Owndom
Log in Sign up

Owndom Hardware Fund

Help Owndom get its own AI server

Owndom is an independent AI workspace, and this fund buys the hardware it runs on. Today everything shares one consumer graphics card. The goal is one machine built for the job: a 96-core AMD Threadripper PRO workstation with two NVIDIA H200 GPUs. It lets Owndom serve many people at once, start research on multi-agent systems, and train its own model, OmniMind, from scratch.

Contributions aren't open yet. Until then: read the plan, tell people, and follow what gets built.

Goal 1

Goal · $125,000

Threadripper PRO 9995WX + 2 × NVIDIA H200 NVL

$0 raised. The number here updates once contributions open.

  • 96 cores for the app, the agents and the data work, all at once
  • Two H200 cards: 282 GB of GPU memory joined by NVLink, enough to serve large open models and train at the same time
  • 256 GB of ECC memory, fast NVMe storage, storage redundancy, backup power and its own 240 V circuit
  • Five years of NVIDIA AI Enterprise come with the cards

What it unlocks

Serve more people. Research. Train from scratch.

  • Serve many people at once, with agents running in parallel instead of taking turns.
  • Start research on multi-agent systems: agents that hand work to each other, test their own code in sandboxes, and are measured against a fixed benchmark.
  • Train OmniMind, Owndom's own model, from scratch: a goal this hardware makes possible.

If the goal isn't reached, the order is fixed: the workstation with one H200 first, the second card when the rest arrives. Nothing is bought on credit.

Goal 2, after Goal 1

Goal 2 · price announced when it's available

An NVIDIA Vera Rubin server with 8 GPUs

The next step once Goal 1 is running: a data-center-class server with eight Vera Rubin GPUs, to train a bigger and better Owndom model.

Why it comes second

Prove it on Goal 1 first

Goal 1 serves people and trains OmniMind from scratch. Goal 2 only makes sense once that works, so contributions go to Goal 1 until it is reached; the price and plan for Goal 2 are published before anything is raised for it.

Where it stands today

Today everything shares one consumer graphics card, so it handles one heavy job at a time. That works well for a few people; it can't serve a growing community at the same time.

What already works for members: chat with AI agents, Missions, Messages with voice and direct messages, Image Studio, Design Studio, a code sandbox, web search, phone notifications, and the phone and Windows apps. Follow progress on what's new.

The build

Prices are public listings from mid-October 2026 and change quickly. The goal is this build or equivalent hardware, and parts are bought as soon as the money is there.

PartEstimated cost
2 × NVIDIA H200 NVL 141 GB (NVLink bridge)$69,000–72,000
AMD Threadripper PRO 9995WX (96 cores)$11,500–11,700
256 GB DDR5 ECC memory$7,000–9,000
WRX90 board, GPU server chassis, 2–2.4 kW power supply$3,500–7,000
2 × 4 TB NVMe, CPU cooler$1,100–1,700
240 V circuit (electrician) and a 3 kVA UPS$2,000–4,500
Backup power, storage redundancy and networking$1,100–1,700
Sales tax$6,000–6,800
Parts total$101,000–114,000
Platform fees (~3.5%) and a 5% margin for price changes→ $125,000

What I promise

  • Regular updates: what got built, and the progress number.
  • Receipts for every purchase, photos of the build, and benchmarks once it runs.
  • Honest numbers. If a price changes I say what changed; anything left over goes to the next piece of hardware on the list, and that list is public.

What supporters get

  • Your name on the founding supporters wall, only if you want it there. Names that break the Terms of use are removed.
  • Behind-the-scenes updates as the server comes together, and news about the model when there's something to show.

The fine print

  • Contributions go to Owndom's founder to buy the hardware described here. They are voluntary and are not tax-deductible.
  • A contribution is not an investment, a loan, a purchase or a share in anything, and it doesn't buy a product, a service or a vote.
  • Contributions are non-refundable, except where the law requires or if a payment was processed by mistake. See the Terms of use.
  • To stop a monthly contribution, use the link in your Stripe receipt or write to [email protected].

Help spread the word

The most useful thing right now is people knowing about it. Share this page with anyone who follows open models or wants an independent AI workspace.

Share on X Share on Reddit Hacker News

Questions

Why not just rent cloud GPUs?

Owndom runs its own models on its own hardware, so chats on those models stay there. Renting suits one-off jobs, but an always-on service needs a machine that is always there, and owning one costs less over a few years than renting the same memory around the clock.

Why H200s?

141 GB of memory each at 4.8 TB/s, joined by NVLink, in a form that fits a workstation. That is the smallest machine that can serve a large model and train another one at the same time. Two RTX PRO 6000 cards would cost less but hold a third less memory and cannot be joined; a rack of H100s needs a datacenter.

What if the goal is not reached?

The workstation and one H200 come first; the second card when the rest arrives. Nothing is bought on credit, and every purchase is shown with its receipt.

Can I help without money?

Yes: share this page and follow what's new.