A few years ago, if you wanted to run a very large AI model at home, your options were pretty awkward. You could buy a workstation with expensive NVIDIA graphics cards, build a second-hand multi-GPU setup, or rent a machine in the cloud and watch the bill grow. Even after spending thousands, you might discover that the model you wanted to test wouldn’t fit into your GPU memory. The processor was rarely the only problem. Getting enough fast memory was the headache.
ASUS has put a rather different computer on the table. Its Ascent GX10 packs 128GB of unified memory, an NVIDIA Blackwell GPU and CUDA support into a case measuring roughly 150 × 150 × 51mm. That’s about 1.15 liters in volume. It runs NVIDIA’s DGX software environment, targets local AI development and uses a USB-C connection for power. ASUS says the platform can handle AI models with as many as 200 billion parameters under suitable conditions. This is the sort of specification sheet that would usually belong to a machine taking up far more space.
The part that caught my attention was the combination rather than any one number. Lots of mini PCs are small. Some have unusually large memory options. NVIDIA hardware has supported CUDA for years. Putting 128GB of shared memory and NVIDIA’s AI software in something that looks like a little desktop box is much less common. And that USB-C power connection sounds almost too convenient, until you check what kind of USB-C adapter it needs.
Access without medium partner: ASUS Put 128GB of Memory In CUDA
A Full AI Development System in a Very Small Case
The Ascent GX10 is built around NVIDIA’s GB10 Grace Blackwell Superchip, the same general platform behind NVIDIA’s DGX Spark. It combines a 20-core Arm CPU with a Blackwell GPU designed for AI work. The CPU uses ten Cortex-X925 cores and ten Cortex-A725 cores, rather than the Intel or AMD x86 processors you probably expect in a desktop. ASUS lists up to one petaflop of FP4 AI computing performance, a theoretical low-precision figure under specific conditions, not a promise that every AI program will run at that speed.
That difference matters. A petaflop is a huge number, but it doesn’t tell you how many words a local language model will produce each second. It doesn’t tell you how long an image model takes to generate a picture either. Both depend on the model, precision, memory use, software and many other settings. Manufacturers like a single big compute number because it fits nicely in a product announcement. Developers need benchmark results with the settings written down.
Still, the engineering is worth paying attention to. The GX10 doesn’t have a conventional full-size graphics card hanging off a motherboard. The chip and memory are part of a tightly integrated platform, and ASUS uses active cooling to keep that hardware running inside the tiny enclosure. You get HDMI, USB-C data connections, networking and a dedicated high-speed interface for linking the machine with another GX10. The box weighs about 1.5kg, which is light enough to carry between two desks without turning it into a project.
And the operating system is part of the story. ASUS ships the GX10 with NVIDIA DGX OS, a Linux-based environment that comes with NVIDIA’s AI software stack. That makes more sense for machine learning than installing Windows and hoping every tool behaves the way it does on your gaming PC. The catch is that some users will need to adjust to Linux and Arm64 packages. This is a specialist desktop, even if the case looks friendly enough for a normal office.
The 128GB Memory Is More Useful Than the Petaflop Claim
I think ASUS picked the right headline feature with the 128GB memory configuration. Local AI runs into memory limits very quickly. A GPU with 16GB or 24GB of VRAM may be fast, but a sufficiently large model will refuse to fit. You can offload some weights to ordinary RAM, although moving them back and forth can slow things down badly. You end up owning a fast GPU that spends too much time waiting for data.
The GX10 uses unified memory, meaning its CPU and GPU share one large memory pool. The processors communicate through NVIDIA’s NVLink-C2C connection. It is different from a conventional desktop with separate system RAM and graphics memory, and the design gives the GPU access to far more capacity than a typical consumer graphics card carries on its own. For developers experimenting with large language models, that changes which models they can attempt to load locally.
Take a dense 70-billion-parameter model. Its weights alone would take about 140GB at 16-bit precision. A 4-bit quantized version needs roughly 35GB for raw weights, with additional memory used for metadata, attention caches and runtime buffers. A 24GB graphics card cannot hold that whole model in its dedicated VRAM. A 128GB shared memory system has considerably more room, although the amount you can actually use depends on the software and other memory demands.
Or think about a 120B model. Quantization can make one fit in a memory pool like this, but long context windows create additional pressure. Different model architectures use memory differently, and a mixture-of-experts model can have a large total parameter count without activating all those weights on every token. That is why a claim such as ‘supports 200B models’ needs context. It is a statement about what the system can accommodate with the right model format and setup, not proof that every 200B model will run quickly.
The machine has 128GB. That doesn’t mean all 128GB are free for model weights, and it certainly doesn’t mean the GPU behaves like a desktop card with 128GB of dedicated GDDR memory. The operating system and running applications use part of the pool. The memory is shared, and the access pattern is different. I would much rather know a model’s actual memory use before buying hardware based only on its parameter count.
CUDA Is the Reason This Is Different From Other High-Memory Mini PCs
Large-memory mini PCs are no longer quite as unusual as they used to be. AMD-based machines have made high-capacity shared memory more accessible, and Apple Silicon systems have developed a decent following among people running models locally. Some of those computers make more sense for particular workloads. The ASUS GX10 has a specific attraction, though: CUDA.
CUDA is NVIDIA’s platform for running compute workloads on its GPUs. Many AI applications and libraries have been written with NVIDIA hardware in mind. PyTorch, TensorRT and countless third-party projects have CUDA-based paths, so developers often find NVIDIA easier to support than alternative GPUs. On the GX10, that software runs alongside a much larger shared memory pool than you’d get from a normal consumer graphics card.
I wouldn’t pretend that CUDA fixes everything. The processor here is Arm-based, and that makes compatibility worth checking before spending money. A Python package may offer a ready-made wheel for x86 Linux but not for Arm64. A container might include CUDA libraries and still fail because one dependency was compiled for the wrong CPU architecture. This is exactly the sort of boring problem that can eat an entire afternoon. If your project uses custom extensions, verify them on the GB10 platform rather than assuming ‘CUDA supported’ means ‘everything supported.’
Still, for teams already using NVIDIA tools, the platform is easier to understand than a completely new GPU software stack. They can use familiar libraries and keep their development work close to the environment used on larger NVIDIA servers. That’s useful for prototypes and testing. A model might still need performance tuning when moved to another machine, but at least the basic tooling doesn’t start from zero.
The USB-C Power Detail Needs an Asterisk
Yes, ASUS powers the GX10 through USB-C. More specifically, its product specifications describe a USB-C input supporting USB Power Delivery 3.1 Extended Power Range at up to 180W. The bundled adapter is designed for the machine’s power needs. This is very different from plugging in a random 65W phone or laptop charger and expecting a local AI workstation to run at full speed.
USB-C describes the connector shape, not a universal guarantee that every charger can supply every device. Power Delivery negotiates voltage and current between the charger and the machine. High-wattage EPR operation also needs the right cable and compatible equipment. A USB-C charger can physically fit and still be the wrong charger. That sounds obvious, but the words ‘runs on USB-C’ can make the machine seem far more casual about power than it really is.
ASUS also provides separate USB-C connections for data and display output. Those aren’t all interchangeable with the power input. The published port list includes a USB-C connection with 180W PD input, three further 20Gbps USB-C connections, HDMI, 10Gb Ethernet and NVIDIA ConnectX-7 networking. Before placing it on a cramped desk or building it into a setup, I’d check the rear-panel arrangement and leave enough room for the cables. A tiny machine with several thick cables connected to the back can occupy more useful desk space than its dimensions suggest.
And no, this doesn’t make it a laptop replacement. The GX10 has no internal screen or battery. You still need its compatible power adapter and a way to access the desktop, whether that’s a monitor or a remote connection. What USB-C does is give a compact workstation a cleaner power arrangement than the large desktop power connectors people associate with high-end graphics cards.
The Catch: 273GB/s of Memory Bandwidth
Here’s the specification I would look at before the one-petaflop figure. ASUS lists memory bandwidth at 273GB/s. That’s useful bandwidth, but some high-end discrete GPUs have much faster dedicated VRAM. And language model generation can be limited by memory bandwidth because the hardware repeatedly needs to read large amounts of model data as it produces tokens.
There is a rough way to think about it. A 4-bit quantized 70B model has around 35GB of raw weights, before other memory overhead. Dividing 273GB/s by 35GB gives about 7.8 complete weight reads per second under ideal conditions. That’s not a reliable real-world tokens-per-second prediction. Kernels, caches, quantization details and batch size change the result, and actual operation isn’t one clean memory read per token. But it helps explain why a huge memory pool and a very fast model response are separate goals.
If a 14B or 32B model fits inside the memory of a faster desktop GPU, that GPU could deliver better inference speed than the GX10. The GX10 becomes more interesting as the model grows beyond what one mainstream GPU can hold. At that point, capacity may be more useful than peak speed because the alternative involves offloading, dividing work across GPUs, or using a server somewhere else.
This is why I wouldn’t buy the GX10 for small-model chatbot use alone. Plenty of less expensive systems handle that job. I’d want to see measured results for the exact model, quantization and context length I intended to run. For large models, I would also check performance after the conversation has grown to thousands of tokens, because long-context memory demands are easy to overlook in a quick benchmark.
There is no magic workaround for memory bandwidth. Software improvements can help, batching can raise total throughput, and some model architectures behave more efficiently than others. Yet the physical memory system still sets limits. ASUS has solved a capacity problem in a tiny case; that doesn’t mean it has solved every speed problem.
You Can Connect Two of Them, but That Is Not One Giant Memory Chip
The GX10 includes an NVIDIA ConnectX-7 interface that lets compatible systems exchange data over a high-speed link. ASUS advertises configurations connecting two units, giving a combined 256GB of memory when two 128GB models are used. The combined theoretical FP4 compute figure also rises. That could help developers test larger models or distribute multiple jobs across the machines.
But two computers still contain two separate physical memory pools. The software must divide the workload and handle communication between them. A program written for a single GPU doesn’t automatically become twice as fast when you connect a second box. Network overhead and how the model is split can make a big difference. For some jobs, two units are useful; for others, a simpler single machine may be easier to operate.
The idea is attractive if you start with one system and later discover that you need more capacity. You don’t necessarily have to throw away the original computer. On the other hand, buying two GX10 units is expensive enough that you should compare the total cost with a larger workstation or server. The network hardware may be neat, but the purchase still needs to make sense on a spreadsheet.
ASUS Has Already Added a Cheaper-Memory Option
There is a fresh development here. On October 5, 2026, ASUS published an announcement for a 64GB version of the Ascent GX10, expanding the lineup alongside the original 128GB model. ASUS is aiming the new configuration at people developing local AI agents and running open models in roughly the 30B to 35B parameter class. It uses the GB10 Grace Blackwell platform, so the difference most buyers will notice is the reduced memory capacity.
That complicates the buying decision. If your models fit within 64GB, paying extra for 128GB may not produce a better result in every workload. Memory capacity isn’t the same as memory bandwidth or raw compute speed. Developers who mostly test 14B and 32B models could find the smaller configuration enough, provided they leave room for context caches and the rest of their tools.
ASUS also promotes pairing two of the 64GB units through ConnectX-7, giving 128GB of combined memory and more aggregate compute. It’s worth comparing that idea with one 128GB unit, but do not confuse the two arrangements. A single 128GB machine is generally simpler for a model that needs one large memory pool. Two smaller machines may do better at tasks that divide cleanly between devices, with the extra work of coordinating them.
The 64GB news also tells us where ASUS thinks this market is going. The first wave of GB10 systems was priced for serious AI developers. A smaller-memory option broadens the range for people who want the software platform without necessarily loading the largest possible models. Prices and availability still vary by region, so the value of the new model depends on what local sellers charge rather than the memory number alone.
Would I Buy This Instead of a Conventional GPU Workstation?
For running smaller local models as fast as possible, probably not. A regular desktop with a capable NVIDIA graphics card may cost less and offer stronger performance on models that fit entirely in VRAM. It may also be easier to upgrade, repair and use for gaming or video work. Those are real advantages, not minor inconveniences.
For a developer who wants to test much larger quantized models, needs CUDA and has limited room, the GX10 becomes much more interesting. A conventional multi-GPU workstation adds complexity quickly. Several graphics cards bring larger power requirements, more heat, and decisions about how to divide model weights. The ASUS box packages a large shared memory pool and NVIDIA tooling into one small machine. That convenience has a price, but I can see why someone would pay for it.
Storage and hardware maintenance deserve a look too. ASUS offers SSD configurations that vary by region, but large model collections consume disk space at an alarming rate. A few different quantizations, datasets, containers and checkpoints can fill a drive faster than expected. ASUS’s support guidance also places limits on customer-performed hardware modification, so this isn’t the kind of PC I’d buy assuming I can replace every internal part later.
I would also separate local privacy from local security. Running a model on the GX10 means inference can stay on your own equipment instead of sending every prompt to an external API. That’s useful for sensitive development work. But you still have to manage access, storage encryption, network exposure and updates. A small box on a desk doesn’t automatically make confidential data safe.
What I like about this computer is that it makes a previously awkward setup easier to picture. One box, one memory pool, NVIDIA’s software and a high-wattage USB-C power connection. What I dislike is that the headline specifications can make it sound faster and simpler than it will be for every user. The person buying it for 7B chatbot responses and the person testing a quantized 120B model are shopping for two completely different things.
The 1.1-Liter Part Might Be the Biggest Achievement
The Ascent GX10 won’t replace an RTX 5090 gaming workstation, and 128GB of unified memory won’t make every AI model respond instantly. Its bandwidth, Arm software compatibility and price all deserve serious attention. The USB-C input is a real feature, although it needs the right high-power adapter rather than whatever cable and charger happen to be nearby.
Even with those limits, the physical package is hard to ignore. ASUS has placed a 20-core Arm CPU, Blackwell GPU, 128GB unified memory and CUDA-capable AI software inside a box that takes up barely more than one liter. A few years ago, getting access to that much model memory on an NVIDIA-based local system meant planning around a much bigger computer.
If ASUS and NVIDIA want to convince more developers, the next thing they should publish is simple: repeatable model benchmarks with token speeds, context lengths, memory use and wall-power measurements. I’d take that over another petaflop graphic any day. The tiny hardware is already real. What buyers need now is a clearer picture of what it can do after the model has loaded.