Enterprise AI Server
Model Comparison
22 named server models and rack-scale systems from 5 OEMs, compared by accelerator support, form factor, and cooling. Each entry shows the OEM page it came from and the date it was checked.
Updated
This page compares specific, named AI server models: which accelerators each supports, what chassis it comes in, how it is cooled, and whether it is bought as a server or as a whole rack. It carries no price, lead time, or availability, because those change constantly and SLYD has no governed public record for them. Choosing which vendor to build a relationship with is a separate question.
Two different questions
Which system should we buy? That is this page. It compares orderable and announced models on the attributes that determine whether a system fits your workload and your facility.
Which vendor should we buy it from? That is OEM comparison. Procurement path, support model, geographic reach, and lifecycle fit are properties of the relationship, not of the chassis, and they often point somewhere different than the specification sheet does.
Accelerator specifications themselves are in the GPU database. This page does not restate them, so the two cannot disagree.
What Each OEM Publishes
Grouped by vendor. Read the Level column first: a rack-scale system and a single server are not comparable line items, and comparing them as though they were is the most common way a table like this misleads.
Dell
8 entries Detail: Dell hardware
| Model | Level | Form factor and cooling | Accelerators | Host processors | Source |
|---|---|---|---|---|---|
| PowerEdge XE9712 | Rack-scale system | 48U, ships in IR9000 rack | 72x NVIDIA B300, as GB300 NVL72 | 36 Arm-based NVIDIA Grace processors | Dell page |
| PowerEdge XE8712 | Server | 1OU, ships in IR7000 rack | 4x NVIDIA B200, as GB200 NVL4 | 2 Arm-based NVIDIA Grace processors | Dell page |
| PowerEdge XE9780 | Server | 10U | 8x NVIDIA HGX B200 or B300 | 2x Intel Xeon 6 6700P | Dell page |
| PowerEdge XE9780L | Server | 3OU | 8x NVIDIA HGX B300 | 2x Intel Xeon 6 6900P or 6700P | Dell page |
| PowerEdge XE9785 | Server | 1OU | 8x AMD Instinct MI355X | 2x AMD EPYC 9005 | Dell page |
| PowerEdge XE9680L | Server | 4U | 8x NVIDIA HGX B200 | 2x 5th Generation Intel Xeon Scalable | Dell page |
| PowerEdge XE9680 | Server | 6U | 8x NVIDIA H100 or H200, or 8x AMD Instinct MI300X, or 8x Intel Gaudi 3 | 2x 5th Generation Intel Xeon Scalable | Dell page |
| The widest accelerator choice in the Dell AI range, including a non-NVIDIA and non-AMD option. | |||||
| PowerEdge XE7745 | Server | 4U | Accelerator options not resolved from the catalog | 2x AMD EPYC 9005 series | Dell page |
| Dell positions this model for generative AI fine-tuning and inferencing. Its accelerator column could not be read unambiguously from the catalog, so it is described rather than specified. | |||||
HPE
3 entries Detail: HPE hardware
| Model | Level | Form factor and cooling | Accelerators | Host processors | Source |
|---|---|---|---|---|---|
| ProLiant Compute XD685 | Server | 5U for direct liquid cooling, 6U for air | 8x NVIDIA B300 HGX, B200, or H200, or 8x AMD Instinct MI355X | 5th Generation AMD EPYC | HPE page |
| The form factor changes with the cooling method, which is a useful signal that cooling choice is a chassis decision rather than an option. | |||||
| Compute XD690 | Server | Not stated on the cited page | 8x NVIDIA B300 HGX | Not stated on the cited page | HPE page |
| HPE describes this model as part of NVIDIA AI Computing by HPE. | |||||
| Cray XD670 | Server | 5U, single 2-CPU node, direct liquid cooling option | 8x NVIDIA H200 SXM5 | 2 CPUs in a single node | HPE page |
Supermicro
3 entries Detail: Supermicro hardware
| Model | Level | Form factor and cooling | Accelerators | Host processors | Source |
|---|---|---|---|---|---|
| Liquid-cooled HGX B300 8-GPU, 4U | Server | 4U for 19-inch EIA racks, front I/O | 8x NVIDIA HGX B300 | Not stated on the cited page | Supermicro page |
| Supermicro states its DLC-2 technology on this front-I/O 4U system captures up to 92 percent of the heat generated by server components. That figure applies to this configuration. | |||||
| Liquid-cooled HGX B300 8-GPU, 2-OU OCP | Server | 2-OU for 21-inch OCP ORV3 racks | 8x NVIDIA HGX B300 | Not stated on the cited page | Supermicro page |
| Supermicro states this OCP form factor enables up to 144 GPUs per rack. That density figure applies to this configuration in an ORV3 rack, not to the portfolio. | |||||
| NVIDIA Blackwell portfolio | Vendor portfolio | Multiple | HGX B300, HGX B200, and GB200 NVL72 solutions | Varies by system | Supermicro page |
| Listed as a portfolio because Supermicro publishes the range at this level. Individual model numbers should be confirmed against a current Supermicro configuration. | |||||
Lenovo
5 entries Detail: Lenovo hardware
| Model | Level | Form factor and cooling | Accelerators | Host processors | Source |
|---|---|---|---|---|---|
| Lenovo NVIDIA GB300 NVL72 | Rack-scale system | Rack-scale system | 72x NVIDIA B300, as GB300 NVL72 | NVIDIA Grace | Lenovo page |
| ThinkSystem SR680a V3 | Server | Not stated on the cited page | NVIDIA B200 | Not stated on the cited page | Lenovo page |
| ThinkSystem SR685a V3 | Server | Not stated on the cited page | AMD Instinct MI300X | Not stated on the cited page | Lenovo page |
| Documented by Lenovo in a published customer story rather than a specification sheet on the cited page. | |||||
| ThinkSystem SR675 V3 | Server | Not stated on the cited page | NVIDIA H100 Tensor Core and NVIDIA L40S | Not stated on the cited page | Lenovo page |
| Documented by Lenovo in a published customer story rather than a specification sheet on the cited page. | |||||
| ThinkSystem SC777 V4 Neptune | Server | Not stated on the cited page | Not stated on the cited page | Not stated on the cited page | Lenovo page |
| Published as a Neptune server, which is Lenovo's liquid-cooling technology line. | |||||
GIGABYTE
3 entries Detail: GIGABYTE hardware
| Model | Level | Form factor and cooling | Accelerators | Host processors | Source |
|---|---|---|---|---|---|
| GIGAPOD, air-cooled configuration | Rack-scale system | Scalable unit of nine air-cooled racks, using an 8U G893-series server | 8 GPUs per server | Not stated on the cited page | GIGABYTE page |
| GIGABYTE states each rack in this configuration requires support for 50 kW of power for IT hardware. GIGAPOD scales to eight racks with 32 GPU nodes, a total of 256 GPUs. | |||||
| GIGAPOD, liquid-cooled configuration | Rack-scale system | Scalable unit of five liquid-cooled racks, using a 4U G4L3-series server | 8 GPUs per server with cold plates on all eight GPUs and both CPUs, up to eight servers and 64 GPUs per rack | 2 CPUs per server, both cold-plated | GIGABYTE page |
| GIGABYTE states this configuration runs at 100 kW per rack. The higher per-rack power is the point of the liquid-cooled variant: fewer racks for the same GPU count. | |||||
| GIGAPOD AI DLC Rack, NVIDIA GB300 NVL72 | Rack-scale system | Rack-scale, direct liquid cooling | 72x NVIDIA B300, as GB300 NVL72 | NVIDIA Grace | GIGABYTE page |
What Actually Separates These Systems
In rough order of how often each one eliminates a candidate.
| Criterion | Why it decides things |
|---|---|
| Accelerator support | The narrowest constraint. Some chassis take one accelerator family only; the Dell XE9680 takes NVIDIA, AMD, or Intel parts. If you want optionality across vendors later, the chassis has to allow it now. |
| Cooling method, and what it does to the chassis | Not a configuration option. The HPE XD685 is 5U for direct liquid cooling and 6U for air, so cooling changes the rack arithmetic. Some rack-scale systems are liquid-cooled only, which means the site has to be able to reject the heat before the system is even a candidate. |
| The level you are buying at | A rack-scale system is specified, sited, powered, and commissioned as a rack. That is a different procurement, a different facility conversation, and a different failure domain from buying servers. Decide which you are doing before comparing. |
| Rack density and power per rack | GIGABYTE publishes 50 kW per rack for the air-cooled GIGAPOD and 100 kW for the liquid-cooled variant, which is the whole reason the liquid variant needs five racks where air needs nine. Density is what converts a GPU count into a floor plan. |
| Host processors and platform | Intel Xeon, AMD EPYC, and Arm-based NVIDIA Grace all appear across these models. This affects the software stack, the memory architecture, and what else the node can usefully do between accelerator jobs. |
| Rack standard | Supermicro publishes the same HGX B300 system in a 4U for 19-inch EIA racks and a 2-OU for 21-inch OCP ORV3 racks. If your facility is not OCP, half of a vendor's density story may not be available to you. |
Then check the facility, before the shortlist hardens
Server input power, rack design load, and heat rejection decide whether a system is deployable at your site, and they are the constraints most often discovered last. Work them through with the power and cooling calculator and the cooling requirements guide. The governing thermal figure is the OEM system specification for the configuration you intend to order, not the accelerator's board power.
Why There Are No Prices Here
Enterprise AI server pricing depends on configuration, quantity, channel, geography, negotiated terms, and the week you ask. A static range with no currency, configuration, date, or source is not something a buyer can act on, and SLYD has no governed public price record to publish in its place. The previous version of this page carried 77 currency figures, a price-position ranking, and a five-year cost table, none of them sourced or dated.
What is useful instead is the shape of the cost and which inputs move it. Model that against your own assumptions in the TCO calculator, which shows every assumption it uses and lets you replace the ones that are wrong for you. For a configuration-specific price, ask.
Lead time and availability are absent for the same reason. They are the most volatile facts on a page like this, they move with allocation, and a figure that is stale by a week is worse than no figure at all.
Who Actually Supports the System
Support on an AI server can come from four different places, and which one you get depends on how you buy rather than on which chassis you pick.
| Path | What it covers | What determines it |
|---|---|---|
| OEM support | The hardware, under the OEM's own service organization and entitlement | The service level you purchase with the system, and whether the OEM operates in your country |
| Distributor support | Fulfilment and first-line handling, escalating to the OEM | The channel the order goes through |
| Integrator support | The system as deployed, including the parts the OEM does not own such as the fabric or the cooling loop | Whether an integrator built the deployment and what the statement of work covers |
| SLYD | Sourcing, structuring, and coordination of a transaction | What is agreed in writing for a specific deal |
SLYD is an authorized reseller for Dell Technologies, HPE, Supermicro, Lenovo, Cisco and IBM, and sources through authorized distribution including Arrow, Ingram Micro and TD SYNNEX. Configuration availability and lead times are quoted at time of order. SLYD publishes no support SLA, response time, coverage window, or geographic service guarantee on this page, and states no factory-direct status. No contract record establishes those, so they are absent rather than approximated. Confirm the entitlement in writing against the specific order, in the specific country, before it matters.
Frequently Asked Questions
Common questions about comparing AI server models
Which AI server is best?
None of them, in the abstract. The choice is decided by which accelerator you have settled on, whether your facility can take the cooling method that accelerator requires in that chassis, and whether you are buying servers or a whole rack. Those three constraints usually leave one or two candidates, which is why this page compares attributes rather than publishing a ranking.
Why does this page show no prices or lead times?
Because a static figure for either would be wrong within weeks and cannot be checked. Server pricing depends on configuration, quantity, channel, geography, and negotiated terms, and lead time moves with allocation. SLYD has no governed public record for either, so publishing a number would mean inventing one. Ask for a price against a specific configuration instead.
What is the difference between this page and OEM comparison?
This page compares systems: what a given chassis supports and requires. OEM comparison covers the relationship you buy it through, which means procurement path, support model, geographic reach, and lifecycle fit. They frequently point in different directions, because the best chassis for a workload is not always available on the terms or in the country you need.
Why do some entries say a field is not stated?
Because the OEM does not publish it on the page that entry was read from. The alternative would be to infer it from a sibling model or take it from a reseller listing, and both produce a table that looks complete while carrying details the vendor never committed to. A gap that is visible is more useful than a gap that is filled in.
Does cooling method change which server I can buy?
Often it decides it outright. Some rack-scale systems are liquid-cooled only, so a site that cannot reject the heat cannot deploy them at all. Where an air-cooled variant exists it is usually a different chassis height rather than an option on the same one, which changes the rack arithmetic. Establish what your facility can deliver before shortlisting.
Can I mix accelerator vendors across a deployment?
At the chassis level that depends on the model. Most take one accelerator family, though some are published with a choice across NVIDIA, AMD, and Intel parts. Across a deployment it is more a software and operations question than a hardware one: two accelerator families means two software stacks, two sets of kernels to keep current, and scheduling that understands the difference.
How this comparison is built
Every model detail is read from the OEM's own page for that product, listed in the register below. Reseller listings, distributor catalogs, and comparison articles are not used as the source for any detail here.
What is not done
- No filling in gaps across models. Where an OEM does not state a form factor or a processor on the cited page, the entry says so. It is not inferred from a sibling model in the same family.
- No portfolio-wide figures. Density and heat-capture numbers appear only against the specific configuration the OEM states them for. A 144-GPU-per-rack figure for one OCP form factor is not a property of the vendor.
- No price, lead time, or availability. Volatile commercial facts with no governed public record behind them.
- No ranking. There is no best server in this set. Which one fits depends on your accelerator choice, your facility, and the level you are buying at, and a ranking would hide exactly the constraints that decide it.
- No market share. It does not help you choose between two models, and SLYD has no primary source for it.
Retrieval limitation, stated plainly
Dell's AI server catalog and Lenovo Press respond to direct retrieval. HPE times out, and Supermicro and GIGABYTE return an access error, from the environment this page was compiled in. Details for those three OEMs come from the vendors' own published pages reached through search, and each entry carries the vendor URL it came from. They are cited to the vendor, not to a third party, but they warrant a direct re-check against the vendor sites before release.
What this page is not
A quote, an availability statement, or a substitute for the OEM configuration guide and specification sheet for the system you intend to order. Confirm the exact configuration, its thermal and electrical requirements, and its support entitlement with the vendor before committing.
Sources and basis
One entry per OEM page behind the catalog, listing the models it supports.
PowerEdge XE9712, PowerEdge XE8712, PowerEdge XE9780, PowerEdge XE9780L, PowerEdge XE9785, PowerEdge XE9680L, PowerEdge XE9680, PowerEdge XE7745
https://www.dell.com/en-us/shop/servers/sf/poweredge-ai-servers
GIGAPOD AI DLC Rack, NVIDIA GB300 NVL72
https://www.gigabyte.com/Enterprise/GIGAPOD-Rack-Scale/AI-DLC-Rack_NVIDIA-GB300-NVL72
GIGAPOD, air-cooled configuration, GIGAPOD, liquid-cooled configuration
ProLiant Compute XD685
https://www.hpe.com/us/en/compute/proliant-compute-xd685.html
Compute XD690
Cray XD670
Lenovo NVIDIA GB300 NVL72, ThinkSystem SR680a V3, ThinkSystem SR685a V3, ThinkSystem SR675 V3, ThinkSystem SC777 V4 Neptune
NVIDIA Blackwell portfolio
Liquid-cooled HGX B300 8-GPU, 4U, Liquid-cooled HGX B300 8-GPU, 2-OU OCP
The ordering of the selection criteria, the four support paths, and the guidance on what separates these systems.
SLYD's framing. Reviewed for accuracy, but it is a way of approaching the decision rather than a sourced fact.
Accelerator specifications referenced anywhere on this page.
This page holds no accelerator specification table of its own so the two cannot disagree.
The absence of price, lead time, availability, market share, efficiency percentages, support SLAs, and any factory-direct status.
Volatile commercial facts and contractual claims. SLYD publishes no governed public record for any of them. Authorized reseller agreements are executed through vendor partner portals; no factory-direct claim appears on this page.
That HPE, Supermicro, and GIGABYTE pages could not be retrieved directly from the build environment, so their entries were read from vendor pages reached through search.
Recorded so the difference in retrieval confidence between OEMs is visible. Those entries should be re-checked against the vendor sites directly before release.
Need a Configuration Priced?
Tell us the accelerator, the count, and what your site can deliver, and we will come back with a specific configuration and a price against it.