Build lists
A build list is a table of one AI setup: the computer, the model, the software, and what each line costs. You can fill that table in today; saving it on Polaris, a permanent link, other members' lists, and the Polaris Verified Build mark open later.
Beginner
Three rows and a total
Start with three rows and a total. Name the list after the job, for example "Home chat on my own computer".
Each row needs five fields: the kind of thing, its name, what you use it for, the price or the word Free, and the date you checked that price. Use one row for the computer, one for the model, and one for the program that runs the model. A free program still gets a row, with a price of zero, so the figure at the bottom is a real total.
VRAM is the memory on the graphics card. The model file is a set of numbers called weights. Hugging Face's deployment guide gives a rough size for those weights in 16-bit form, bfloat16 or float16: about 2 GB of VRAM for each billion parameters. Their worked example is Llama 2 70B at about 140 GB for the weights alone, before extra memory for a long conversation. Write your card's memory and that estimate on the computer row. If the estimate is larger than the card, say so in the notes.
Quantization stores each weight with fewer bits, often 8 or 4, so the same model uses less memory. The guide shows the cut can be large, and that answers can shift, so the model row should name the bit level.
For an ordinary desktop, open a public parts list on PCPartPicker, read the current total, and copy that figure into your computer row with the date beside it. Prices move, so the date is part of the fact.
Three-row starter list
Make a starter AI build list with exactly three rows: computer, model, and the program that runs the model. My budget is [amount and currency]. My graphics card has [X] GB of memory. I want to [task]. Use about 2 GB of 16-bit weight memory per billion parameters, and say whether 8-bit or 4-bit is what makes the model fit. For each row give the name, the role, the price or Free, and the price date. Where you do not know today's street price, write Unknown and "check before buying". End with a one-off total and a monthly total. Do not invent prices.
Intermediate
A list you can paste and share
A list that you will share needs a name, a currency, and a one-line purpose. Add a type to every row: machine, model, tool, course, or prompt. A tool is software. LocalAI is one example: an open-source engine under the MIT license that runs models on your own computer, with a graphics card or on the processor alone. A course row is a class you still have to take, with its fee. A prompt is a saved instruction you reuse. Its price is usually zero, and the note should say which model it expects.
Keep two money columns. One is a one-off cost, such as a graphics card. The other is a monthly fee, such as a hosted API. Sum them separately. A single total that adds a card to a subscription treats different kinds of money as the same thing.
You can paste the list today. Use the same columns in both formats: type, name, role, one-off price, monthly price, price date, and link. A Markdown table works on many modern forums. A BBCode table works on older boards. The blocks below are the whole export. Copy them into a text file, fill the blanks, and keep the price dates.
Before you send the file to anyone, remove serial numbers, account names, and anything that identifies a person. A link to a product page is fine. A photo of an invoice with your address is a data leak.
Markdown and BBCode tables
Turn the rows I paste below into two tables with these columns: type (machine, model, tool, course, or prompt), name, role, one-off price, monthly price, price date, link. Output a Markdown table first, then a BBCode table. Keep my figures exactly. If a date or a link is missing, write MISSING in that cell. After the tables, list only the rows that are missing a date or a link. Do not add products I did not name. Rows: [paste rows here]
Expert
A production list and the DGX Station Polaris plans to raise money for
A production list is the company record of the machine and of any personal data it processes. Personal data, in the EU General Data Protection Regulation, is information that can identify a living person. Name the controller (the organization that decides why the data is processed), the city where the data is stored, and whether prompts or files leave the building. Article 32 of Regulation (EU) 2016/679 requires security matched to the risk. Link the regulation, and keep a "GDPR-safe" label off unless a qualified adviser has signed it for this system.
The Open Compute Project publishes specifications for rack chassis and sleds. Link the matching specification on the hardware row.
NVIDIA's DGX Station product page is the source for the community reference rows: one Blackwell Ultra GPU with 252 GB of HBM3e (high-bandwidth memory on the GPU), a 72-core Grace CPU with 496 GB of LPDDR5X, NVLink-C2C at 900 GB/s, up to 748 GB of memory shared by CPU and GPU, four M.2 Gen5 slots, ConnectX-8 networking up to 800 Gb/s, Ubuntu with NVIDIA AI Developer Tools, up to seven isolated GPU slices, and 1,600 W on a 20 A circuit. Disks, cooling, and price wait for a written quote. The development guide names DeepSeek-R1, GPT-OSS-120B, and a 235-billion-parameter Qwen model and discusses models up to about 1 trillion parameters. Add a model row only after a logged run, with the precision written down.
The Polaris Verified Build mark requires a source and date on every paid row, the power figure, the memory split, and a line for what is still unknown.
Review a production list
Review the AI build list I paste below for a system that may process personal data in the EU. Check for: controller, city of the disks, whether prompts or files leave the building, a link to Regulation (EU) 2016/679, power in watts, GPU memory, CPU memory, storage on its own rows, cooling, and the precision of each model. Quote the list when a field is present. List missing fields as gaps. Do not add the label GDPR-safe. Do not invent prices, disk sizes, or benchmark scores. If the list mentions a DGX Station, compare the figures with NVIDIA's published deskside specification and flag any number that has no source line. List: [paste list here]
Columns that keep a list honest
Use one row per item. Leave a cell blank only when you still have to look it up, and write the word Unknown there so a later reader can see the hole.
| Column | What to write |
|---|---|
| Type | Machine, model, tool, course, or prompt |
| Name | Product or model name, plus the size in parameters if it is a model |
| Role | The job this row does in one short clause |
| One-off | Purchase price, or 0 if free. Currency in the list title |
| Monthly | Recurring fee, or 0 |
| Price date | The day you read that price |
| Link | The page the price or the specification came from |
| Notes | Bit level for a model, wattage and memory for a machine, city of the disks for a company system |
Add the one-off column. Add the monthly column in a separate sum. Write the budget in the list title if you have one, in your own currency.
Markdown and BBCode you can copy now
Markdown:
| Type | Name | Role | One-off | Monthly | Price date | Link |
| --- | --- | --- | --- | --- | --- | --- |
| Machine | | | | 0 | | |
| Model | | | 0 | 0 | | |
| Tool | | | 0 | 0 | | |
BBCode, for boards that use a table tag. Preview it, because board software differs:
[table]
[tr][td]Type[/td][td]Name[/td][td]Role[/td][td]One-off[/td][td]Monthly[/td][td]Price date[/td][td]Link[/td][/tr]
[tr][td]Machine[/td][td][/td][td][/td][td][/td][td]0[/td][td][/td][td][/td][/tr]
[tr][td]Model[/td][td][/td][td][/td][td]0[/td][td]0[/td][td][/td][td][/td][/tr]
[/table]
Under the table, add two lines: one-off total, and monthly total. Those two lines are the list's total cost.
What a Polaris Verified Build has to show
The mark is a checklist, applied to a list that already has sources. A list is ready to submit when every line below is filled or marked Unknown on purpose.
- Every paid row has a link and a price date.
- One-off costs and monthly fees are summed apart.
- The machine row states power in watts, GPU memory, and CPU memory, each tied to a manufacturer page or a quote.
- Storage, cooling, and network gear are their own rows when they are not included in the base machine.
- Each model row states the parameter count, the bit level, and whether someone has loaded it.
- A company list names the controller, the city where the data is stored, and whether any prompt or file is sent outside that building, with a link to Regulation (EU) 2016/679.
- A final row, titled Unknowns, lists every figure that is still blank, including disk size and street price when no quote exists yet.
If the community DGX Station is bought, start from NVIDIA's published deskside figures and leave disks, cooling, and price on the Unknowns row until a written quote arrives.
A shared community DGX Station would be the pinned machine on this list, with the published 252 GB of GPU memory for models whose weight files are larger than a home graphics card, while everyday work stays on the short list each person keeps. About the shared station
Sources
- NVIDIA DGX Station product pageManufacturer specification: 252 GB HBM3e, 496 GB LPDDR5X, up to 748 GB coherent memory, four M.2 Gen5 slots, ConnectX-8 up to 800 Gb/s, up to seven GPU slices, and 1,600 W on a 20 A circuit. Drives and street price are not on this page.
- NVIDIA DGX Station Development GuideWhat the deskside Grace Blackwell machine is built to do, including local inference and fine-tuning and the large models NVIDIA names there, up to about 1 trillion parameters.
- Optimizing LLMs for Speed and MemoryHugging Face guide with the 16-bit weight rule of about 2 GB of VRAM per billion parameters, the Llama 2 70B example at about 140 GB, and measured drops when the same model is loaded in 8-bit and 4-bit. Summarized here; the page keeps the worked examples.
- PCPartPicker completed buildsPublic desktop parts lists with prices and totals you can read and date. Prices belong to that site and change; copy the figure into your own row with the day you checked it.
- LocalAIOpen-source engine for running models on your own hardware, including CPU-only machines. MIT license: https://github.com/mudler/LocalAI/blob/master/LICENSE — this page links the project and does not copy its code.
- Open Compute Project, ServerPublic project page for open rack and server specifications, including chassis and sleds. Link a specific accepted spec from a hardware row. OCP publishes those files under its own IP policy; this page does not reproduce them.
- Regulation (EU) 2016/679, General Data Protection RegulationOfficial EU text. Article 4 defines personal data and the controller. Article 32 requires security measures appropriate to the risk. A machine in your building still needs this reading if it processes personal data.