How Much Does It Cost to Run AI on Your Own Hardware?

Business owners keep asking us the same thing: our data is sensitive and the usage bill grows every month, so can we run the model ourselves instead of depending on an external service?

An honest answer starts with a correction: the weights of Claude and GPT are not published, so nobody runs them at home whatever the hardware. What you genuinely can run are open-weight models that today reach 200 billion parameters and more, and that come close on many practical tasks to what the paid services give you: summarising documents, answering questions about your own data, writing code, translating. The gap narrows every few months.

So the real question is not whether it is possible, but how many parameters your machine holds. That leads to the single rule that decides everything.

The rule that decides it: memory before speed

A model must fit entirely in memory to run. The arithmetic is simpler than it looks:

• At full precision: two bytes per parameter.
• At 8-bit quantisation: one byte per parameter.
• At 4-bit quantisation: half a byte per parameter — the practical choice for most uses.

Then add roughly 20 to 30% for context and runtime overhead. The result is this estimating table, which you can measure any machine against:

The figures include runtime headroom and are estimates; they shift slightly with the quantisation method and context length.

That table alone shortcuts most comparisons: look at memory capacity first, because it decides what you can run. Bandwidth and processor type decide how fast.

The options available today

Apple — the new Mac Studio

On 25 August 2026 Apple announced a new Mac Studio generation shipping from 22 September, with the 512GB configuration arriving in late October. It is currently the widest option on capacity:

• M5 Max: an 18-core CPU, a GPU of up to 40 cores with Neural Accelerators built into each core, and up to 128GB of unified memory at up to 614GB/s of bandwidth.
• M5 Ultra: up to a 36-core CPU, up to an 80-core GPU, and up to 512GB of unified memory — enough to run enormous models entirely on device.
• Thunderbolt 5 allows multiple units to be clustered, reaching up to 3x the performance of a single system for distributed inference.

Its real advantage is that memory is unified between CPU and GPU, so you do not need a separate card with its own memory. It ships with two developer frameworks: the new Core AI and the open-source MLX.

AMD — the Ryzen AI Max+ 395

It arrives in small desktop machines with 128GB of unified memory, runs Windows or Linux, and costs less than its direct Nvidia rival. It is the practical choice for anyone who wants 128GB at the lowest cost, or whose team prefers a familiar Linux environment.

Nvidia — DGX Spark

Roughly a one-litre box carrying a Grace Blackwell chip and 128GB, and the product that created this category. Its most important advantage is not the specifications but the CUDA ecosystem: most AI tooling and training frameworks were written for it first. If your team will train or fine-tune models rather than only run them, that is a decisive consideration.

The traditional route — a server with discrete GPUs

Still the right answer when you serve several users at once. Server cards deliver far higher bandwidth than unified-memory machines, and that shows up in concurrent requests rather than model size. The downsides: higher cost, plus power, heat and noise that need a suitable room.

The machines, side by side

Exact prices and where to buy

We took Apple's prices straight from its own US and Saudi stores on the date this article was published:

Buy from the Apple Saudi store or the Apple US store. These are starting prices per configuration; maxing the memory raises them considerably, and the 512GB option arrives in late October rather than at launch. Note that the Saudi price includes VAT while the US price is before tax.

AMD and Nvidia do not sell these machines directly at a single price: the AMD chip ships inside boxes from different manufacturers at roughly $1,500 to $4,000 depending on maker and configuration, and DGX Spark launched near $4,000 and has risen on the street. We found no official Saudi retail listing for either as of publication, so buying means going through a distributor or importing — in which case add VAT, customs and shipping. Their official pages: NVIDIA DGX Spark and AMD Ryzen AI Max+ 395.

Which open model do you run, and on what?

The machine is half the decision; the other half is the model itself. These are the leading open-weight models you can download and run today, with the machine that holds each at 4-bit and an estimate of its practical capability:

What does the capability percentage mean here? How close the model gets, on everyday business work — summarising, answering questions over documents, writing code, translating — to the strongest paid closed models. It is an Origami editorial estimate based on published specifications and the vendors' own reported figures, not the result of a single neutral test. The reason is plain: every lab publishes its numbers with a different harness and different settings, so no one neutral benchmark ranks all of these models today. Treat it as a guide, not a verdict, and test the model on your own tasks before committing.

Two notes that compress the whole table: models marked "active" use only a small share of their parameters per request, so they are faster than their size suggests — but they still need the full memory. And the difference between 72% and 90% disappears entirely in most office work, while showing clearly on long, complex tasks. Do not pay for 512GB to gain a margin you will never notice.

How to choose, practically

• One or two users? A unified-memory machine is enough and saves a great deal.
• A team using the system simultaneously? You will need discrete GPUs, because the problem is now concurrent requests, not model size.
• Training or fine-tuning? Lean towards CUDA.
• Want the largest possible model with the least complexity? The largest capacity today is Apple's.
• Limited budget? 128GB covers models up to 200 billion parameters at 4-bit, which genuinely meets most company needs.

The trade-off to understand before buying: unified-memory machines give you large capacity at lower bandwidth than server cards. They load the big model without trouble, but generate text more slowly. For chat and summarisation you will barely notice; for heavy batch processing you will.

When a private server is the wrong decision

We say this plainly because we see it often: the hardware is not the end of the project but the beginning. You are also buying responsibility for updates, backups, monitoring, and whoever fixes it when it stops on a Thursday. If your usage is intermittent and small, an external service's monthly bill is cheaper than the machine, its electricity and the time of whoever runs it.

A private server earns its place when three things meet: data that must not leave, steady heavy usage, and a person or party who will operate it. If one of the three is missing, redo the calculation.

The Origami view

We think the important question is not which machine to buy, but what you will run on it. Most people who ask us about a private server actually need a mid-sized model that answers questions about their company's documents, and that runs comfortably on 128GB — sometimes on less. Buying 512GB to run a 70-billion-parameter model is spending without return.

Our practical advice: define the task, then pick the model that does it, then buy the machine that holds that model with headroom. We help in that order — defining the task, selecting the model, and building the layer that connects it to your data and systems, because a model with no connection to your data stays a general tool rather than a system that serves your company. You will find us on the contact page.

Conclusion

You will not run Claude at home, but you will run capable open models that cover most business work. Size your need in parameters first: half a byte per parameter at 4-bit, plus headroom. Then choose between unified memory, which buys capacity at a reasonable cost, and discrete cards, which buy speed and concurrency. The most important decision still comes before the hardware: what you will run, and why.

Sources

• Apple's press release, 25 August 2026 — Mac Studio specifications with M5 Max and M5 Ultra, unified memory capacities and bandwidth, Thunderbolt 5 clustering, and the Core AI and MLX frameworks.
• Independent reviews from StorageReview and HotHardware of the AMD Ryzen AI Max+ 395 and Nvidia DGX Spark, including their capacity and price comparisons.
• The memory arithmetic follows the standard rule for running models: two bytes per parameter at full precision, scaling down with quantisation.
• Official Hugging Face model cards for Qwen3.8, Llama 4, Mistral, gpt-oss, DeepSeek V4, GLM-5.3 and Kimi K3 — the source of the parameter counts and licences. The capability percentages are an Origami editorial estimate, as explained above.