There are three practical ways to run AI models on your own desk in 2026. Each one gives up something different. An Apple Mac Studio is the turnkey option: large unified memory, low power and very little to set up, but nothing can be upgraded and there's no CUDA. An NVIDIA RTX Spark PC, such as the upcoming ASUS ProArt GR1X, brings unified memory and CUDA to Windows, but its memory tops out at 128 GB and its software is still maturing. A custom PC with a discrete GPU such as the RTX 5090 is the fastest per token and the only one you can upgrade, but its 32 GB of graphics memory limits which models fit. Pick by the largest model you need to run, how fast it must respond, and whether you want to upgrade later.
How do the three approaches compare?
| Approach | Mac Studio (M5 Max / M5 Ultra) | RTX Spark PC (e.g. ASUS ProArt GR1X) | Custom PC with RTX 5090 |
|---|---|---|---|
| Memory the GPU can use | Up to 128 GB (Max) or 512 GB (Ultra) unified; macOS lets the GPU use about 75% by default | 64 or 128 GB unified; the GPU sees about 55 GB on 64 GB models, and about 64 GB by default (adjustable to about 96 GB) on 128 GB models (laptop data) | 32 GB on the card; anything larger spills to much slower system RAM |
| Memory bandwidth | 460 or 614 GB/s (Max); 1.2 TB/s (Ultra) | "Up to 300 GB/s" (ASUS) | 1,792 GB/s on the card |
| Operating system and AI software | macOS; Metal, MLX, llama.cpp; no CUDA | Windows 11 on Arm; CUDA as a developer preview | Windows or Linux; full CUDA stack |
| Upgrades after purchase | None; memory and storage are fixed | Memory is soldered | GPU, RAM, storage and power supply can all be replaced |
| Power | Maximum 200 W (Max) or 385 W (Ultra) | 140 W sustained TDP; 240 W adapter | Card rated 575 W; whole PC peaked at 567–577 W in MacStories' tests |
| Price guide | Apple Singapore's list price from S$3,499 (Max) and S$7,999 (Ultra) | About S$6,200–7,200 for 64 GB/1 TB (SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price) | Card US$1,999 MSRP; US street prices about US$6,500–8,500 (Oct–Nov 2026), plus the rest of the PC |
Source: Apple Singapore Mac Studio specifications and newsroom (accessed 8 Nov 2026); ASUS ProArt GR1X product page and Notebookcheck (Oct 2026); RTX Spark GPU memory from laptop reports (Oct 2026); NVIDIA RTX 5090 specifications; MacStories (Sep 2026); Tom's Hardware price tracker (26 Oct 2026) and eBay sold listings (Nov 2026) for US street prices.
What does the turnkey Mac Studio approach give you?
The Mac Studio's main advantage is capacity with very little effort. An M5 Ultra can be configured with 256 GB or 512 GB of unified memory, and the GPU can use most of it. Setup is short: install LM Studio, Ollama or Apple's MLX tools, download a model and start. It idles at 7–9 W according to Apple, it's quiet, and there are no drivers, power supplies or cooling to choose.
The trade-offs are real, though:
- You buy the memory once. Nothing inside can be upgraded, so the configuration has to cover the life of the machine.
- Prompt processing is slower than on NVIDIA discrete GPUs. MacStories ran Qwen3.8-27B with a 6,091-token prompt. The M5 Ultra processed 1,701 tokens per second against the RTX 5090's 3,031, and took 4.0 seconds to the first token against 2.0.
- There's no CUDA. CUDA-only tools such as TensorRT-LLM won't run. Our post on CUDA versus Apple's MLX covers what that means.
Where does an RTX Spark PC fit?
RTX Spark is NVIDIA's chip for Windows PCs, built around a unified memory pool shared by the CPU and GPU. It's the middle option: more AI memory than a gaming graphics card, plus CUDA on Windows. The ASUS ProArt GR1X mini PC uses it. So do laptops such as the ASUS ProArt P16, MSI Prestige N16 and Lenovo Yoga Pro 9n.
The constraints are bandwidth and software. ASUS quotes "up to 300 GB/s", far below the M5 Ultra's 1.2 TB/s or an RTX 5090's 1,792 GB/s. Token generation for large models is limited mainly by bandwidth, so expect it to be closer to an NVIDIA GB10 system than to a Mac Studio Ultra. In HardwareZone's review (6 Nov 2026), an M5 Max MacBook Pro generated tokens about twice as fast as the ProArt P16 on a 9B model in llama.cpp, while the P16 was faster at prompt processing. No independent benchmarks of the GR1X desktop have been published yet. On software, CUDA on Windows on Arm is a developer preview, vLLM has no native Windows support, and some tools need Arm64 builds (as of November 2026). Our RTX Spark versus GB10 guide compares the Windows and Linux routes, and our post on macOS versus Windows on Arm compares it directly with the Mac.
The real choice isn't Apple versus NVIDIA. It's how big your largest model is, how quickly it has to answer, and whether you'll want to upgrade in two years.
Why would anyone still build a custom PC?
A custom PC wins on two things the other two approaches can't match: speed when the model fits and the ability to upgrade.
- Speed. With 1,792 GB/s of memory bandwidth and NVIDIA's tensor cores, an RTX 5090 is the fastest of the three on models that fit in 32 GB, both for prompt processing and generation. In MacStories' Qwen3.8-27B test it generated 59 tokens per second against the M5 Ultra's 48.
- Upgrades. You can swap the GPU for a larger card later, add a second one, or move to an RTX PRO workstation card with more memory. Neither the Mac nor an RTX Spark PC allows that.
- Full CUDA. Training, fine-tuning and vLLM serving work as they do on NVIDIA servers.
The weakness is capacity. When a model doesn't fit in 32 GB, part of it runs from system memory. MacStories measured the RTX 5090 PC falling to between 1.5 and 4.6 tokens per second in those cases. It also draws several times the power of a Mac Studio, and RTX 5090 cards are hard to find in late 2026. Our Mac Studio versus RTX 5090 PC comparison and our hidden costs of building versus buying go into both.
Which approach suits which workload?
- Very large models (above about 100 GB in memory): a Mac Studio M5 Ultra with 256 GB or 512 GB is the only one of the three that runs them on its GPU.
- Models of 70B to 120B parameters: a Mac Studio with 128 GB or more, or an RTX Spark PC with 128 GB. The Mac has much more bandwidth, while RTX Spark offers CUDA on Windows.
- Models under about 30 GB and long prompts: a custom PC with an RTX 5090 is the fastest, because prompt processing is compute-heavy and favours NVIDIA's tensor cores.
- Image and video generation: like-for-like Mac versus RTX 5090 timings are scarce. In Blender Open Data, a GPU rendering benchmark, the M5 Ultra scored 15,350 against about 16,000 for the RTX 5090, but tools such as ComfyUI are built for CUDA first.
- CUDA development, training and fine-tuning: a custom PC, or a Linux GB10 system such as the DGX Spark, keeps you closest to the servers you'll deploy on.
- A quiet, low-power, low-maintenance desk machine: a Mac Studio or an RTX Spark mini PC.
Our guide to the best PC to run an LLM locally in Singapore has more configurations, and our post on unified memory versus GPU power explains how to read the specifications.
Buying in Singapore
SourceIT supplies all three approaches. Mac Studio models such as the M5 Max with 128 GB and the M5 Ultra with 256 GB are price on request; for reference, the 256 GB/2 TB M5 Ultra is about S$16,699 configured (calculated from Apple Singapore's upgrade pricing — confirm on apple.com/sg). The ASUS ProArt GR1X is price on request; our estimate for the 128 GB/2 TB model is S$9,000–10,800 (SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price). RTX Spark laptops are available now, including the ASUS ProArt P16 128GB at S$10,773.20 (listed 8 Nov 2026) and the MSI RTX Spark range. For a custom build, the AMD Ryzen 9 9950X3D is S$1,108.02 and the RTX PRO 4500 Blackwell is S$8,880; RTX 5090 cards aren't in stock, so ask us about availability. Prices change; 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
Which is fastest for local AI: Mac Studio, RTX Spark or an RTX 5090 PC? For models that fit in 32 GB, the RTX 5090 PC. For larger models, the Mac Studio M5 Ultra, because it has far more memory and bandwidth than RTX Spark and the model stays on the GPU.
Can I upgrade the memory in a Mac Studio or an RTX Spark PC later? No. Both use soldered memory, so choose the capacity you'll need for the life of the machine. Only a custom PC lets you add memory or change the GPU.
Does RTX Spark run CUDA? Yes, but on Windows on Arm CUDA is still a developer preview, and tools such as vLLM and TensorRT-LLM aren't supported natively (as of November 2026).
Is a custom PC cheaper than a Mac Studio for AI? It depends on the GPU. RTX 5090 cards sell well above their US$1,999 list price in late 2026, and the rest of the PC adds to that. A custom PC gives more speed per dollar on small models; the Mac gives more memory per dollar.
Which option uses the least electricity? The Mac Studio and RTX Spark PCs. Apple rates the M5 Ultra at a maximum of 385 W and 9 W at idle, and ASUS quotes a 140 W sustained TDP for the GR1X. An RTX 5090 card alone is rated at 575 W.
