Y Custom · built to order
Built for purpose.
When a stock chassis isn't the right shape, we build the machine around the work. Three tiers, each assembled, burned in and benchmarked as one system with Y OS — each with one advantage the others can't match.
Personal128GBY Computer Mini 128
The most model per dollar and watt.
One quiet desk box, 128GB of unified memory, and serious local models for a single power user. The best first Y Computer.
- Processor
- AMD Ryzen AI Max+ 395
- AI memory
- 128GB unified memory
- Accelerator
- Radeon 8060S integrated GPU
- Storage
- 2TB NVMe
- Format
- Compact, single-box desktop
A good fit for
- 35B-class coding and multimodal models
- 120B sparse reasoning models
- Private chat, documents, voice and one agent loop
Optimized for one primary user; long context and concurrent jobs share the same memory pool.
$5,999
complete system · Y OS installation included · tax and shipping extra
Team256GBY Computer Pro 64
Serves the whole studio at once.
Two high-memory accelerators tuned for throughput — several assistants answering in parallel without a queue forming.
- Processor
- AMD Ryzen Threadripper 9960X
- AI memory
- 256GB ECC system memory
- Accelerator
- 2× Radeon AI PRO R9700 · 64GB aggregate VRAM
- Storage
- 4TB NVMe
- Format
- Tower workstation · dedicated circuit recommended
A good fit for
- Several simultaneous local assistants
- 70B-class quantized models across both GPUs
- Parallel chat, code, transcription and image services
The 64GB figure is aggregate capacity. Models must be explicitly sharded; two GPUs do not automatically mean twice the speed.
$11,999
complete system · Y OS installation included · tax and shipping extra
Maximum compatibility96GBY Computer Max 96
96GB on one GPU. Nothing translated.
The cleanest single-GPU path to high-quality local inference, with the broadest application support in the line.
- Processor
- AMD Ryzen Threadripper 9960X
- AI memory
- 256GB ECC system memory
- Accelerator
- NVIDIA RTX PRO 6000 Blackwell Max-Q · 96GB ECC
- Storage
- 4TB NVMe
- Format
- Tower workstation · dedicated 120V/20A circuit recommended
A good fit for
- 120B sparse models on one GPU
- High-fidelity 70B-class models
- CUDA-first agents, serving and creative applications
For models larger than 96GB or heavier concurrency, a dual-GPU Max 192 is quoted after workload validation.
$24,999
complete system · Y OS installation included · tax and shipping extra
Limited edition · QM Edition · founding run of 25
The multiplayer agent machine.
QM is the open-source agent harness Y Combinator's software team built for its own company and released under MIT — one agent floor for a whole team, in Slack and on the web. Deploying it normally takes a cloud account, a Postgres database and a platform engineer. This edition is a Pro 64 with all of that done: QM and Postgres on the box, agents running against local models through the Y OS API, Slack and web wired up at install.
- Per-person workspaces — isolated memory, files, permissions and sandboxes
- One identity across Slack and the web UI
- Shared skills, crons and watches for background work
- Strict / Auto security postures with hard denials, enforced on your hardware
$11,999
Pro 64 hardware · QM configured · Y OS installation included · tax and shipping extra
QM is open source (MIT), released by Y Combinator's software team. Y builds and supports this machine independently — not affiliated with or endorsed by Y Combinator.
Every build arrives working.
Assembled, burned in and benchmarked as one system — then shipped with the report from your exact machine.
01
Y OS, installed
Managed runtimes, private chat, local knowledge, voice, agent controls and one authenticated local API.
02
Models for your work
We choose formats that fit the memory and duty cycle, then pin the exact model and runtime versions.
03
Measured before handover
Your report records model hash, quantization, context, runtime, power profile and single- and multi-request results.
04
A system you can recover
Encrypted storage, signed pinned releases, pre-update snapshots and a documented rollback path.
Custom builds · order path
Order by card. Facts before it ships.
Order by card at the listed system price, including Y OS, or take a written quote first. These are built-to-order machines: we source the components, build and burn in, install Y OS and benchmark — 6–8 weeks from order to handover, acknowledged at checkout. Before your machine ships you get the exact components, lead time, warranty owner, return terms, shipping and estimated sales tax in writing.
- 1
Fit call
Workload, users, data boundary and the models you actually need.
- 2
Written build
Final BOM, price, lead time, warranty and benchmark profile.
- 3
Build and handover
Burn-in, Y OS onboarding and your machine-specific report.
Questions, answered plainly
- Are these final retail prices?
- The listed founding-batch price includes the stated hardware, build, Y OS installation and configuration, selected models, burn-in and machine-specific benchmark. Shipping and state sales tax are separate. Because high-memory component supply moves quickly, we confirm the exact bill of materials and warranty owner in writing before shipping and get your approval before any material change.
- Custom versus a GB10 desk box — which do I want?
- The GB10 boxes on the home page are stock chassis that ship in 3–5 weeks; pick one when 128GB unified and a standard shape fit the job. The custom line exists for the shapes stock cannot take: maximum capacity per watt, team throughput, or a single big CUDA GPU. Custom builds are assembled to order in 6–8 weeks.
- Does all data stay local?
- Local models, document search, transcription and the local API do not require an external AI provider. Optional web search, cloud models, email and other connectors can send data outside your network; Y OS labels those surfaces as External before they are enabled.
- Which one should I buy?
- Mini 128 is the best first system for one power user. Pro 64 is for parallel services and small teams willing to use a qualified AMD software image. Max 96 is for maximum CUDA compatibility and models that need up to 96GB on one GPU.
Founding-batch prices include the stated hardware, build, Y OS installation, selected models, burn-in and machine-specific benchmark. Representative product images shown.
Custom & big builds