MyriadPulse

Local AI, compared

What private AI costs, from a desk to a data centre.

We looked at every way a South African firm could run AI on its own documents without sending them anywhere: from the smallest machine that does useful work to the data-centre racks that now sit within a few points of the best chatbots in the world. This is where each option lands, and why the appliance is the one we sell first.

Prices checked on 2 September 2026 from South African retailers where they exist, converted at R16.72 to the dollar where they do not — the average over the past twelve months rather than today’s rate, so these figures do not swing with the currency. Rounded, and excluding VAT unless marked. Apple prices are from Apple’s published list for the M5 machines announced on 25 August 2026, which ship from 22 September; the 128GB figure is still the outgoing generation on South African shelves and we will requote it when the M5 rand price lands. How we judged capability, and the numbers behind it, are explained under the table.

The ladder

Nine ways to get AI onto your files

Public AI subscription

ChatGPT Business, Claude Team or Microsoft Copilot on a per-seat plan. Nothing to buy, nothing to run.

CostR0 to ownR330 – R420 per user per month, before VAT, set by the vendor
How capable is it?The best there is, by definition — whichever model the vendor happens to be serving this month.
Documents leave the building?Yes. Every file and every question is processed on the vendor’s own infrastructure, wherever it keeps it.

Rented GPU server

A slice of an RTX PRO 6000 in a Johannesburg data centre, unmanaged, in three sizes: 24GB, 48GB or 96GB. You install and administer everything, and it is never yours. The rent covers the data centre, the power and the connectivity, which an office machine does not need.

CostR0 to own, everR17,500 a month at 24GB, R34,000 at 48GB, R67,000 at 96GB
How capable is it?At 24GB, identical to the appliance below. At 96GB it runs what the 128GB desk two rungs down runs, for R67,000 every month instead of once.
Documents leave the building?Yes, to a local data centre under your own contract.

The MyriadPulse appliance

A 24GB Mac mini in your office, configured, managed and supported. It ships with a small, fast model tuned to answer from your documents and cite what it used. The same box can also run a model roughly seven times larger, which is markedly better at reasoning and reads back at about the speed you would read it.

CostR25,000 to ownR7,000 per month for 24 months, then optional support and licence at R5,000. Larger appliances are quoted on the same basis
How capable is it?Weaker than a chatbot at open-ended reasoning, and we would rather say so. Answering from your own files with citations is a narrower job, and the gap on that job is much smaller. The larger model it can also run is a big step up again.
Documents leave the building?No.

128GB desktop

An NVIDIA DGX Spark, a Ryzen AI Max mini PC or a Mac Studio with 128GB of memory. Runs the appliance’s larger model uncompressed, and models several times larger again, faster and for more people at once.

CostR65,000 – R95,000Bought once. Power, plus whoever runs it — or we supply and manage it as a larger appliance
How capable is it?Running the same model as the appliance, the same answers — arriving faster, and for more people at once. Its extra memory also holds models several times larger, and those do reason better.
Documents leave the building?No.

256GB desktop

A Mac Studio M5 Ultra with 256GB, shipping 22 September 2026. It reads memory about half again as fast as the model it replaces, which is what makes answers appear quicker. Runs the strong middleweight open models, compressed to fit. The very best open models are larger than this and need the rung below.

CostAbout R181,000Bought once, with 1TB of storage. More storage takes it to R306,000. We can supply and manage this too
How capable is it?A real step up in reasoning over the appliance, and still a step below the best chatbots you could subscribe to.
Documents leave the building?No.

512GB desktop

The same Mac Studio at 512GB, which Apple brings back in late October 2026. This is the smallest machine that runs GLM-5.3, which ties for the best AI in the world that anyone can own outright. It has to be compressed to fit — 365GB even then, so 256GB will not do — and compressing a model can cost it a little quality, so treat the published score as the full model’s rather than a promise about the build on your desk. It answers at a working pace because only a small fraction of it is used for any given word. This is the top of the ladder on the scores — and it is a desk.

CostAround R325,000 – R370,000Our estimate: Apple has not priced 512GB yet, so we step up from its published 256GB machine at R181,000 by one memory tier. Confirmed on release
How capable is it?Level with the best AI anyone can own at any price, and a short step behind the best you can subscribe to — the closest those two have ever been. On the published scores for the full model; a compressed build may come in a little under.
Documents leave the building?No.

Desktop supercomputer

An NVIDIA DGX Station GB300 with 748GB of memory. Runs GLM-5.3 with room to spare, and the 256GB tier’s models at full precision, for dozens of people at once.

CostR1.5 million – R2.1 million, plus shipping, clearing and import VATPower, cooling, a specialist
How capable is it?No higher on the benchmark than the desk above it. What you are buying here is speed and the number of people it can serve at once.
Documents leave the building?No.

Eight-GPU server

An NVIDIA HGX B300 with eight data-centre GPUs and 2.3TB of high-bandwidth memory. The first rung that can hold Kimi K3, the other model tied for best-you-can-own. It is roughly four times the size of GLM-5.3 and needs about 1.4TB even compressed, before the working memory each conversation takes.

CostR6.7 million – R8.4 million, plus shipping, clearing and import VATAbout R890,000 for each of the eight GPUs. Liquid cooling, three-phase power, a specialist
How capable is it?Also no higher on the benchmark than that desk. It holds a far bigger model, which is not the same thing as a better-scoring one.
Documents leave the building?No.

Seventy-two-GPU rack

An NVIDIA GB300 NVL72: a whole rack wired as one machine, 72 GPUs and 20TB of high-bandwidth memory, drawing over 100kW. This is what Kimi K3’s own authors recommend for running it properly, and it is one of the largest single-rack systems on the market. It is a data-centre installation, not a room in an office.

CostR62 million – R109 million, plus shipping, clearing and import VATA data centre, over 100kW of power and cooling, and a team
How capable is it?No higher on the benchmark than that desk. What a rack of this size buys is throughput and the number of people it can serve at once.
Documents leave the building?No.

How we judged “capable”

There is an independent benchmark, the Artificial Analysis Intelligence Index, that scores every serious AI out of 100 on reasoning, coding and research-style tasks. It is the fairest public yardstick there is, and it is what the column above is based on. Two things are worth knowing before you lean on it.

It measures open-ended cleverness, not the job you would buy this for. Reasoning from a blank page is a harder and broader test than answering from a specific set of documents and showing you where each answer came from. On the second job the differences between these machines are much smaller than the scores suggest, and that second job is the whole purpose of the appliance. Every model here, ours and the chatbots, can still be wrong; a person still checks the work.

The gap has closed a lot, and recently. The best AI you can buy outright and keep in your building now scores 60. The best chatbot in the world scores 66. A few months ago that gap was nine points, and it has never been narrower. The models that reach 60 — Kimi K3 and GLM-5.3 — need the 512GB desk to run. On this benchmark today, that is where the ladder stops climbing: the R62-million rack scores exactly the same. What the rungs above buy is throughput and the number of people served at once, not a higher score.

Scores checked 3 September 2026. On the same scale: the ChatGPT most people first tried 8, a strong 2025 reasoning model 31, the appliance’s shipping model 20 and its larger model 52, GLM-5.3-Flash 57, Kimi K3 and GLM-5.3 both 60, GPT-5.6 Sol 61, Claude Opus 5 63, and the leader, Claude Fable 5.1, at 66.

Reading the ladder

More money buys three different things

Machines at this level are sold on one headline number, and it is usually the wrong one to compare. Three separate things go up as the price goes up, and only one of them decides which AI you can actually run.

Memory

Decides which models you can run at all

A model has to fit in memory to run, and that behaves like a wall rather than a slowdown: one step below the line it will not load, one step above it runs fine. There are ways to spill a model across a machine, or across several, but they cost speed and complexity, so treat memory as the thing that decides what you can practically run. It is also the reason a cheaper machine with more memory can run models a dearer one cannot — memory and speed are sold separately, so compare capacity before you compare prices.

Memory bandwidth

Decides how fast the answer appears

Producing each word means reading the working model out of memory again, so for the large models on this page the speed you feel is set mostly by how fast memory can be read, rather than by the graphics chip. This is the number that separates two machines holding the same amount of memory, and it is rarely the one on the price tag.

Graphics cores

Decide almost everything except the above

More cores speed up reading a long document in, generating images and training, and they are worth having. What they do not do is unlock a model the memory cannot hold, and they move answer speed far less than the jump in price suggests. The top chip in a tier usually runs exactly the same models as the middle one.

There is a fourth use for memory that matters more to a firm than any of this: it does not have to go on a bigger model. The same memory can hold a modest model plus room for several people asking at once and for long documents held open in context. For most firms that is worth more than loading the largest model the machine can physically hold, and it is how we size an appliance.

The question everyone asks

Rent the same memory and you have paid for the machine in six weeks

Renting a GPU server is the first thing most firms price up, so it is worth doing the arithmetic properly. The machine is a real one in a real data centre and somebody else keeps it running — but you are renting memory you could buy outright, and the crossover comes fast.

MemoryRented, monthlyBought, onceRent reaches the purchase price inA year of rent
24GBR17,500about R21,300five weeksR210,000 — about ten times the machine
96GBR67,000about R95,000six weeksR804,000 — about eight times the machine

Rental prices from a South African host’s published list, excluding VAT. Purchase prices are iStore retail converted to exclude VAT. Two things this table is not saying. The rented server is a bigger machine overall — far more ordinary memory, an enterprise graphics card, redundant power and a fast link — and the rent pays for a data centre, which a machine in your office does not need and does not have. And renting suits work that stops: a burst of research, a one-off project, a proof of concept. What it does not suit is running the same workload for years, and it does not answer the question this whole page is about — the files still leave your building, for a data centre and an administrator outside the firm.

Where the appliance sits

The cheapest rung that does real work, and the only one that comes managed.

Everything below it sends your files away. Everything above it needs someone to run it, and past the 512GB desktop the scores stop climbing and only the speed goes up. The appliance is the smallest machine we could make do useful, checkable legal work, and we manage it so nobody in the firm has to become a systems administrator. When a firm outgrows it, the next rung is a one-off purchase we can quote, not a subscription.

A frontier subscription for five peopleR20,000to R25,000 a year, and risingAt R330 to R420 per seat per month, before VAT, on the business tiers. The price and the model are set by the vendor and can change at any renewal, and the files leave.
The appliance, 24-month pathR96,000the first year, then R84,000R12,000 setup and R7,000 a month, for up to five people, with no per-seat fees — R180,000 over the 24 months, and the machine is yours at the end. Then R5,000 a month, optional.

So the appliance costs somewhere between R60,000 and R75,000 a year more than putting the same five people on a subscription, and we would rather put that number in front of you than round it in our favour. The question it leaves is the real one: is keeping matter files in your own office worth that to your firm? What the money buys is set out below.

What you are paying for

Three things a subscription cannot give you

Secure and private

The files stay in the room

Your documents, your questions and the answers are processed on the machine in your office, and are not sent to any external AI service. No third party is handed a matter to process, so what you tell a client about where their information goes is a fact about your own office. That is the whole reason the appliance exists.

Stable performance

Nobody changes the model under you

A public AI vendor retires models, swaps them, and tunes them for its own economics, not for your practice. The model on your appliance changes only when we update it, on a schedule you know about, and we test the legal workflows before it ships.

Stable cost

Your price is not somebody else’s pricing strategy

Subscription AI is priced for growth today and repriced when the market allows. The appliance is a fixed monthly for your term, then an optional support and licence fee adjusted by CPI and capped at 6%. If cloud AI prices double in three years, it does not reach you.

The honest trade

A smaller model on your desk, or a bigger one on someone else’s.

The best chatbots are still better at open-ended reasoning than anything you can own, though the distance is far shorter than it was a year ago and still closing. For a firm bound by a confidentiality duty, that was never really the question. The question is what you can put the files into at all. The appliance answers from your documents, cites what it used, and does not send them to an external AI service. When a bigger private machine makes sense for your firm, we will say so and quote it.

Want the comparison for your firm's numbers?

Tell us how many people and how many matters, and we will show where each option lands for you.

Talk to us See pricing