This document updates and narrows the personal server procurement target while preserving the five-year CareTV reference architecture.
Immediate predecessor:
https://8k.art/beta/_daily-builds/2026-08-17/caretv-reference-server-requirements-updated-owned-hardware-and-five-year-buying-rules-2026-08-17-v001.html
Canonical CareTV architecture:
https://8k.art/beta/_daily-builds/2026-08-15/caretv-canonical-spec-2026-08-15-v009.html
The personal purchase is intentionally less expensive than the eventual primary CareTV compute server. It is expected to remain useful later as a storage, index/database, surveillance-recording, transcoding, or secondary-node server.
Target: approximately $2,000 USD for the immediate server platform. A lower-cost system is preferred when it meets all hard requirements. The long-run expectation is that substantially more compute will become available for less money; therefore today's purchase should emphasize durable connectivity, expansion, serviceability, and rack infrastructure rather than maximum bundled GPU performance.
| Area | Hard requirement |
|---|---|
| Chassis | 4U rackmount. Standard 19-inch rack compatibility. No tower substitution. |
| Serviceability | RAM and primary M.2 NVMe storage must be accessible without motherboard removal or near-complete chassis disassembly. Side/top access should be straightforward. A competent technician should be able to perform ordinary RAM/NVMe/NIC/GPU service rapidly. |
| CPU / board | Current-generation or near-current high-performance platform, delivered with CPU and motherboard already installed. The owner is not required to install the CPU. |
| GPU | GPU-upgradable. The system must physically and electrically support a full-size discrete GPU. It should be possible to reuse an existing RTX 4070 Ti when desired. The immediate purchase does not need to include an expensive GPU. |
| Memory | DDR5 preferred/current platform; easy user access. 64GB can be acceptable initially if expansion to at least 128GB is straightforward. Avoid paying a premium for bundled memory when owned compatible RAM can be reused. |
| NVMe | Accessible M.2 NVMe slots. Must support reuse of owned 8TB WD_BLACK SN850X PCIe Gen4 x4 NVMe and 8TB Lexar NM790 PCIe Gen4 x4 NVMe where platform compatibility permits. |
| Thunderbolt | Thunderbolt 5 strongly preferred. Native/supported motherboard implementation, not dependent on a proprietary single-vendor add-in card or disappearing accessory. Thunderbolt 4 may be considered only if it materially improves price and all other hard requirements are satisfied. Existing OWC ThunderBay 8 storage is a 40Gb/s-class Thunderbolt 3 device and does not itself require Thunderbolt 5 bandwidth. |
| Fiber networking | Immediate PCIe capacity for a dual-port 100GbE QSFP28 fiber NIC. Slot must be physically usable and electrically appropriate after accounting for any installed GPU and motherboard lane sharing. |
| Copper networking | At least one ordinary RJ-45 copper Ethernet port. 10GbE RJ-45 copper Ethernet is preferred where practical; 2.5GbE RJ-45 copper Ethernet is acceptable as secondary/local connectivity if 100GbE QSFP28 fiber is present. |
| Power | Standard, replaceable, adequately sized PSU with headroom for future GPU and 100GbE NIC. Avoid proprietary PSU lock-in where practical. |
| OS | Windows 11 Pro supported. For Thunderbolt 5 platforms, Windows 11 support must be documented by the motherboard/system vendor. |
| Recovery | Documented BIOS/UEFI ability to restore power automatically after AC loss. |
Disqualifying rule: no essential CareTV capability may depend on a proprietary single-vendor expansion card, private header, proprietary cable, mandatory vendor cloud, or discontinued accessory for basic operation.
If removing one vendor from the market makes a required CareTV capability impossible to repair or reproduce, that platform is disqualified from future CareTV reference procurement.
The existing ASRock X570 Taichi workstation is a useful negative reference. Its practical Thunderbolt path depended on ASRock's proprietary Thunderbolt add-in-card ecosystem. Once the required card became discontinued/unavailable, the workstation ceased to be a suitable future CareTV reference platform even though the computer itself remained otherwise powerful.
Future CareTV systems should therefore prefer native, documented, standards-based interfaces that can be serviced and reproduced by multiple vendors.
Stage 1 — now: approximately $2,000 4U rackmount infrastructure/server platform that immediately gets CareTV storage, Thunderbolt JBODs, 100GbE QSFP28 fiber, databases/indexing, surveillance ingest, and local media services online.
Stage 2 — later: approximately $6,000 primary AI/compute server when next-generation CPU/GPU hardware offers materially more performance per dollar. The Stage-1 machine remains a permanent useful node rather than being discarded.
A candidate is not approved for purchase until the exact SKU's motherboard, Thunderbolt implementation, PCIe slot electrical widths/lane-sharing, RAM access, M.2 access, PSU, GPU clearance, RJ-45 copper Ethernet speed, Windows 11 Pro status, and AC-recovery behavior have been verified.
CareTV / Open Access Media Lab — open-platform hardware procurement requirements — 2026-08-17-v002