Almost everything you’ve seen from us runs on one computer: the images and video, the 3D avatar, the videos that produce themselves every morning at 7am, and for most of this year her brain. Not a server rack. A prebuilt gaming PC sitting in a home office. Two things live elsewhere: the narration in our videos is ElevenLabs, a voice we designed once, and the brain behind her live face now runs on a second PC in the house.
This is that machine, and an honest breakdown of what matters.
Update, 30 September 2026. The build hasn’t changed; what runs on it has. Aillex’s live face is now SoulX-FlashHead Pro, about 8 GB on this card, shown on the case’s side touchscreen. The brain behind it moved to a second PC in the house (two 2016 graphics cards, Qwen 3.5 9B). On the live face her voice comes from ElevenLabs temporarily, because the local voice was too slow sharing the card with her face; her other apps still speak locally. The VRAM advice below still stands. The full story.
The exact build
An iBUYPOWER prebuilt, the RDY Y70 TI B05:
| Part | Spec |
|---|---|
| GPU | GeForce RTX 5090, 32 GB GDDR7 |
| CPU | Intel Core i9-14900KF (24 cores) |
| RAM | 64 GB DDR5-6000 |
| Storage | 4 TB NVMe Gen4 SSD |
| PSU | 1000 W Corsair RM1000e (Gold, fully modular) |
| Board | MSI PRO Z790-P WiFi |
| Cooling | 360 mm AIO liquid cooler |
| Case | HYTE Y70 Touch Infinite |
| OS | Windows 11 Home (+ WSL2 Ubuntu) |
What actually matters for local AI (ranked)
1. VRAM is the whole game, 32 GB of it. Every capability on this site is ultimately a VRAM budget line: the 26B brain (~17.6 GB), speech recognition (~1.5 GB), real-time lip-sync (~8.5 GB). They run simultaneously because 32 GB gives room to co-resident models. With 16 GB you’d run a smaller brain OR the avatar, not both. If you take one thing from this article: buy VRAM, not framerate.
2. System RAM, 64 GB is the quiet hero. WSL2, model conversion, video pipelines, a dozen Python environments, we routinely sit above 40 GB. 32 GB works; 64 GB removes a whole class of mystery failures.
3. Storage, 4 TB fills faster than you think. Models are huge: our LoRA library, three video generators, TTS models and checkpoints currently occupy well over a terabyte. Gen4 speed matters less than capacity, models load once and stay warm.
4. The 1000 W PSU earns its keep. A 5090 under sustained video-generation load is not a gaming duty cycle. Generation runs hammer the card for 40 to 60 minutes flat. Headroom = stability.
5. CPU matters less than you’d expect, with one exception. Inference lives on the GPU. But our text-to-speech runs entirely on CPU (that story here) precisely to keep VRAM free, and the 24-core i9 doesn’t blink at it.
Honest notes on prebuilt vs self-built
We didn’t plan an “AI workstation.” This was bought as a gaming PC, the AI obsession came after. That’s actually the point: the machine you may already own (or can order assembled) is enough. No used datacenter cards, no Threadripper tax. One warranty, one box, GPU arrives packaged separately (install it yourself, it’s four screws and a cable).
Could you part-pick the same spec cheaper? Somewhat: Newegg (referral link, disclosed) is where we’d do it. Would we trade the 3-year warranty and the working-out-of-the-box on a machine that now runs a 7am production pipeline unattended? No.
Considering a build? Our referral link is here, it supports the channel at no cost to you.
What we’d change (2026 edition)
- More VRAM, always, if a 48 GB consumer card existed at sane pricing, we’d own it. Watch this space.
- Second SSD from day one. You will fill the first one.
- Everything else? This box has rendered thousands of images, hundreds of video clips, and an entire self-producing YouTube channel without a single hardware complaint.
What it runs: the full local AI companion architecture →
This article contains referral links, we only link what we actually use. See our disclosure.