or a way to add GPUs later. Our RTX PRO 6000 workstation build guide covers the costs.
• Consider a GB10 system if you want the GR1X's memory size but with Linux, NVIDIA's DGX software and the option to link two units. For the long view, read is this the death of the traditional AI desktop.

Buying in Singapore

The ASUS ProArt GR1X is listed at SourceIT as price on request. For a tower, the RTX PRO 6000 Blackwell Max-Q is S$29,100 and the RTX PRO 5000 Blackwell 48GB is S$14,971.20, and we can quote complete workstations around them. The ASUS Ascent GX10 128GB/1TB is S$10,350 (all listed 8 Nov 2026). RTX 5090 cards are not in stock at the time of writing; ask us about availability. Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Is the GR1X faster than a GPU workstation for AI? Only when the model doesn't fit on the workstation's GPU. For models that fit in a card's memory, a high-end GPU with about 1,792 GB/s of bandwidth generates tokens several times faster than the GR1X's roughly 300 GB/s.

Can I add a graphics card to the GR1X? No. It has no PCIe slots for a GPU, and no Thunderbolt or USB4 for an external one. The only internal upgrade is storage.

Why is unified memory useful if it's slower? Because capacity decides whether a model runs at all. A 120B-class model can't load on a 32GB card, but it can on a 128GB unified-memory machine, at a lower speed.

Does the GR1X use less power than a GPU tower? Yes. ASUS rates it at 140W sustained, while an RTX 5090 alone is rated at 575W and an RTX PRO 6000 Workstation Edition at 600W.

What is the GR1X's price? ASUS hasn't published one at the time of writing. SourceIT's estimate based on comparable RTX Spark devices is about S$9,000–10,800 for 128GB/2TB, including GST. It is not an official ASUS price.