Home/The machines/Custom builds
For the jobs outside
the standard box.
Strix Halo and DGX Spark are already main choices in our standard range. Custom work starts when the brief calls for two appliances, a full RTX workstation, CUDA-only software or a setup that cannot be priced sensibly from a four-line tier list.
Paired appliances · RTX and CUDA · quoted to the actual job
First, check the normal range.
Most offices do not need a custom workstation. The main machines already cover macOS, Windows and Linux, from quiet desk appliances through to Strix Halo and DGX Spark.
- Mac mini or Mac Studio
- Our usual recommendation for speed, quiet operation and straightforward setup
- AMD Strix Halo
- Windows or Linux, with up to 128 GB unified memory on supported systems
- NVIDIA DGX Spark
- DGX OS, CUDA and NVIDIA's local development stack
Sized to the work, not to the ceiling.
These are the systems we recommend most often, and for a lot of offices they leave room to spare. If what you actually need is a 16 GB Mac running a small model and transcription, we will quote that instead and say so plainly. The right machine is the one that matches the work — not the largest one we could sell you.
Those are standard products, mixed into the same tiers and starting prices. Compare the current machines and their trade-offs →
Two appliances, for two different reasons.
Sometimes the point is one larger NVIDIA-supported workload. Sometimes the point is keeping two busy workloads from competing for one box.
A supported DGX Spark pair
NVIDIA equips DGX Spark with ConnectX-7 networking and documents dual-system configurations for selected models up to 405 billion parameters. That is a maker claim for supported workloads, not a promise that every model or application will divide neatly across two boxes.
We check the exact model, framework and network path before quoting the pair. If the software is not designed for it, more hardware will not fix the software.
Separate appliances by workload
A practice may be better served by one appliance for document-heavy staff and another for coding, research or a separate team. Each system keeps its own model load and maintenance window.
That can be done with Mac, Strix Halo or Spark hardware. We size each appliance for its own work instead of pretending two unrelated boxes become one large machine.
The quote records which users, documents and external connections belong on each appliance.
RTX workstations: when CUDA is the requirement.
A discrete NVIDIA card can be the right answer for CUDA software, local fine-tuning, computer vision or high-throughput work. Its VRAM is also a hard ceiling: system memory does not turn a 24 GB card into a 48 GB card.
RTX PRO 6000 Blackwell
NVIDIA's Workstation Edition carries 96 GB of GDDR7 memory with ECC and is rated at 600 W. It is a serious card for CUDA work and larger GPU-resident jobs, but the rest of the workstation has to be designed around its power, cooling, noise and physical size.
- GPU memory
- 96 GB GDDR7 with ECC
- Power
- 600 W for the Workstation Edition
- Bus
- PCIe Gen 5 x16
- Best reason to choose it
- CUDA work that genuinely needs this memory class
GeForce RTX 4090
The RTX 4090 remains a practical way to run CUDA on hardware with a known second-hand market. NVIDIA specifies 24 GB of GDDR6X memory, 450 W total graphics power and an 850 W minimum system power recommendation for its Founders Edition.
- GPU memory
- 24 GB GDDR6X
- Total graphics power
- 450 W
- Reference system power
- 850 W minimum
- Main constraint
- The whole model and working headroom must fit in 24 GB
What makes a build genuinely custom.
The hardware brand is rarely the unusual part. The custom work is fitting the whole system around software, data and operating conditions that a standard appliance quote does not cover.
CUDA-only software
A research, vision, engineering or data package may depend on NVIDIA drivers and libraries. We verify the supported versions before choosing the GPU.
Local fine-tuning
If training must stay at your premises, the workload may need more GPU memory, storage and cooling than the everyday Harness does. That becomes a separate scope.
Storage and network constraints
Large document stores, shared imaging data, unusual backup targets or a pair of high-speed appliances can change the chassis and network design.
Windows or Linux policy
Some businesses have a fixed operating-system standard. We quote the engine and model formats that are actually supported there, rather than assuming the Mac setup carries across unchanged.
Heavy concurrent use
Several people asking for large jobs at once is different from one person asking a hard question. We size for the busiest realistic period, not the staff total on paper.
Rack or multi-GPU work
Once the brief reaches server power, heat and support requirements, we treat it as infrastructure work and quote the operating environment as carefully as the cards.
How we get to a useful quote.
- Bring the work
- The models, documents, software and busiest real use
- Check compatibility
- Operating system, engine, model format, VRAM, power and network
- Name the exact build
- Hardware, storage, warranty, configuration and any on-site work
- Test before handover
- The agreed jobs, on the quoted configuration, with the limits recorded
Hardware availability and supplier pricing can move. The quote is the current number; this page is the explanation of why we would choose the parts.
Questions about custom hardware
Are Strix Halo and DGX Spark custom-only machines?
No. Both are part of our standard range. Strix Halo is the main Windows or Linux choice, and DGX Spark is the main NVIDIA and CUDA choice. See the Strix Halo machines or see DGX Spark.
This page is for paired systems, full RTX workstations and jobs that need a non-standard hardware or software scope.
Can two machines work together?
For supported NVIDIA workloads, two DGX Spark systems can be connected over their ConnectX-7 interfaces. NVIDIA describes dual-Spark support for selected models up to 405 billion parameters.
A different option is one appliance per team or workload. That does not turn them into one larger computer, but it can be the cleaner way to separate load, documents and maintenance. We quote either arrangement only after checking the exact job.
When does an RTX workstation make sense?
When the software requires CUDA, the work needs a discrete NVIDIA GPU, or a standard desktop appliance does not have the required GPU memory or throughput. The model still has to fit in the card's VRAM, with room for the running job.
We check the model, software, power, cooling and noise before naming a card. A larger number on the box is not enough on its own.
Why is there no fixed custom-build price?
The standard tiers have published starting prices on the pricing page. A custom build may include two appliances, a workstation chassis, a particular GPU, extra storage or workload-specific setup, and those costs move.
The written quote names the exact hardware, operating system, configuration work and current supplier assumptions before anything is ordered.
Can we supply our own machine?
Yes. Bring us the machine and we will assess what it can genuinely run — memory, storage, operating system, and how much headroom is left once a model is loaded.
If it suits the work, the hardware line comes off the quote entirely and you pay for the software and configuration only: the Harness, the local models, your own templates and the setup. If it is not up to the job, we will say so plainly and tell you what it would take, rather than install something that will disappoint you.
Bring the awkward requirement.
Tell us which software has to run, which model you are trying to fit and who will use it. We will start with the standard range and move to a custom build only when the job gives us a reason.
Bring us one job your team repeats every week. We will show you the system doing it →