Barebone vs complete GPU server: what the package contains, levels L6 to L12 and warranty

Barebone or complete GPU server: the difference

A barebone GPU server is a chassis with motherboard, power supplies, cooling and backplane, sold without processors, memory, drives and GPU cards. A complete server has all of these installed, configured and tested. Gigabyte's glossary of server manufacturing levels names the two stages L6 and L10 and adds L11 for a rack of servers and L12 for a cluster. Which form is offered depends on the maker and the model: Gigabyte gives a barebone part number for its GPU servers, HPE and Dell configure their servers to order, and Supermicro marks some HGX B300 systems "(onsite service required)" or "(complete rack and onsite service are required)". We deliver complete servers built to order, assembled and tested before delivery, on one EU contract.

The difference decides three things a buyer has to settle before the order: who installs and tests the parts, whose warranty covers each part, and what the quote has to list. The makers' own wording, read on 12 November 2026, answers most of it.

What a barebone GPU server package contains

Gigabyte's glossary describes the barebone stage in one sentence: "L6 generally refers to the server barebones stage, where the motherboard, chassis, power supply, cooling components, backplane, and basic I/O are integrated into a hardware platform." It continues: "At this stage, the system may not yet include a complete set of CPUs, GPUs, memory, storage devices, or expansion cards."

Gigabyte's product pages show what that means for a GPU server. Each page carries a part number line and a list headed "Packaging Content".

Gigabyte product pages for the G494-ZB4-AAP2 and the G894-SD3-AAX7, © 2026, read 12 November 2026. Processors, memory, drives and GPU cards are not in either packaging list.

Three details on the G494-ZB4-AAP2 page matter to anyone who plans to populate the chassis. The power cables for eight GPUs are in the box, the mains cords are not: "Power cords are not included with server packages." The fan kit that Gigabyte names for 600 W GPUs is an option with its own part number, as is the TPM 2.0 kit. And Gigabyte limits what it has validated: "The system is only validated with a uniform GPU model." For the HGX model the part number line says that the NVIDIA module belongs to the order; the page does not explain further.

ASUS documents its GPU servers on the same basis. Its datasheet for the ESC8000-E12 gives the weight "excluding CPU, DRAM & HDD" and marks the power supply row "(following different configuration by region)". The figures of the three makers' eight-card chassis are side by side in our comparison of Supermicro, Gigabyte and ASUS 8-GPU servers.

L6, L10, L11 and L12: the integration levels

Level numbers describe the degree of integration. Gigabyte presents them as a convention in use: "In server manufacturing and data center system integration, some ODMs, OEMs, and system integrators use Levels 1 through 12 to describe different levels of manufacturing and integration." And: "The exact scope of each level may vary depending on the manufacturer, product design, and project requirements."

GIGABYTE glossary, "What Are Server Manufacturing Levels? Understanding L6, L10, L11, and L12 Integration", © 2026, read 12 November 2026. The page defines these four levels and no others.

Gigabyte leaves the order of work between the barebone and the complete server to the supplier: "Because manufacturers may define the intermediate levels differently, the exact installation sequence and testing scope should be determined by the supplier or project specifications." A level number in an offer is therefore a starting point for a question: what exactly is installed, and what exactly is tested.

Supermicro uses the two upper levels in its rack brochure (code MKT-02-RackScale-2025-R19). It describes L11 as design, assembly, configuration and testing of each server at node and rack level, with all nodes tested together, and writes about the next step: "The manufacturing process at L12 involves networking from the rack to multi-rack level, then loading the customer's application software for complete, real-world testing at rack scale." Supermicro's pages do not use the labels L6 or L10.

Complete systems, complete racks and onsite service: the makers' wording

People search for the phrase "Supermicro complete system only". On the Supermicro product pages read for this article on 12 November 2026, and on Supermicro's GPU systems overview, those words do not appear. What the pages do say differs by model.

Supermicro product pages (undated) and HPE Compute XD690 QuickSpecs, a50009220enw, version 7, 2 November 2026; all read 12 November 2026. The Supermicro pages do not explain what "(onsite service required)" obliges the buyer to order.

On the pages read, the notes appear on HGX systems only: onsite service on the air-cooled 8U Intel model, a complete rack on the liquid-cooled 4U models. The 8U AMD model and the PCIe servers for eight cards carry none. HPE adds for the XD690: "Orders without confirmation of installation capability may be subject to configuration restrictions." What these service conditions mean in practice is the subject of our article on on-site service for GPU servers; the Supermicro 8U system is described in Supermicro SYS-822GS-NB3RT: HGX B300 server specs.

HPE and Dell describe configured systems throughout. HPE's QuickSpecs for the XD690 speak of "all CTO models", of "a factory-integrated model" and of options marked FIO, which "indicates that this option is available only as a factory-installable option". Dell's product pages for the XE7745 and XE9780 have a section "Tech Specs & Customization" and the entry "Contact Sales", and no barebone offer. NVIDIA's DGX B300 is a complete NVIDIA system whose service manual says: "When replacing a component, use only the replacement supplied to you by NVIDIA."

We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project, as complete servers built to order; send us the model or the cards you have in mind and we reply with a configuration and quote within one business day.

Warranty on a barebone build and on a complete server

A maker's warranty is written for the maker's own products. Supermicro's Limited Warranty, an undated page read on 12 November 2026, defines it this way: "'Product' is defined as a product developed and manufactured solely by Supermicro and does not include products, components, or software supplied by third parties." Its table has a row for GPU systems and a row for stand-alone components, and no row for barebones.

"Supermicro Limited Warranty", table in section 3, undated, read 12 November 2026. Stand-alone components are motherboard, power supply, add-on card and chassis.

The same page lists what voids the warranty, among other things "repair work performed or damage caused by a party other than Supermicro". Who warrants GPU cards that an integrator fits into a Supermicro chassis is not stated there.

Gigabyte's enterprise warranty page, read the same day, grants "a three (3) year limited warranty" on servers and barebone systems alike and starts the period at the invoice date unless agreed otherwise in writing. For claims it sends the buyer first to the distributor or system integrator from whom the product was purchased. HPE writes for the XD690: "Hardware diagnostic support and repair are available for three years from the date of purchase." For certified systems NVIDIA states that "Hardware support for NVIDIA-Certified Systems is available from the system vendors", as our article on what the NVIDIA-Certified label covers explains.

For a buyer this means that a barebone build has as many warranties as it has makers: chassis, processors, memory, drives and cards each follow their own terms. The quote should name each of them, and who answers for the assembly. Terms and periods in the EU are compared in GPU server warranty in the EU: OEM vs integrator.

Who tests the server

For a complete server Gigabyte's glossary describes tests at L10: "After hardware assembly, the system undergoes server-level functionality, compatibility, and stability testing." It names what this "may include", from boot validation and component detection to stress testing of processors, GPUs, memory and the power system. A barebone leaves the maker before processors, memory and cards are fitted, so the test of the final configuration falls to whoever assembles it.

Two other kinds of test are often confused with this one. Certification is a test of a model with named GPUs in NVIDIA's programme and is no test of the individual server that is delivered. An acceptance test is the buyer's own check after delivery, with the buyer's models and load; our guide to GPU server acceptance testing describes one.

What the quote lists in each case

• Barebone. The chassis part number and the options the maker lists separately, such as a fan kit for 600 W cards, a TPM kit and power cords. Processors, memory, drives, GPU cards and network cards are separate lines, possibly from separate suppliers. Ask who assembles, who tests and who takes a warranty case.
• Complete server. Every component of the configuration as a line, the tests before delivery, the warranty per component and the party that handles it, and onsite service where the maker requires it.
• Rack or cluster. In addition the rack, power distribution units, switches, cabling and, for liquid cooling, the rack-level parts, with the level of integration spelled out in words.

A request that names the cards exactly makes all three easier, as our article on NVIDIA GPU part numbers and what to state in a quote shows; the remaining lines are in what to specify before buying an AI server.

How we deliver a GPU server

We deliver complete servers, built to order. The steps are those on our AI servers page. You tell us the model, the users, the data and the deadline, and we propose a configuration free of charge. The quote follows within one business day and states price, lead time and warranty terms in writing, with EU invoicing from Austria. The server is assembled and tested before delivery, with a power-on and basic load test. It is delivered anywhere in the EU on one contract, with manufacturer warranty on every component; DOA units are replaced, the assembly itself is covered by us, and commissioning on site is available on request. An operating system, drivers, CUDA and a container runtime are installed on request.

If you already own a chassis or parts, we build on them after a compatibility check of the platform, power and cooling, and we say in advance if something will not work. We sell new equipment only, as new vs refurbished GPU servers explains. Other ways to order in Europe are compared in where to buy an AI server in Europe, and a public buyer finds the wording for these points in our GPU server tender specification.

To start, write to us through the form or by e-mail to request@eurokommerz.at with the workload, the cards and the rack position.

FAQ