On-site service for GPU servers: what next business day and 4-hour mean at each maker

What on-site service for a GPU server covers

On-site service for a GPU server comes from the server maker: on some models as part of the warranty, otherwise as a service contract with a named response level such as next business day or four hours. In the documents of Dell and HPE that time runs until a technician arrives, and it starts once the fault has been diagnosed as a hardware fault. NVIDIA's standard support for DGX systems counts the shipment of the replacement part instead. Dell's service description of July 2026 also names GPUs among the parts that are deemed non-essential or are not stocked in its four-hour parts locations.

This guide sets out what Supermicro, Dell, HPE, ASUS, Gigabyte and NVIDIA publish about on-site service, as read on 12 November 2026. Warranty periods are a subject of their own, covered in our guide to GPU server warranty in the EU. Every response time below is a term of a maker's contract. We quote these terms and give no response-time promise of our own.

Supermicro, Dell ProSupport, HPE Tech Care and NVIDIA DGX service levels

Supermicro "Onsite Services" page (no date); Dell "ProSupport Infrastructure Suite Datasheet" (April 2025), row "Onsite parts and labor response objective"; HPE "HPE Tech Care Service" data sheet a00108652ENW, Rev. 7 (2026); NVIDIA "Services for DGX Products – Terms and Conditions" (24 October 2023); ASUS "RMA Service" page for servers; GIGABYTE "Enterprise Products Warranty Information" (2026). All read 12 November 2026.

Supermicro sends field engineers at every level "with the exception of Bronze Level", so the lowest level is listed under on-site services and comes without a technician. A footnote in Dell's datasheet says that "Basic Hardware Support extensions and upgrades are not available in Germany" and that availability and terms vary by region and by product. Of HPE Complete Care we read only the page-level description, so its response levels are left out here. Which cards the two makers' GPU servers take is covered in our articles on Dell PowerEdge XE servers and HPE ProLiant servers for AI.

NVIDIA's DGX page says the platform "includes enterprise support", and the terms and conditions of 24 October 2023 require that a support service is purchased together with the system. At the standard level the replacement part is shipped; a technician for field-replaceable units may be sent "subject to the availability of technicians". ASUS names an Onsite Service and an Advance Replacement Service for servers and publishes no hours or days for either. Gigabyte's warranty page puts on-site service among the extended services that "require a separate quotation and a formal written agreement".

What next business day and four hours measure

The name of a level states a time. The service descriptions state when that time starts and what ends it.

Dell service description "ProSupport for Infrastructure" v17 (July 2026); HPE Tech Care data sheet a00108652ENW, Rev. 7; NVIDIA "Services for DGX Products – Terms and Conditions" (24 October 2023); Supermicro datasheet "Global Services" (March 2026). All read 12 November 2026.

Dell writes both definitions with the word "typically". For the higher level: "A technician typically arrives on-site within 4 hours after completion of troubleshooting and isolating the issue." A call that comes in after the local cut-off time "may require an additional business day" at the lower level, and the four-hour level exists only "within defined four (4) hour response locations". At Dell and at HPE remote diagnosis comes first, and the time it takes lies outside the four hours.

NVIDIA's standard DGX support counts the shipment of the part. Its Enterprise Support and Services User Guide adds in a footnote: "Onsite Engineer is dispatched once replacement part is in possession of the customer." Supermicro states the start of its clock and, on the pages we read, gives no definition of what a four-hour or next-business-day label ends with.

A deadline for the repair itself exists only as a separate, higher level. HPE's Tech Care Critical carries the call-to-repair commitment: "For severity 1 and 2 incidents, HPE returns the covered hardware to operating condition within six hours." The six hours hold within 50 miles (80 km) of an HPE support hub; between 51 and 100 miles HPE states eight hours, and beyond 100 miles the commitment is not available. Dell's counterpart is the Mission Critical option of ProSupport Plus, a four-hour on-site response with a six-hour hardware repair service. In Dell's wording the technician "often repairs the hardware within 6 hours of dispatch", and the option is offered "within a 50 mile or 80 kilometer distance" of a Dell support hub and "Only for Equipment connected to the secure connect gateway."

We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project; send us the address of the site and the maker's response level your operation needs and we reply with a written quote within one business day.

What Dell ProSupport says about GPUs and four-hour parts

For a GPU server one sentence in Dell's service description weighs more than the names of the levels. ProSupport for Infrastructure (v17, July 2026) says: "Parts deemed non-essential and/or not stocked in four-hour (4) parts locations include, but are not limited to: bezels, mechanical chassis, hard drive blanks, rail kits, cable management accessories and Graphical Processing Units (GPUs)." The same sentence stands in the description of ProSupport Plus for Infrastructure (v16.5, July 2026). As parts that "may be deemed essential" Dell lists "motherboards, CPUs, select memory modules and hard disk drive that may impact array and production environment".

The four-hour objective concerns the technician's arrival. Whether the part the technician needs lies in a four-hour location depends on the part, and Dell puts GPUs on the list of parts that are deemed non-essential or are not stocked there. The document gives no time for a replacement GPU, so that question belongs in the request for the service quote.

From HPE we read one line about GPUs, in the services list of a Gen12 QuickSpecs document: under HPE Comprehensive Defective Material Retention it says "Excluded: GPUs and superchips.", so that retention option does not extend to a failed GPU.

A node that must not stand idle while a card is on its way therefore needs a spare card on site or a second node that carries the load. NVIDIA offers on-site spares for DGX systems as an option and writes: "The customer will manage the onsite spares." How two nodes share a service is the subject of our guide to high availability for LLM inference with two nodes.

Which parts the customer fits and which the technician fits

Makers divide parts into those a customer may replace and those a technician replaces. Dell's description says: "Customer has option to perform installation of Customer Replaceable Units (CRUs)." Which parts those are in a PowerEdge server is not listed in Dell's warranty table of 29 May 2026: "CRUs for Server Products vary by region and Product type."

HPE calls the same thing customer self-repair and ties a condition to it that touches the response level: "If Customers agree to the recommended CSR and a CSR part is provided to return the system to operating condition, the on-site service level shall not apply."

NVIDIA lists the customer-replaceable components of each DGX system in its service manual.

NVIDIA "DGX B300 Service Manual" (last updated 30 April 2026), "DGX B200 Service Manual" and "DGX H100/H200 Service Manual" (both last updated 13 September 2026), Introduction pages, read 12 November 2026.

No GPU, GPU tray or baseboard appears in any of the three lists, and the chapters of the B300 manual contain no procedure for a GPU tray. The manuals do not forbid the work in so many words. They say: "Contact NVIDIA Enterprise Support for replacement instructions and guidance for specific components if those instructions are not included in this document." On-site labour for customer-replaceable units exists as an option of the Business Critical level; in the Business Standard column of NVIDIA's support table that row reads "Not Available". Supermicro publishes no division into customer and field parts on the pages we read.

What the maker asks of the site and the customer

• Remote diagnostics. Dell requires the customer to run its "remote IT support and monitoring software (implemented as a secure connect gateway)" for the whole term. HPE customers "will be required to support HPE remote resolution efforts" and to "Provide remote connectivity through an approved communications line". NVIDIA asks for "embedded diagnostic information associated with the Product" and for access over the internet or a VPN connection. A server in a closed network needs this point cleared with the security team before the contract is signed.
• Access. Dell expects "free, safe and sufficient access to Customer's facilities and environment, including ample working space, electricity, safety equipment (if applicable)". NVIDIA's DGX service manuals add a rule for the room: "Hearing protection is mandatory when working with or in the vicinity of this equipment". None of the documents we read asks for a spare power feed.
• Backup. The data stays the customer's job. HPE asks for "a separate backup system or procedure that is not dependent on the products under support".
• Registration. At HPE the customer "must activate the hardware product to be supported within 10 days of purchase of this service". NVIDIA: "You are responsible for registering the Product to receive these Services".
• Distance. HPE's response times "apply only to sites located within 100 miles (160 km) of an HPE designated support hub". Dell defines its "Dell Service Area" as a location within 160 drivable kilometres of a Dell service location and in the same country; for EMEA the same document says that "on-site service is available within a distance of up to 150 kilometers from nearest Dell Logistics location".
• Moves. After a move Dell "must first re-qualify Your Product" before the response times apply again, and NVIDIA must be notified when a DGX system is relocated to a new site.

What the room itself must offer, from rack depth to power feeds, is the subject of our site survey for GPU servers.

Supermicro on-site service in Europe

Supermicro's global page lists four levels. Its own European eStore, on the page "Service offering", sells two: "Bronze on-site service" with a response SLA of "3 Business Days" and without field engineers, and "Silver on-site service" with "Next Business Day" and with field engineers. For both, the help desk works "Monday - Friday, 9am-5pm local (Excludes holidays)". Gold and Mission Critical are not on that eStore page, which says nothing about whether they are sold in Europe by another route.

Supermicro eStore Europe, page "Service offering" (no date), read 12 November 2026.

One note on the eStore page reads "Please note: On-site service overrides Warranty."; the page does not explain the sentence, and we quote it without reading more into it.

Eligibility is stated in Supermicro's datasheet "Global Services" of March 2026: "your purchase must consist of a complete Supermicro server with CPU, memory, and hard drives all provided by Supermicro and integrated at Supermicro." A server that an integrator builds on a Supermicro barebone does not meet that sentence as worded. Our article on barebone and complete GPU servers explains the two ways these systems are sold.

Five decisions before a service contract is signed

1. Arrival or repair. Decide whether a technician's arrival within a stated time is enough or whether the operation needs a repair deadline. The second is a different level with a shorter permitted distance to the support hub.
2. The GPU. Ask in writing where a replacement card comes from and in what time, and plan a spare card or a second node if the answer is too slow for the service.
3. Who touches the hardware. Clarify which parts count as customer-replaceable. At HPE, agreeing to a self-repair part sets the on-site level aside for that incident.
4. Remote access and location. Check that the maker's remote tool is permitted in your network, measure the distance to the support hub and check the country coverage.
5. Data carriers. If drives may not leave the building, order a retention service with the system. Supermicro names "Digital Media Retention" for disks and a "Parts Retention Program" for other components.

Our own part is stated on our pages for AI servers and GPUs and goes no further than that: manufacturer warranty on every component, handled through us; units that arrive dead are replaced; the assembly of a server we build is covered by us; commissioning on site is available on request. A maker's response level is tied to that maker's system and its service contract, as Supermicro's eligibility sentence shows. If you are weighing such a system against a server built to order, describe the site, the workload and how long the service may be down, and the written quote states the warranty terms for each variant.

The other points of a server request are in our guide on what to specify when buying an AI server, and the checks on delivery in our guide to GPU server acceptance testing. How chassis height affects access to the cards is covered in 4U and 8U GPU servers compared.

FAQ