DGX Spark vs Mac Studio M5 Ultra for Private LLMs (2026)

Every figure below comes from the vendor's own product page or announcement. Where a vendor publishes no number, the cell says so rather than borrowing one from a review.

Three entries need a caveat before they mislead anyone.

The DGX Spark's 1 PFLOP is FP4 with sparsity, labelled theoretical by NVIDIA. NVIDIA does not publish a dense figure on the product page, so do not halve it and print the result as a spec.

The 128 GB M5 Max is a 614 GB/s machine. Apple ties the 128 GB option to the 40-core GPU, which is the faster memory configuration. On the M5 Ultra, 1.2 TB/s applies to both the 30-core and 36-core chips, but 256 GB and 512 GB require the 36-core CPU and 80-core GPU. A 96 GB base unit cannot be upgraded to 256 GB later on the same chip, so order the capacity you need on day one.

Strix Halo's 256 GB/s is arithmetic: 8000 MT/s times 256 bits, divided by 8. AMD's press material also caps the memory available to graphics at 96 GB of the 128 GB, which matters more for model sizing than the bandwidth figure does.

LLM inference has two phases with different limits. Splitwise, a 2023 paper on phase-split serving, describes them as "a compute-intensive prompt computation, and a memory-intensive token generation" (arXiv 2311.18677). A later survey uses a roofline model to analyse why LLM inference is memory-bound (arXiv 2402.16363).

For decode the consequence is simple arithmetic. Generating one token for one user reads every active weight once, so the ceiling is memory bandwidth divided by the bytes of weights read per token. A 4-bit model stores about half a byte per parameter.

The table below applies that division to two generic model shapes: a 70B dense model at 4-bit (about 35 GB read per token) and a mixture-of-experts model with about 10B active parameters at 4-bit (about 5 GB per token). It is a theoretical ceiling from published bandwidth. These are not measurements: real single-stream decode lands below them because the KV cache also has to be read, kernels never hit peak bandwidth, and the software stack adds overhead.

Two readings follow. First, the ranking is fixed by the bandwidth column: whatever the software does, a Spark will not out-decode an M5 Ultra on a single stream of a model that fits both. Second, model architecture moves the numbers more than hardware choice does. A sparse MoE on a Spark has a higher ceiling than a dense 70B on an M5 Ultra, which is why sparse MoE models are the natural fit for all of these boxes.

Prefill, the prompt-processing phase, is compute bound, and it cannot be compared across these vendors from spec sheets. NVIDIA publishes only the sparse FP4 peak for GB10. Apple publishes no TFLOPS for M5 Max or M5 Ultra, only relative claims: up to 4x faster LLM prompt processing in LM Studio than the M3 Ultra and up to 9.8x faster than the M1 Ultra, from Apple's own July 2026 testing. AMD's announcement gives NPU TOPS but no GPU TFLOPS. What can be said from first principles is that Spark's Blackwell Tensor Cores with native FP4 are built for exactly that compute-bound phase. For the latency targets a shared deployment needs to hit, see what TTFT and tokens per second targets to set.

All three boxes start at or near 128 GB, and the arithmetic of weights times bytes per parameter says what fits. 128 GB holds a 70B dense model at 8-bit (about 70 GB) or roughly 200B parameters at 4-bit (about 100 GB) with room left for KV cache and the operating system. Past about 230B parameters at 4-bit you need the 256 GB tier. The 512 GB tier is for the largest open-weight MoE models at low-bit quantization. On Strix Halo, size against AMD's 96 GB graphics figure, not the full 128 GB.

The paths up from 128 GB are different in kind:

The Spark price has moved twice in 2026. In late February NVIDIA raised the Founders Edition MSRP from $3,999 to $4,699 across all regions, citing worldwide memory supply constraints, with no hardware change. On October 2, press coverage reported a second increase for the 128 GB model to about $7,000, the same day NVIDIA announced a 64 GB configuration sold only through OEM partners from $4,999, on sale from October 23. Any $3,999 or $4,699 quote predates those changes. The 512 GB Mac price comes from launch coverage of the October 28 store opening: the 36-core CPU and 80-core GPU chip at $6,799, plus $15,500 for 512 GB. Treat $22,299 as a reported list price at order opening. The 512 GB model is orderable; this comparison makes no claim about delivery dates.

Per gigabyte the two paths are close up to 256 GB, about $55 for Sparks against $57 for the base M5 Ultra, and at 512 GB the single Mac is cheaper: $22,299 against about $28,000 for four Sparks before the switch. The gigabytes are not equivalent either. Four linked Sparks are four 128 GB pools joined by 200 Gb/s links, which is 25 GB/s per port, about a tenth of each box's local memory bandwidth. A model split across them pays that cost on every crossing. One 512 GB M5 Ultra is one pool at 1.2 TB/s. Apple's own figure for a four-Mac cluster, up to 3x faster AI inference than one system over Thunderbolt 5 at 120 Gb/s, is the same lesson from the other vendor: linking scales sub-linearly. Link boxes for capacity, not for speed.

For a worked example of a very large MoE squeezed into a 256 GB Mac Studio, see our GLM-5.2 local runbook. If the real question is whether owning any of this beats paying per token, that math is in self-hosted LLM vs API break-even, and we will not redo it here.

This is the section that should decide the purchase for a regulated team.

A desk box in a regulated environment produces evidence: accuracy on your evaluation set at a given quantization, refusal behaviour, latency at a given context length. That evidence feeds model risk review and, eventually, a production sign-off. It is only worth what it predicts about the production system.

The second question is who owns the box. A machine holding real customer, patient or claims data is in scope for the same controls as any other system that touches it: patching, access, logging, encryption at rest, and an inventory entry.

NVIDIA documents DGX Spark as an enterprise endpoint managed with an appliance mindset. Its fleet lifecycle guide covers agentless management over SSH, provisioning with cloud-init, updates in maintenance windows rolled out in staged rings with pre-check and post-check evidence, and Canonical Landscape as the recommended management platform, with Ansible, Puppet and Tanium usable but supported by their own vendors. The manageability guide covers air-gapped installation from a local package repository and firmware mirror. Storage is self-encrypting NVMe up to 4 TB. DGX OS major updates arrive around February and August for the first two years after release, with security patches in between. NVIDIA is also explicit about what stays with you: identity and access control, secrets, change management, artifact retention and legal hold, and network segmentation. Once the stack is validated, running it fully offline follows the same pattern as our air-gapped vLLM deployment guide.

A Mac Studio is managed as an endpoint. Apple's deployment guide describes mobile device management that configures devices remotely, updates software and settings, monitors compliance with organisational policies, and can lock or wipe a machine. That fits a single developer's machine neatly, because it rides on the same MDM that already manages your laptops. It fits a shared server less well, and Mac clustering adds a manual step: MLX's distributed documentation says RDMA over Thunderbolt, needed by its low-latency backend for tensor parallelism, must be enabled from macOS recovery on each machine and cannot be turned on remotely, even with administrator rights.

A Strix Halo box is a commodity mini PC or workstation. Firmware, BIOS updates and manageability features depend on the OEM, so the controls you get depend on the model you buy. Check the vendor's firmware update channel and remote management support before the data owner asks.

Write down three answers before the box arrives, whichever you buy: who patches it and on what schedule, where its access and inference logs go, and whether more than one person logs into it. The third answer decides between endpoint management and server management.

Usage pattern changes the hardware answer as much as carry-over does.

NVIDIA's RTX Spark is a different product with a similar name. It is a Windows 11 platform for slim laptops and compact desktops, launched at events on October 16, 2026, with laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI. NVIDIA lists up to 128 GB of unified memory for both laptops and desktops, with NVIDIA claiming up to 1 petaflop of FP4 AI performance; NVIDIA publishes no memory bandwidth for it. It is a Windows developer endpoint that runs models locally, not a DGX OS box and not a shared server. If your engineers need a laptop that runs a mid-size model offline, it belongs on that shortlist. It does not replace a DGX Spark as a validation rig.

Our LLM and models guide collects the rest of our work on sizing, quantization and serving open-weight models on your own hardware.

This week, write two things on one page: the serving stack your production deployment will run (engine, GPU vendor, quantization format) and the largest model you must evaluate before sign-off. Match the box to the first line, then size it to the second. If the first line says NVIDIA and vLLM, the Spark is the default and the M5 Ultra has to justify itself on capacity alone.