AI & MACHINE LEARNING WORKSTATIONS
AI workstations, built for training, inference and generation
Your graphics card’s memory decides which models you can run at all. We build machines with enough of it to serve models, fine-tune on your own data and generate images and video — no job queue, no hourly cloud bill, and data that never leaves your building.
- 5.00 / 5.00 RATEDFrom 2.8K+ customer reviews
- READY TO SHIPAI builds held in stock
- CUSTOMISE ANY BUILDCPU, GPU, RAM, case & PSU
- AUSTRALIAN TEAMSeven Hills, NSW
- PAY YOUR WAYAfterpay, PayID & cards
Select your software
Pick one. The machine we’d start you on appears here, and the four machines below re-describe themselves for it.
WHAT IT DEPENDS ON
VRAM before anything else
The card’s memory decides which models fit; its compute decides how long they take.
HOW IT SHOWS UP
An out-of-memory error rather than a slow run — the job doesn’t start at all.
WHAT IT DEPENDS ON
VRAM, then the input pipeline
CPU cores and fast NVMe keep the card fed between batches.
HOW IT SHOWS UP
A GPU sitting idle between batches because data loading can’t keep up.
WE’D START YOU ON
CORE FOCUS
A 16-core Ryzen 9 9950X3D, 64GB DDR5 and 2TB NVMe behind two GPUs.
VIEW CORE FOCUSWHAT IT DEPENDS ON
The card’s memory is the ceiling
Quantised weights lower it; a longer context window raises it again.
HOW IT SHOWS UP
A model that runs on one machine and refuses to load on another — even one with a faster card.
WE’D START YOU ON
Size VRAM to the model
CORE LITE for 7–14B models. ARTISAN MAX for 32 GB on a single card.
SIZE IT IN 10 SECONDSWHAT IT DEPENDS ON
GPU speed, once the model fits
Most image models fit in 16 GB; the card’s compute sets the time per image.
HOW IT SHOWS UP
Long waits per generation, and heavier video workflows that won’t load.
The machines that run them
Four machines on one ladder, ordered by GPU memory. CORE LITE is the development seat. ARTISAN is one faster card. CORE FOCUS adds a second card when fine-tuning is the wait. ARTISAN MAX puts 32 GB on a single card. Every one can be customised before we build it.
-
01 · 16 GBOUR PICKCORE LITE
The development seat
GPU memory 16 GB
- Intel Core Ultra 7 270K Plus
- ASUS NVIDIA GeForce RTX 5060 Ti Dual OC - 16GB
- 32GB Lexar THOR G2 RGB (2x16GB) 6000MHz DDR5 - Black
- 1TB Gen5 Samsung 9100 PRO M.2
FOR
Prototyping, notebooks and small fine-tunes on one 16 GB card.
Writing and debugging pipelines before the long runs.
7–14B models with room left for context.
Image generation on a budget — slower per image.
-
02 · 16 GBOUR PICKARTISAN
One faster card
GPU memory 16 GB
- Intel Core Ultra 7 270K
- GIGABYTE GeForce RTX 5080 WINDFORCE SFF - 16GB
- 32GB Lexar THOR G2 RGB (2x16GB) 6000MHz DDR5 - Black
- 1TB Lexar NQ790 M.2 NVMe
FOR
Faster single-card training for models that fit 16 GB.
Quicker epochs when the model fits on one card.
Faster responses on models that fit 16 GB.
Fast generations on a single RTX 5080.
-
03 · 2 × 16 GBOUR PICKCORE FOCUS
Dual-GPU AI workstation
GPU memory 32 GB total
- AMD Ryzen 9 9950X3D
- 2x ASUS GeForce RTX 5060 Ti Prime OC - 16GB (32GB VRAM)
- 64GB Lexar THOR G2 RGB (2x32GB) 6000MHz DDR5 - Black
- 2TB Gen5 Samsung 9100 PRO M.2
FOR
Two cards to split one job, or run two experiments at once.
Distributed training in one box, with 64GB RAM to feed it.
30B-class models split across two 16 GB cards.
Two generations at once — one per card.
-
04 · 32 GBOUR PICKARTISAN MAX
32 GB on a single card
GPU memory 32 GB
- AMD Ryzen 9 9950X3D2
- GIGABYTE GeForce RTX 5090 WINDFORCE OC - 32GB
- 64GB Lexar THOR G2 RGB (2x32GB) 6000MHz DDR5 - Black
- 2TB Gen5 Samsung 9100 PRO M.2
FOR
Larger models on one card, with no sharding to manage.
The fastest single-card training in the range.
30B-class models on one card — no splitting.
The fastest generations and the heaviest video workflows.
BEYOND 32 GB · QUOTE ON REQUEST
Custom AI server
For 70B-class models, inference served to a whole team, or training runs measured in days. Tell us the model, the precision and how many people will use it — we’ll come back with a spec and a price.
See every model side by side in the full workstation range.
What each workload needs
Training hits a wall at video memory. Data work asks for cores and RAM instead. Here’s where each one puts the load, what it looks like when a machine is short of it, and where we’d start you.
| WORKLOAD | WHAT IT DEPENDS ON | HOW IT SHOWS UP | START WITH |
|---|---|---|---|
| PyTorch | VRAM before anything else. Memory decides which models fit; compute decides how long they take. | An out-of-memory error rather than a slow run. | CORE FOCUSFine-tuning and parallel sweeps |
| TensorFlow | VRAM, then the input pipeline. Cores and NVMe keep the card fed. | A GPU idling between batches while data loading catches up. | CORE FOCUSDistributed training in one box |
| Running models locally | The card’s memory is the ceiling. Quantised weights lower it; longer context raises it. | A model that loads on one machine and not on another with a faster card. | Size VRAM to the modelNot to the benchmark |
| Image & video generation | GPU speed, once the model fits. Most image models fit in 16 GB. | Long waits per image; heavy video workflows that won’t load. | ARTISANFast single-card generation |
| Data science | Cores and system RAM. Many libraries never touch the GPU. | Jobs that scale with core count, and datasets that won’t fit in memory. | CORE FOCUS16 cores and 64GB DDR5 |
Tell us the models and frameworks you use and we’ll size the card to them.
WHERE YOUR WORK SITS
Running published models
Bounded by the card’s memory alone. One strong GPU covers it — CORE LITE, ARTISAN or ARTISAN MAX depending on model size.
THE PROFESSIONAL LOAD
Fine-tuning on your own data
Where a second card stops the waiting. Multi-GPU training and long sweeps are where owning the machine starts to beat renting it by the hour.
What makes a run faster
Five parts decide how fast AI work feels, in the order they matter — what each does, and what you notice when there isn’t enough of it.
Graphics
VRAM first, speed second. Memory decides whether the job runs at all; the card’s compute only decides how long it takes once it does.
Processor
Loading, decoding, tokenising and augmenting all happen on the CPU. Too few cores and the GPU waits for the next batch.
Memory
Keep system RAM at or above total VRAM. It holds the dataset between epochs and catches any layers offloaded from the card.
Storage
Training reads millions of small files in random order and writes checkpoints over and over — exactly what NVMe is for.
Power & cooling
A run is hours or days at full load. Throttling doesn’t throw an error; it just adds time to every epoch.
CORE LITE · ARTISAN · CORE FOCUS · ARTISAN MAX
VRAM is the first hard limit
Everything a run needs lives in the graphics card’s own memory: the model weights, the activations and, if you’re training, the gradients and optimiser states on top. If it doesn’t all fit, the job doesn’t start. VRAM is the first number to settle and the last one to compromise on — which is why our machines are ordered by card memory before anything else.
When a model doesn’t fit, every workaround costs something. Quantising to 8-bit or 4-bit gives up a little quality. Offloading layers to system RAM slows things sharply, because data crosses the PCIe bus for every token. Shrinking batch size and context makes training slower and lets the model see less at once.
- Batch size and context length are VRAM questions. Both cost memory before they cost time.
- Training costs far more memory than inference. A model you can serve comfortably may be out of reach to fine-tune at full precision.
- Two cards is more total VRAM, not one big pool. A single model only spans both if your framework splits it — otherwise each card holds its own copy.
Size it in 10 seconds
What’s the largest model you’ll run?
Estimated VRAM at 4-bit≈ 5–6 GB≈ 9–10 GB≈ 19–21 GB≈ 40 GB+
7–8B: Estimated VRAM at 4-bit ≈ 5–6 GB.
WE’D START YOU ON
CORE LITE$3,199
RTX 5060 Ti 16GB · 32GB DDR5
One 16 GB card holds a model this size with plenty left for long context — or run it at 8-bit for better quality.
VIEW CORE LITE13–14B: Estimated VRAM at 4-bit ≈ 9–10 GB.
WE’D START YOU ON
ARTISAN$5,399
RTX 5080 16GB · 32GB DDR5
Fits a single 16 GB card with headroom. The RTX 5080 answers noticeably faster than an entry card.
VIEW ARTISAN30–34B: Estimated VRAM at 4-bit ≈ 19–21 GB.
WE’D START YOU ON
CORE FOCUS$6,199
2× RTX 5060 Ti 16GB · 64GB DDR5
Past one 16 GB card. Two cards give 32 GB when your runtime splits the model across them. Prefer one card? ARTISAN MAX has 32 GB on an RTX 5090.
VIEW CORE FOCUS70B+: Estimated VRAM at 4-bit ≈ 40 GB+.
WE’D START YOU ON
Custom AI serverQuote
Multi-GPU, sized to your model
More memory than any single desktop card holds. We spec multi-GPU builds around your model, precision and number of users.
REQUEST A QUOTERough guide for running models with 4-bit quantised weights. Longer context, higher precision and fine-tuning all need more memory.
Why choose Nebula PC?
Anyone can order the same parts. What you’re really choosing is who builds the machine and who picks up the phone when you have a question.
-
5.00 / 5.00
from 2.8K+ customer reviews
-
Ready to ship
AI-capable builds held in stock
-
1300 553 985
talk to our Australian team
Real advice before you buy
Call, email or live chat with our team and we’ll match the machine to your models and frameworks.
Customise any build
Start from a ready-to-ship configuration and change the case, CPU, motherboard, RAM, GPU or power supply.
Pay your way
Afterpay, PayID, Shop Pay, Apple Pay, Google Pay and all major cards at checkout.
AI workstation FAQ
How much VRAM do I need?
Start from the largest model you plan to run and the precision you’ll run it at. As a rough guide at 4-bit, an 8B model needs about 5–6 GB, a 14B about 10 GB, a 32B about 20 GB and a 70B about 40 GB — before allowing for long context. Fine-tuning needs several times more.
One big GPU or two smaller ones?
One card is simplest: every framework can use all of its memory. Two cards give more total VRAM and let you run two jobs side by side, but a single model only spans both if your framework splits it across them.
Can I fine-tune on this, or do I need the cloud?
LoRA and QLoRA fine-tuning runs well locally on models that fit your VRAM. Full-precision fine-tuning of larger models is where a custom multi-GPU build comes in.
Do the GPUs need NVLink?
No. The GeForce RTX cards in these builds don’t support NVLink; multi-GPU frameworks communicate over PCIe, which works well for data-parallel training and for splitting a model for inference.
Can the team share it, or does it have to sit on my desk?
A workstation can serve models over your network to several people. Tell us how many users and how they’ll connect, and we’ll size the build — or quote a custom AI server.
How can I pay?
Afterpay, PayID, Shop Pay, Apple Pay, Google Pay, Visa, Mastercard and American Express are accepted at checkout.
Ready to ship
BUY NOWCurated AI-capable configurations held in stock. Professionally built and backed by Nebula PC — no configuration required.
Custom workstations
SEE THE MACHINESStart with the machine built for your software, then change what you like before we build it.
Talk to an expert
TALK TO A SPECIALISTTell us the models, frameworks and how many seats. We’ll come back with a spec and a price.
TALK TO AN EXPERT
Tell us what you’re running. We’ll size the machine.
The most useful detail is the largest model you plan to run, and at what precision. With that, we can size the card to the job rather than to a price bracket.