On local inference for one person, NVIDIA's GB10 is now behind the Mac Studio M5 Ultra. The Mac has about 4.5 times GB10's memory bandwidth (1.2 TB/s against 273 GB/s), up to four times the memory in one box, and Tom's Hardware measured it ahead of a DGX Spark on both prompt processing and token generation. But "losing the local AI war" overstates it. GB10 runs the full Linux CUDA stack, has 200 Gb/s ConnectX-7 networking, has better fine-tuning tools and gives a direct path to NVIDIA's data-centre platforms. The two machines are winning different buyers.
How do the M5 Ultra and GB10 compare?
| Specification | Mac Studio M5 Ultra | NVIDIA GB10 (DGX Spark, ASUS GX10) |
|---|---|---|
| Memory | 96 GB, 256 GB or 512 GB unified | 128 GB unified (64 GB versions also sold) |
| Memory bandwidth | About 1.2 TB/s | 273 GB/s |
| GPU memory available by default | About 75% of RAM (raisable with an unsupported setting) | Most of the 128 GB, under Linux |
| Operating system | macOS 27 | NVIDIA DGX OS (Ubuntu Linux) |
| AI software | MLX, Metal, llama.cpp, Ollama (MLX preview), vllm-metal | CUDA, TensorRT-LLM, vLLM, SGLang, NVIDIA NIM |
| Networking for clusters | Thunderbolt 5 with RDMA (macOS 26.2+), direct cables, up to 4 in a full mesh | ConnectX-7, 200 Gb/s per port |
| Power | 9 W idle, 385 W max (Apple) | 240 W external supply; about 160 W at the wall under GPU load (Tom's Hardware) |
| gpt-oss-120b, prompt / generation | About 2,829 / 109 tok/s (oMLX community run, 4K prompt) | 2,444 / 58.7 tok/s (llama.cpp, 2K prompt) |
Source: Apple Mac Studio specifications and Apple Support power page (accessed 8 Nov 2026); NVIDIA DGX Spark documentation; Tom's Hardware DGX Spark testing (27 Jan 2026) and M5 Ultra review (21 Sep 2026); oMLX community benchmark (24 Sep 2026, M5 Ultra 80-core, quantisation not stated, unverified); NVIDIA / llama.cpp DGX Spark figures. The gpt-oss rows use different engines and prompt lengths, so treat them as indicative, not a head-to-head.
Why is the Mac faster at local inference now?
For text generation, memory bandwidth sets the speed limit. Each new token requires reading the model's active weights from memory, so a machine with 4.5 times the bandwidth can, in principle, generate tokens up to 4.5 times faster on the same model. Tom's Hardware reported the M5 Ultra producing "almost four times" the DGX Spark's generation rate on Qwen 3.8-27B, close to that ratio.
The bigger surprise is prompt processing, which depends on compute rather than bandwidth. The M3 Ultra lost to the DGX Spark here: in EXO Labs' 2025 test the Spark processed an 8K prompt in 1.47 seconds against the M3 Ultra's 5.57. The M5 generation added matrix-multiply hardware to every GPU core, and Tom's Hardware found the M5 Ultra ahead of the Spark on prompt processing too. Nobody has yet published both machines running the same engine on the same model, so the exact margin is still open.
What does GB10 still do better?
It runs NVIDIA's full software stack. GB10 machines ship with NVIDIA's Linux DGX OS. vLLM, SGLang, TensorRT-LLM, NIM containers, Docker with GPU access and most research code work as they do on NVIDIA's data-centre GPUs. On the Mac, Docker containers can't use the GPU, PyTorch's Metal backend still lacks some operations, and the Mac's official vLLM is a separate plugin, vllm-metal, released in September 2026.
It's built to move work to the data centre. NVIDIA designed DGX Spark for developing locally and deploying on NVIDIA servers. The same CUDA code, containers and NVFP4 model formats run on both. A team that will put its model into production on NVIDIA GPUs avoids a porting step.
Fine-tuning tools are more mature. NVIDIA publishes DGX Spark fine-tuning figures, for example about 760 tokens per second for QLoRA on Llama 3.3 70B in its own tests. MLX supports LoRA and full fine-tuning, but tools such as Unsloth, TRL and DeepSpeed are built around CUDA. Our GB10 fine-tuning guide covers what fits.
Networking. GB10's ConnectX-7 runs at 200 Gb/s per port, and NVIDIA supports linking DGX Sparks for bigger models. The Mac's RDMA over Thunderbolt 5 works without a switch for up to four machines in a full mesh, and has to be switched on from Recovery mode on each Mac. Jeff Geerling measured real-world Thunderbolt 5 throughput at roughly 50–60 Gb/s.
Entry price. A 64 GB ASUS Ascent GX10 costs S$6,670, below the cheapest M5 Ultra. A 128 GB GX10 or DGX Spark costs about S$9,300–10,350, against S$7,999 at Apple Singapore's list price for a 96 GB M5 Ultra.
If "local AI" means one person running the biggest model possible as fast as possible, the Mac is ahead. If it means building something that will later run on NVIDIA servers, GB10 is still the more natural tool.
Is NVIDIA really losing?
In the specific contest of desk-side inference speed and capacity per box, GB10 has fallen behind the M5 Ultra. NVIDIA has partly answered with RTX Spark Windows PCs and the much larger DGX Station GB300, but neither competes directly on the Mac's ground of a quiet, efficient single box with 256–512 GB.
The "war" framing also ignores what each side is selling. Apple sells a workstation that happens to be very good at local AI. NVIDIA sells an on-ramp to CUDA and its data-centre platforms. Developers often own both. Our earlier comparison of GB10 and the Mac Studio covers model sizes. For 70B models specifically, see Mac Studio vs GB10 on a 70B model, and for the bigger strategic picture, whether NVIDIA should be worried.
Which should you choose?
Choose the Mac Studio M5 Ultra if you want the largest models on one machine, fast generation for one or two users, low idle power and a quiet desktop, and your tools run on MLX, llama.cpp or Ollama. Consider 256 GB or more if you plan to run 200B-plus mixture-of-experts models.
Choose GB10 (DGX Spark or ASUS Ascent GX10) if you need CUDA, Linux and Docker GPU support, plan to fine-tune models, or will deploy on NVIDIA servers. It is also the cheaper way into 64–128 GB.
Buying in Singapore
SourceIT supplies both. The Mac Studio M5 Ultra 96 GB / 1 TB (S$7,999 at Apple Singapore's list price) and the 256 GB configuration are price on request; ask us for a quotation. On the NVIDIA side, the NVIDIA DGX Spark 4TB is S$9,300, the ASUS Ascent GX10 64 GB / 1 TB is S$6,670 and the GX10 128 GB / 1 TB is S$10,350 (listed 8 Nov 2026). See all GB10 systems in our NVIDIA DGX Spark collection. 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 Mac Studio M5 Ultra faster than a DGX Spark? In Tom's Hardware's testing, yes, on both prompt processing and token generation. No test has yet run both on the same engine and model, so exact margins vary by setup.
Why is the M5 Ultra faster at generating text? Mainly memory bandwidth: about 1.2 TB/s against GB10's 273 GB/s. Text generation speed rises roughly with bandwidth.
Can a Mac Studio run the same software as a DGX Spark? Partly. Ollama, llama.cpp and LM Studio run on both. CUDA-only tools such as TensorRT-LLM and NIM don't run on a Mac, and Docker containers can't use the Mac's GPU.
Which is better for fine-tuning? GB10, for most teams. It runs CUDA-based tools like Unsloth and TRL, and NVIDIA publishes fine-tuning figures for it. MLX supports LoRA on the Mac, but the tooling is narrower.
Which is cheaper? GB10 at the entry level: a 64 GB GX10 is S$6,670. At 128 GB, the DGX Spark (S$9,300) and the 96 GB M5 Ultra (S$7,999 at Apple Singapore's list price) are close, while the Mac offers more memory at higher prices.
