Choosing the hardware for a home AI assistant: RTX 5090 vs RTX PRO 6000 and everything in between

The configurator offers a lot of choices. This is the guide to the ones that matter.

The accelerator decides almost everything

A home AI assistant lives or dies by its graphics card, and specifically by how much memory is on it. Model weights have to fit in that memory, along with the context the model is reading. More memory means larger, more careful models and more of your household in view at once.

Accelerator memory	Typical models	What that feels like
16 GB	14B-class	Triage and short drafts. Long documents read in sections.
32 GB	32B-class	Confident triage, drafting with facts, calendar reasoning.
48 GB	70B-class	Multi-step tasks with low false-alarm rates.
96 GB	120B-class (gpt-oss-120b)	The reference tier. Whole household in one pass.
192 GB	235B-class	Frontier-adjacent quality offline, or two households.

RTX 5090 or RTX PRO 6000?

These are the two cards most people end up choosing between.

The GeForce RTX 5090 has 32 GB, is very fast, and costs about a quarter of the PRO 6000. It is the heart of our Essential build and it does the household job well. Its drawbacks are noise and heat: it draws up to 575 W and is audible under load, so it belongs in a study or cupboard rather than a living room, and the quiet package is worth adding.

The RTX PRO 6000 Blackwell has 96 GB with error-correcting memory, and is the reference Majordomo accelerator. It runs gpt-oss-120b natively with room to spare, which is what makes the long-context tasks reliable. The Max-Q version delivers the same 96 GB at half the power, which makes it the right choice for a quiet room. It costs more, and it is the single biggest line in the Household build, but it is also the component that most changes what the assistant can do.

Between them sit the RTX PRO 4500 (32 GB, 200 W, blower-cooled and quiet) and the RTX PRO 5000 (48 GB), both good choices if you want workstation reliability without the flagship price. The AMD Radeon AI PRO R9700 is a strong-value 32 GB card for people comfortable with the ROCm software stack.

Processor and system memory

The processor matters less than you would think; it indexes documents and runs transcription while the graphics card does the thinking. A Ryzen 7 is plenty for one card. Two cards need a Threadripper platform, which also brings error-correcting memory.

For system memory the rule is simple: match or exceed the accelerator memory. Models stage through system RAM as they load, and a 96 GB card paired with 64 GB of RAM will load slowly and feel sluggish. The configurator warns you if you do this.

Storage

Four terabytes suits most households: models, a decade of email, an indexed document archive and transcripts fit comfortably. Add the archive drives if you want scanned paper, photos and recordings on the same machine.

Noise, power and where it lives

A Household build draws about 168 W on a typical day and peaks near 800 W when it is thinking hard. It will not trouble a standard socket. Noise is the thing to plan for: a Max-Q card or the quiet package makes the machine comfortable in a hallway cupboard; a 5090 without them wants a door between it and you. Battery backup is worth it anywhere power cuts happen.

Three starting points

• Essential, around $4,000: RTX 5090, Ryzen 7, 64 GB, 2 TB. Capable and quiet enough for one or two people.
• Household, around $11,000: RTX PRO 6000, Ryzen 9, 128 GB, 4 TB. The reference build; runs the full model set.
• Estate, around $26,000: two Max-Q cards, Threadripper, 256 GB, 8 TB plus archive, fully managed. Several properties or a family office.

Prices are current retail estimates before tax and delivery; your estimate confirms them. Every part is yours to change in the configurator, and if you would rather not think about any of this, opt in to the white-glove service and we will choose, build, install and look after it for you.