For one person chatting with a large model, the new Mac Studio with M5 Ultra is likely to win on speed, because its memory runs at 1.2 TB/s against the GB10's 273 GB/s, and Tom's Hardware's September 2026 review found it generating text nearly four times faster than a DGX Spark on a 27B model. NVIDIA's GB10 wins on the things that matter to AI developers and teams: the CUDA software stack used in production, NVIDIA's serving and fine-tuning tools, 200 Gb/s clustering, and a lower entry price for 128 GB. Note that there is no "M4 Ultra": Apple went from the M3 Ultra to the M5 Ultra, which shipped on 22 September 2026. At SourceIT the ASUS Ascent GX10 starts at S$8,100, with the NVIDIA DGX Spark at S$8,800 (listed 28 Oct 2026). SourceIT does not sell the Mac Studio.
GB10 vs Mac Studio: the specifications
| Specification | NVIDIA GB10 (DGX Spark, ASUS GX10) | Mac Studio M5 Max | Mac Studio M5 Ultra |
|---|---|---|---|
| Unified memory | 128 GB | 36, 48, 64 or 128 GB | 96, 256 or 512 GB (no 128 GB option) |
| Memory bandwidth | 273 GB/s | 460 or 614 GB/s | 1.2 TB/s |
| GPU | Blackwell, up to 1 PFLOP FP4 (sparse) | 32 or 40-core Apple GPU | 64 or 80-core Apple GPU |
| AI software | CUDA, PyTorch, TensorRT, vLLM, NIM on DGX OS (Linux) | MLX, llama.cpp, Ollama, PyTorch (Metal) on macOS | Same as M5 Max |
| Clustering | ConnectX-7, 2x 200 Gb/s; 2 to 4 units supported, up to 8 manually | Thunderbolt 5 (120 Gb/s) | Thunderbolt 5; Apple claims a four-Mac RDMA cluster gives up to 3x faster inference |
| Power | 240 W adapter | 480 W maximum continuous (Mac Studio) | 480 W maximum continuous |
| Singapore price | From S$8,100 (GX10 1 TB, SourceIT) | From S$3,499 (36 GB); S$4,399 (64 GB) | From S$7,999 (96 GB) |
Mac specifications from apple.com/sg, checked 28 Oct 2026. Mac Singapore prices as reported by HardwareZone at launch (September 2026); confirm in Apple's configurator, as higher-memory configurations cost substantially more. GB10 specifications from NVIDIA.
Why is the Mac faster at generating text?
Generating each token means reading the model's active weights from memory, so speed rises with memory bandwidth. The M5 Ultra's 1.2 TB/s is more than four times the GB10's 273 GB/s, and Tom's Hardware's review of an 80-core, 256 GB M5 Ultra found token generation "nearly 4x" a DGX Spark on Qwen3.8-27B at 4-bit. Even Apple's laptop chip is quick: Techtesters measured a MacBook Pro M4 Max at 86.2 tokens per second on GPT-OSS-120B in October 2026, against 58.7 for DGX Spark in llama.cpp's February 2026 benchmark, with different tools and settings.
The GB10's strength is compute. Before generating, a model must process the prompt, and that stage depends on raw compute. In EXO Labs' October 2025 test of the previous M3 Ultra against a DGX Spark, the Spark processed an 8,000-token prompt nearly four times faster (1.47 s against 5.57 s), while the Mac generated faster. Tom's Hardware found the M5 Ultra faster at prompt processing too, so Apple has narrowed that gap with its new chip.
The Mac is the faster chat machine. The GB10 is the closer match to where AI actually gets deployed.
Why do developers still choose GB10?
- CUDA. Most AI research code, training scripts and serving engines target NVIDIA GPUs. TensorRT-LLM, NVIDIA NIM and many vLLM optimisations are CUDA-only.
- Production parity. DGX OS is the same software family as NVIDIA's data-centre systems, so what works on a GB10 deploys upward.
- Fine-tuning. NVIDIA publishes LoRA and QLoRA playbooks and rates DGX Spark for fine-tuning up to 70B. See Can a GB10 Fine-Tune Your Own AI Model?
- Serving several users. Batching engines built for NVIDIA GPUs make a GB10 a capable small-team server. See Can One GB10 Support 20 Employees?
- Clustering. 200 Gb/s RoCE links with NVIDIA's NCCL, supported from two to four units and up to eight manually.
- Price for 128 GB. GB10 systems with 128 GB start from S$6,895 at SourceIT, and the GX10 1 TB is S$8,100. The M5 Ultra skips from 96 GB to 256 GB, and the 128 GB M5 Max runs at 614 GB/s rather than 1.2 TB/s.
When is the Mac Studio the better buy?
- You work on macOS and want one machine for daily work and local AI.
- Your main use is single-user chat, coding assistance or summarising with models that fit in memory, where generation speed matters most.
- You need more than 128 GB in one box without clustering: the M5 Ultra offers 256 GB and, once available, 512 GB.
- You use MLX-based tools or creative apps on macOS.
Our guide to the best PC for running LLMs locally sets both options against the rest of the market.
Can you use both?
Yes. NVIDIA PAIR can serve inference from an NVIDIA machine to macOS computers on the same network, so a Mac user can call a GB10 for CUDA-only models or shared team workloads. EXO Labs also showed in 2025 that combining a DGX Spark for prompt processing with a Mac for generation could speed up the whole job, by 2.8 times in its test, though that remains a specialist setup.
Buying in Singapore
SourceIT supplies GB10 systems including the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800) and the HP ZGX Nano G1n (S$11,499), with local warranty and a GST invoice. We do not sell Apple products; buy the Mac Studio from Apple or an Apple reseller. Browse the NVIDIA DGX Spark collection, or compare GB10 with Windows options in DGX Spark vs RTX Spark. For a 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 there a Mac Studio M4 Ultra? No. Apple's Mac Studio moved from the M3 Ultra to the M5 Ultra, which was announced on 25 August 2026 and shipped on 22 September 2026.
Is the Mac Studio M5 Ultra faster than DGX Spark for local AI? For generating text, yes. Tom's Hardware measured nearly 4x a DGX Spark's token generation on Qwen3.8-27B. The Mac's memory bandwidth is 1.2 TB/s against 273 GB/s.
Why buy a GB10 instead of a Mac? For the CUDA software stack used in production, NVIDIA's serving and fine-tuning tools, 200 Gb/s clustering, and 128 GB of GPU-addressable memory from S$6,895 at SourceIT.
How much memory does the Mac Studio M5 Ultra offer? 96 GB, 256 GB or 512 GB, according to Apple's specifications. The M5 Max offers up to 128 GB.
Does SourceIT sell the Mac Studio? No. SourceIT supplies NVIDIA GB10 systems from ASUS, NVIDIA, HP, Dell and other brands.
