Local AI, compared
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 supercomputers that approach the frontier chatbots. 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 R16.07 to the dollar where they do not, rounded, and stated excluding VAT unless marked. Model comparisons use the Artificial Analysis Intelligence Index, a single score across nine reasoning, coding and agentic tests. It measures open-ended ability, not grounded document work, and every model on this page, local or cloud, can make mistakes.
The ladder
ChatGPT Business, Claude Team or Microsoft Copilot on a per-seat plan. Nothing to buy, nothing to run. Enterprise tiers cost two to three times more and want 150 seats.
A 24GB slice of an RTX PRO 6000 in a South African data centre, unmanaged. You install and administer everything, and it is never yours.
A 24GB Mac mini in your office, configured, managed and supported. Ships with a 4-billion-parameter model tuned for fast, cited answers from your documents (index 20). The same box can run a 27-billion-parameter model that scores 52, at reading speed.
An NVIDIA DGX Spark, a Ryzen AI Max mini PC or a Mac Studio with 128GB of memory. Runs the same 27B model at full precision and 35 to 120-billion-parameter models, faster and for more people.
A Mac Studio M3 Ultra with 256GB. Runs the best open models that exist at 4-bit: GLM-5.3-Flash (index 57) and DeepSeek V4 Flash (52). This is the open-weight intelligence ceiling; nothing you can buy at any price is smarter than this today.
An NVIDIA DGX Station GB300 with 748GB of memory. Runs trillion-parameter open models such as Kimi K2.6 and the 256GB tier’s models at full precision, for dozens of people at once.
An HGX H200 or B200 server with eight data-centre GPUs. Everything above at full precision, for hundreds of users.
Index scores are the Artificial Analysis Intelligence Index v4.1, September 2026. Reference points on the same scale: GPT-4o 8, o3 31, GPT-5.4 mini 41, GPT-5.4 53, GPT-5.5 60, Claude Opus 4.7 57, the current leader about 66. Two things the ladder shows: the best model you can own scores about 57 against a leader at 66, one tier behind, not an order of magnitude; and that ceiling is reached at the 256GB desktop. Everything above it buys speed and users, not a smarter model. Small models are genuinely weaker at open-ended reasoning; the gap closes sharply when the task is answering from a specific set of documents with citations, which is what the appliance is built to do.
Where the appliance sits
Everything below it sends your files away. Everything above it needs someone to run it, and past the 256GB desktop it stops getting smarter and only gets faster. 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.
The appliance costs more. What it buys is set out below.
What you are paying for
Secure and private
There is no outbound path. No processor agreement to audit, no offshore transfer, nothing to explain to a regulator or a client about where their matter went. That is the whole reason the appliance exists.
Stable performance
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
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 that cannot leak, or a bigger one that might.
The frontier chatbots are better at open-ended reasoning than anything that fits on a desk, and they will stay ahead. For a firm bound by a confidentiality duty, that is not 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 never sends them anywhere. When a bigger private machine makes sense for your firm, we will say so and quote it.
Tell us how many people and how many matters, and we will show where each option lands for you.