It depends on the size of your models. When a model fits in the RTX 5090's 32 GB, a custom 5090 PC is the faster AI workstation. In MacStories' side-by-side test it processed prompts about 1.8 times faster than a Mac Studio M5 Ultra and generated text about 23–28% faster at every context length. When the model doesn't fit, the Mac wins easily: the 5090 fell to 1.5–4.6 tokens per second once it spilled into system RAM, while the Mac Studio holds 96 to 512 GB. The PC is also upgradeable and runs CUDA. The Mac is quieter, draws about half the power and is a single box you buy once. Right now the RTX 5090 is also hard to find.
How do they compare in the same test?
| Measure | Mac Studio M5 Ultra (256 GB) | Custom RTX 5090 PC (Ryzen 9 9950X3D) |
|---|---|---|
| Prompt processing, Qwen3.8-27B, 6,091-token prompt | 1,701 tok/s | 3,031 tok/s |
| Time to first token, same prompt | 4.0 s | 2.0 s |
| Time to first token, 63K-token prompt | 42.8 s | 26.4 s |
| Generation at 64K / 128K / 256K context | 38.9 / 32.4 / 24.3 tok/s | 49.6 / 40.4 / 30.0 tok/s (256K needed an 8-bit KV cache) |
| Models larger than the GPU's memory | Run normally, up to the Mac's RAM | Fell to 1.5–4.6 tok/s when spilling into system RAM |
| Peak power measured | About 300 W in LLM inference (Apple); 385 W max | 567–577 W |
| Blender Open Data (GPU rendering) | About 15,350 | About 16,000 |
Source: MacStories M5 Ultra review (~21 Sep 2026; Mac running MLX with oQ4e quantisation, PC running LM Studio 0.4.24 with GGUF Q4_K_M and CUDA 12; the quantisation formats differ); Apple Support power figures and Apple ML Research (15 Oct 2026); Blender Open Data medians via Notebookcheck (22 Sep 2026). SourceIT has not tested either system.
Why is the RTX 5090 faster on models that fit?
The RTX 5090 has 1,792 GB/s of memory bandwidth, about 1.46 times the M5 Ultra's 1.2 TB/s, and far more tensor compute. Token generation scales with bandwidth, and in MacStories' test the 5090's lead in generation (about 23–28%) sat a little under that bandwidth ratio. Prompt processing scales with compute, and there the gap was wider, about 1.8 times. Apple's M5 Neural Accelerators narrowed the gap a lot compared with older Macs, but haven't closed it.
For work built around 8B to 32B models, such as coding assistants, chat with documents, or image generation in ComfyUI, a 5090 PC gives quicker answers, especially with long prompts.
Why does the Mac win on big models?
Memory capacity. The 5090 has 32 GB. A 70B model at 4-bit needs about 42 GB of weights before any context, and gpt-oss-120b needs around 60–65 GB, so neither fits on the card. Spilling into system RAM over PCIe gave MacStories just 1.5–4.6 tokens per second, too slow to work with. The M5 Ultra's 96 GB base configuration fits both, and 256 or 512 GB fits mixture-of-experts models of several hundred billion parameters. See our post on why 32 GB frustrates AI developers for the details.
Long context is part of this too. The Mac ran Qwen3.8-27B at 256K tokens with room to spare; the 5090 needed an 8-bit KV cache to fit it.
A 5090 PC is a faster engine with a small fuel tank. The Mac Studio is a slower engine with a very large one.
What about cost and availability?
The RTX 5090's US launch price is still US$1,999, but street prices in late October and early November 2026 ran from about US$6,500 to US$8,500 as memory shortages hit graphics cards. SourceIT has no RTX 5090 cards in stock at the time of writing. A complete 5090 PC also needs a strong CPU (the Ryzen 9 9950X3D used in MacStories' test is S$1,108.02 with us), a 1,000 W-class power supply that NVIDIA recommends, a large case, cooling and 64 GB or more of DDR5, which has also risen sharply in price this year.
The Mac Studio M5 Ultra starts at S$7,999 at Apple Singapore's list price with 96 GB and 1 TB. The 256 GB configuration used in MacStories' test comes to about S$16,699 with 2 TB of storage, calculated from Apple Singapore's upgrade pricing.
Running costs differ too. At the measured peaks and Singapore's current tariff of 31.16 cents per kWh, eight hours a day of heavy use would cost about S$519 a year on the 5090 PC against about S$273 on the Mac (our calculation; real average draw is lower for both). That's about S$250 a year in the Mac's favour.
Which is easier to live with?
Upgrades: the PC wins. You can add RAM, a second GPU or a newer card later. The Mac's memory and GPU are fixed at purchase, so buy the memory you'll need for the next few years.
Software: the PC runs CUDA, so vLLM, TensorRT-LLM, Docker GPU containers and nearly every AI tool work without changes. The Mac runs MLX, llama.cpp, Ollama and LM Studio well, but Docker containers can't use its GPU and some PyTorch features are missing. Our post on CUDA versus MLX covers this.
Noise and heat: the Mac wins. Reviewers found it close to silent, while a 5090 PC under load is a 575 W graphics card plus CPU, venting heat into the room. See our look at the RTX 5090's 575 W power draw.
Gaming and Windows apps: the PC wins outright.
Which should you choose?
Choose a custom RTX 5090 PC if your models fit in 32 GB, you want the fastest answers on long prompts, you rely on CUDA tools, or you also game or render on Windows. Be ready to wait for a card.
Choose a Mac Studio if you want to run 70B models and larger, long contexts, or several models at once; you value silence and low power; and your tools run on macOS. For more on this choice, see whether unified memory replaces expensive GPUs and the hidden costs of building versus buying.
Buying in Singapore
The Mac Studio M5 Ultra 96 GB / 1 TB and the M5 Ultra 256 GB / 2 TB are price on request; ask us for a quotation. RTX 5090 cards are out of stock at the time of writing; ask us about availability, or about a complete system quote. For a compact system with RTX 5090-class graphics now, the ASUS ROG NUC 16 Edition 20 (S$11,888.40, listed 8 Nov 2026) uses the RTX 5090 Laptop GPU with 24 GB, which is slower than the desktop card. 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 a Mac Studio faster than an RTX 5090 for AI? Not when the model fits in the 5090's 32 GB: the 5090 was about 1.8 times faster at prompt processing and 23–28% faster at generation in MacStories' test. For larger models, the Mac is far faster because the 5090 has to spill into system RAM.
Can an RTX 5090 run a 70B model? Not fully in its 32 GB. A 70B model at 4-bit needs about 42 GB for the weights, so part must sit in system RAM, which is very slow.
Which uses less power? The Mac Studio. Apple measured about 300 W peak during LLM inference, against 567–577 W for the 5090 PC in MacStories' test.
Can I upgrade a Mac Studio later? No. Memory and GPU are fixed at purchase. A PC lets you add RAM or change the graphics card.
Can I buy an RTX 5090 in Singapore now? Supply is tight and we have no cards in stock at the time of writing. Contact us for availability or a complete-system quote.
