Hardware

Every price here is indicative, in NZD, excluding GST, freight and installation, and must be re-quoted when a proposal is written. GPU and memory prices have moved by tens of percent within a single quarter. The figures below were last reviewed on 2026-10-13, at about 1.72 NZD to the US dollar.

The tiers below are sized for a team using models interactively. If your workload is one automated process, or a small team, you may need much less: see sizing for your workload.

The five tiers at a glance

Data-centre and RTX PRO cards only. GeForce cards are cheaper per gigabyte but the driver licence restricts data-centre use, so Heliast refuses them at hardware acceptance.

What each tier is built from

Indicative NZD. GPU lines use US$13,250 (the June 2026 NVIDIA listing) for the low end and US$16,000 (NVIDIA's marketplace, October 2026) for the high end.

Host memory is sized for serving, not for the model. Weights and the context cache live on the card. Server memory now costs several times what the original design budgeted (about NZ$3,500 for one 64GB module against NZ$625 implied by the original line), so the earlier reference (256GB, 512GB, 1TB) is no longer the default. Use those larger sizes only if you plan to offload to the CPU or run large ingestion jobs on the same host. With the earlier sizes the totals are about NZ$7,000 (Tier 1), NZ$14,000 (Tier 2) and NZ$28,000 (Tier 3) higher.

Not re-checked in this review: CPU, storage, network and chassis lines, which are carried over from the original design. Treat them as older than the GPU and memory lines.

Tier 2 and 3 cards do not pool memory: the RTX PRO 6000 has no NVLink, so a model that must span cards runs over PCIe, and a model larger than one card's memory needs pooled hardware instead.

Tier 0 in detail

Both 128GB options move data at roughly 273GB/s, about a seventh of the Tier 1 card. That is fine for one person exploring and for showing the compliance workflow. It will not hold latency targets under concurrent load.

What a tier can hold

Memory a model needs is its weights, plus a context cache for every conversation in flight. The weights figure is on each model's verdict page. The cache grows with context length and with how many conversations run at once, so it is the part that runs out first.

A worked case, as an estimate and not a measurement: a 70B-class model at FP8 takes about 70GB, which leaves about 26GB on a 96GB card. With the 8-bit cache that such a model typically uses, one 32,000-token conversation needs about 5GB, so a single card holds on the order of four or five such conversations at once. A 27B-class model at FP8 takes about 27GB and leaves room for many more. This is why a well-chosen smaller model is usually a cheaper capacity increase than more hardware, and why Compare exists.

Sizing note. "Concurrent users" and "conversations in flight" are different numbers, because most users are idle most of the time. How the two relate depends on how your staff work. Agentic and long-context workloads need much more cache and are sized separately.

Performance targets

Targets assume FP8 serving at 32,000 tokens of context and a model sized to the tier. Until a Tier 1 run exists, latency is a target to be measured at deployment and re-measured after any model change. It is not a guarantee.

What the console means by "one card", "one server" and "cluster"

Each verdict page and the Admission filters describe a model by how much memory its smallest published weights need:

These describe a model, not a customer's estate. The tier a model fits is shown on its "Use this model" panel.