Skip to content
Small team, full backlog, zero orders dropped. Support replies are slower than we’d like. Read our status update → Zero orders dropped. Status → 📬 Check your spam folder — most of our replies land there. We do answer. Status update → 📬 Check your spam folder. Status →
1325 Rue Bergar, Laval, QC support@d-central.tech 1-855-753-9997 Mon-Fri: 8:30 - 18:00

Which Local AI Computer Should You Buy in Canada? (2026)

Which local AI computer should a Canadian buy? Three honest paths: an NVIDIA DGX Spark at $9,449 CAD (CUDA, Arm, 128 GB unified, 4 TB, DGX OS); an AMD Strix Halo 128 GB desktop at $7,349 CAD (x86, Windows or Linux, Ryzen AI Max+ 395); or a custom GPU inference rig quoted from current GPU street — the same workshop that used to build GPU mining rigs. Capacity and speed are different questions. D-Central sources and first-boot configures the first two in Laval, Quebec. We did not design the silicon.

This page is a buying map, not a leaderboard. Street prices below were checked 24 August 2026. GPU and LPDDR5X prices move; we confirm the chassis before we place a hardware order. Full credit to NVIDIA, AMD, and the open inference stack (Ollama, llama.cpp, vLLM, Open WebUI) that makes any of these boxes useful.

The three paths in one table

Path What it is D-Central list (CAD) Buy it when Do not buy it when
NVIDIA DGX Spark GB10 Grace Blackwell. 128 GB unified LPDDR5X, 4 TB NVMe, ~240 W, DGX OS, CUDA. $9,449 including first-boot Ollama + private UI You need CUDA, a 128 GB pool, NVIDIA’s stack, and a quiet desk circuit. You need Windows, or models that already fit in 24–48 GB GDDR and you want max tokens/s.
AMD Strix Halo 128 GB Ryzen AI Max+ 395, Radeon 8060S, 128 GB LPDDR5X, x86, Windows or Linux, ROCm/Vulkan. $7,349 including first-boot setup You want 128 GB unified memory on x86, and you can live without CUDA. Your workload is vLLM/TensorRT that assumes NVIDIA. We do not sell the US-only Micro Center “Ryzen AI Halo.”
Custom GPU inference rig Discrete GPU in an x86 tower. Owner-serviceable. Quoted from GPU street. Quoted (24 GB class is the usual start) You want a card you can swap, higher GDDR bandwidth, or a used 3090-class build. You want a 128 GB single pool in a 240 W brick, or you refuse to manage drivers.

Capacity vs speed

A 128 GB unified box fits 70B–200B-class open-weight models (quantized; KV cache still eats headroom). A discrete GPU streams smaller models faster because GDDR bandwidth is higher than LPDDR5X. The DGX Spark’s 273 GB/s is not an RTX 4090’s ~1,008 GB/s. If the weights fit in 24 GB, a custom GPU rig often wins tokens per second. If they do not, the discrete card falls over and the unified box is the one that still runs.

We will not invent tokens/s. Dated third-party reviews (ServeTheHome, StorageReview, The Register, Tom’s Hardware) and Artificial Analysis v4.1.1 (already on our comparison, dated 24 August 2026) are the numbers to cite. Prefill and decode are different jobs.

Canadian street vs our list

On 24 August 2026 a bare DGX Spark in Canada was roughly $7,650–$8,450 at ExtremePC, Caretek, Amazon.ca, and Canada Computers. NVIDIA US list was about US$4,499–$4,699. Our $9,449 is not the cheapest listing. It includes Quebec sourcing, first-boot setup, Bitcoin checkout, and a written model-fit note. If you only want the cheapest sealed box and will install DGX OS yourself, buy from a big-box retailer.

The AMD $7,349 is a sourcable 128 GB Max+ 395 system (Framework Desktop class or equivalent), not AMD’s Micro Center exclusive Halo SKU. See DGX Spark in Canada and Spark vs a custom GPU build.

How to choose in five minutes

  1. Name the model and the quant you actually need, not a leaderboard winner. Use GPU ↔ model fit and the open-weight shortlist.
  2. If it needs CUDA (vLLM, TensorRT-LLM, most NVIDIA NIM): Spark or a discrete NVIDIA GPU. Not Strix Halo.
  3. If it needs Windows + 128 GB + x86: Strix Halo.
  4. If you already think in PSU, PCIe, and exhaust: custom GPU rig. That is this shop’s GPU-mining muscle.
  5. If the data cannot leave Quebec: the box is only the start. On-prem install, logs, and a written boundary still matter — on-premises deployment. Residency is not sovereignty; see CLOUD Act.

What we will not sell you

We retired named vapour SKUs (Pleb AI Box, Workstation 24/48, Hashcenter AI Node 80+). They were not real inventory. The three products above are the ladder. We also will not sell a public GPU cloud, an SLA we cannot staff, or the claim that US/Canada goods tariffs are a tax on ChatGPT tokens. CUSMA Chapter 19 generally prohibits customs duties on digital products transmitted electronically.

Frequently asked questions

Can I just buy a used RTX 3090?

Yes — that is often the right 24 GB class card. We still have the 3090 buyer’s guide. If you want it in a quiet, wired, first-boot tower, that is the custom GPU inference rig, quoted from current street.

Is Strix Halo “as fast as Spark”?

No. Independent reviews in 2026 generally give Spark a large throughput lead on CUDA/vLLM jobs. Strix Halo’s job is x86 + 128 GB + Windows/Linux at a lower list price. Read the dedicated comparison when it publishes; until then use Spark vs custom GPU for the discrete path.

Do you photograph the unit in Laval?

Spark and Strix Halo are sourced manufacturer systems. We will not caption a press render as “built in our shop.” The custom GPU rig is assembled here; that page stays quote-first until we have a real tower photo.

Buy DGX Spark — $9,449 CAD · Buy Strix Halo 128 GB — $7,349 CAD · Quote a GPU rig

New in August 2026 ??? DGX Spark + Strix Halo authority: