YouTuber Builds Enterprise Storage Server for $2,300
A DIY enthusiast built a custom E3.S storage server for 80% less than commercial options, hitting 91 GB/s speeds with consumer parts and 3D-printed components.
Written by AI. Mike Sullivan
April 23, 2026

Photo: Raid Owl / YouTube
There's a particular type of hardware that enterprise IT departments buy and home users don't: E3.S NVMe drives. They're the next evolution past U.3 drives, designed for data centers where density and power delivery matter more than whether a regular human can actually use them. Lenovo's cheapest E3.S server starts at $11,000. YouTuber Raid Owl built one for $2,300.
This is either impressively scrappy or worryingly janky, depending on your tolerance for 3D-printed server chassis held together with magnets. Probably both.
The Enterprise Tax
E3.S drives exist because PCIe Gen 5 and Gen 6 need more power than older form factors can deliver. They're smaller than traditional 2.5-inch drives but pack considerably more performance—the Kioxia CD8P-R drives Raid Owl tested are rated for 12 GB/s sequential reads and 2 million IOPS. That's legitimately fast. The drives also come with enterprise features like power loss protection and warranties that assume you'll write the entire drive capacity every day for a year.
The problem is the same problem that's existed since Sun Microsystems was selling pizza-box servers: enterprise gear requires enterprise infrastructure. You can't just plug E3.S drives into a gaming motherboard. The connectors are different. The backplanes are specialized. The entire ecosystem assumes you're buying complete systems from vendors who charge accordingly.
Raid Owl's build attempts to bridge that gap using parts you can actually buy. The core component is a Supermicro backplane (BPN-E3S5-126ESN) pulled from eBay for $240. This thing converts eight E3.S drives into eight MCIO connectors, which are at least somewhat standardized. From there, an ASRock Rack motherboard with seven MCIO ports handles connectivity, paired with an AMD EPYC 8024P processor that provides 96 lanes of PCIe Gen 5.
The total parts cost was around $2,300, not counting the 3D-printed enclosure or the drives themselves. Compare that to Lenovo's $11,000 entry point. The savings are real. So are the tradeoffs.
The Cooling Problem
Here's what enterprise server manufacturers have figured out and home builders keep rediscovering: cramming high-performance components into tight spaces creates heat problems that money can't fully solve. E3.S drives are designed for data center racks with precisely controlled airflow. The drives sit 2-3 millimeters apart. You can't fit traditional drive caddies in that spacing.
Raid Owl's solution involved multiple iterations. Initial attempts with 40mm fans behind the drives failed—the cables created too much clearance. The final design uses a 3D-printed box with two 40mm fans that magnetically attaches to the front of the drive array, creating suction that pulls air through. "It may not be super elegant, but it works well," he notes in the video. Translation: the temperatures stabilized but this thing sounds like a small aircraft.
This is the part where enterprise hardware reveals its actual value proposition. Commercial servers don't just work—they work quietly, reliably, with redundancy and serviceability baked in. A magnetically attached fan array is charming in a YouTube video. In production, it's a liability.
Performance Reality
The benchmark results tell two stories. First, the drives work exactly as advertised. Raid Owl configured eight drives in RAID 0 and hit 91 GB/s sequential read speeds. That's 728 gigabits per second—genuinely absurd throughput for a DIY build.
Second story: you can't actually use most of that bandwidth. Network limitations immediately become the bottleneck. Even with a 100-gigabit network card (already exotic hardware for most users), the practical throughput topped out around 56 Gbps using NVMe over Fabrics with RDMA. The creator suspects power delivery issues with the DGX Spark he was testing against, plus some kernel-level problems with the latest Linux build.
This gap between theoretical and practical performance is worth dwelling on. Those E3.S drives can move data at 728 Gbps, but your network can't receive it. Your CPU can't always process the random I/O fast enough—during testing, the EPYC chip hit 100% utilization trying to keep up with MDADM handling the RAID array. As Raid Owl observes: "It makes me understand why so many of these crazy storage servers come in dual CPU setups."
Enterprise gear isn't just about raw specs. It's about balanced systems where every component can actually handle what the others can deliver. Building from consumer and eBay parts means those balancing acts become your problem.
The Trickle-Down Question
The video's implicit argument is that E3.S drives might eventually become accessible to home users the way previous enterprise technologies did. SAS drives, server-grade RAM, even NVMe itself all started as data center exclusive and gradually filtered down.
Maybe. The AI boom is creating huge demand for fast, dense storage. More demand could mean more production volume, which historically drives prices down. But AI workloads also mean enterprise buyers can justify premium pricing longer. If hyperscalers will pay top dollar, why optimize for the home lab market?
What's actually interesting here isn't whether E3.S drives will get cheap. It's what happens to the secondary market for enterprise gear as data centers refresh their hardware. That Supermicro backplane cost $240 used. The ASRock board was under $500 new because it targets a weird middle ground between consumer and enterprise. This build exists because there's enough volume in professional markets to create peripheral components at almost-reasonable prices.
Raid Owl spent $2,300 building something that performs like a $11,000 server, except for all the ways it doesn't. No redundant power supplies. No proper chassis. Cooling that required multiple revisions and still sounds like a jet engine. No vendor support. But 91 GB/s read speeds are 91 GB/s read speeds, whether Lenovo's logo is on the box or not.
For YouTubers and home lab enthusiasts, that's a fascinating proof of concept. For anyone running actual production workloads, it's a reminder that enterprise pricing isn't just markup—it's insurance against all the things that can go wrong when you're held together with magnets and optimism.
— Mike Sullivan, Technology Correspondent
We Watch Tech YouTube So You Don't Have To
Get the week's best tech insights, summarized and delivered to your inbox. No fluff, no spam.
Watch the Original Video
I built a FAST storage server
Raid Owl
14m 46sAbout This Source
Raid Owl
Raid Owl, operated by tech enthusiast Brett, is a YouTube channel dedicated to diving deep into the intricacies of home labs, networking, and PC builds. Launched in August 2025, the channel has quickly grown to 150,000 subscribers, providing a rich trove of information for tech enthusiasts. Brett's channel is known for its in-depth exploration of consumer electronics and AI hardware, appealing to both hobbyists and professionals.
Read full source profileMore Like This
AI's Impact on Coding Skills: A 17% Decline?
Anthropic's study reveals AI hinders coding mastery by 17%. Explore the implications on skill development.
AI's Two Paths: Safety First or Fast Deployment?
Exploring Altman and Amodei's divergent AI safety strategies.
Lean Tech: Reviving Old School Principles
Exploring lean and agile in tech: a nostalgic journey with a modern twist.
Unreal Engine 5 Still Doesn't Play Nice With Apple Silicon
While most 3D software runs smoothly on M-series Macs, Unreal Engine 5 remains frustratingly unreliable. One creator documents the disconnect.
Claude's Constitution: AI Ethics or 90s Sci-Fi Plot?
Explore Claude's AI Constitution: a guiding doc or a 90s sci-fi plot? We dive into the ethics and implications.
Effect-Oriented Programming: A New Hope?
Exploring Effect-Oriented Programming: simplifying complexity and enhancing reliability in software development.
RAG·vector embedding
2026-04-23This article is indexed as a 1536-dimensional vector for semantic retrieval. Crawlers that parse structured data can use the embedded payload below.