Who Makes AI Servers, and Where to Buy One

Who makes AI servers? The short answer

Who makes AI servers depends on the layer you mean. NVIDIA makes most of the accelerator chips, and AMD and Intel make the main alternatives. Server makers such as Dell, HPE, Lenovo and Supermicro build complete systems around those chips, and most buyers order from one of them or from a reseller. If you only need the compute for a few months, renting a GPU server is often the better first step.

Here, an AI server means a server built around one or more GPUs (graphics processing units) or similar accelerators, used to train or run machine learning models. This guide separates the makers by layer and names the product families to look for. It then covers each way to buy one, plus the power and delivery problems that stop many purchases late.

Why chip makers and server makers are different companies

An accelerator is the chip that does the math. A system is everything around it: the chassis, CPUs, memory, power supplies, fans, network cards and firmware. Different companies build each layer. The warranty and the support contract come from the company that built the system, so that is the company you are really choosing.

• NVIDIA sells data-centre GPUs and also complete systems under its own DGX brand. It publishes reference designs that other makers build on: HGX is the board that carries eight linked GPUs, and MGX is a modular server design.
• AMD sells the Instinct family of accelerators. You buy them inside servers from the system makers, not as a box from AMD.
• Intel sells the Gaudi family of accelerators, also through system makers.
• Cloud chips such as Google's TPU (tensor processing unit) and AWS Trainium are rented inside those clouds. You cannot buy a server built on them.

Most AI servers sold today use NVIDIA GPUs. The reason is software: most AI tools are written for CUDA, which is NVIDIA's programming platform. AMD's ROCm platform runs the major frameworks, but check that every tool you depend on supports it before you order an AMD system.

The system makers: OEMs, NVIDIA DGX and ODMs

An OEM (original equipment manufacturer) designs servers and sells them under its own brand, with its own warranty and support. Model numbers change with every GPU generation, so look for the product family instead. As of October 2026, these are the main OEM families for GPU servers:

• Dell: the PowerEdge XE line holds its GPU-dense servers. Other PowerEdge models take one to four PCIe GPU cards.
• HPE: GPU-dense systems sit in the ProLiant Compute XD line. Its largest liquid-cooled builds carry the HPE Cray name.
• Lenovo: ThinkSystem servers. Neptune is the name of its direct liquid cooling.
• Supermicro: GPU SuperServer models in more chassis variants than most makers, from one-card servers up to full liquid-cooled racks.
• Gigabyte: G-series GPU servers, and GIGAPOD for rack-scale clusters.
• ASUS: the ESC series of GPU servers.

NVIDIA DGX is NVIDIA's own system line. It runs from DGX Spark, a small desktop unit, through DGX Station, up to 8-GPU DGX servers. DGX SuperPOD is a reference design for a whole cluster of them. DGX systems ship with NVIDIA's software stack and NVIDIA support. Several OEMs also sell their own versions of the smaller DGX designs.

An ODM (original design manufacturer) builds servers to a customer's design. The large cloud providers buy most of their servers this way, in very large volumes. Foxconn, Quanta, Wistron (with its server arm Wiwynn) and Inventec are the names you will see in that market. Most ODMs do not sell one server to one buyer. Quanta is the exception you are most likely to meet, because its QCT brand also sells to enterprises.

Who makes AI servers in India?

The names below are the ones a public source confirms. Global OEMs also sell in India through local partners, so this list covers local manufacturing only.

• Netweb Technologies sells AI systems under its Tyrone brand. On 18 February 2026 it launched the Tyrone Camarero GB200 system and Tyrone Camarero Spark, both described as manufactured in India on NVIDIA platforms (Business Today, February 2026).
• VVDN Technologies, in Manesar, Haryana, designed an 8-GPU AI server called Adipoli. Electronics and IT Minister Ashwini Vaishnaw showed it in April 2025 as the first AI server fully designed in India (The Tribune).
• HPE builds some of its servers in India through VVDN's plant in Manesar. HPE announced the plan in July 2023 and reported the first servers in production in April 2024 (HPE press release). That programme covers high-volume servers, so ask HPE which GPU models it includes.

"Made in India" often applies to some models in a range and not to others. If local manufacturing matters for a tender or an incentive scheme, ask the seller for the exact model and the paperwork that proves it.

What kind of AI server are you actually choosing?

The maker matters less than the class of system. There are four classes, and each one has different power needs, buyers and prices.

8-GPU training systems. Eight high-end GPUs on one HGX-style board, linked by NVLink, which is NVIDIA's fast GPU-to-GPU connection. They are built for training and for serving very large models. They need data-centre power and cooling, and they are sold by quote. NVIDIA's DGX B200 and the OEM HGX servers in the families above belong here.

1 to 4 GPU inference servers. Standard rack servers with PCIe GPU cards. Inference means running a trained model to answer requests, and most companies that self-host a model only need this. These servers fit an ordinary rack, and many OEMs let you configure them online.

GPU workstations. Desk-side towers with one to four workstation GPUs: Dell Pro Max, HP Z and Lenovo ThinkStation, plus NVIDIA's DGX Station and DGX Spark. They run on office power and sit under a desk. They suit one developer or a small team testing models. Once it is running, a workstation is a Linux box like any other, so a driver that will not load needs the same fix as on a server: see why the NVIDIA driver does not load on Ubuntu Server.

Turnkey appliances. A server sold with the model runtime, a chat interface and support already installed. You pay for the setup and the support contract on top of the hardware. Some resellers build their whole business on this model. We compare one of them, step by step, in a Zanus AI private server versus self-hosting the same stack. The useful question for any appliance is what you could do yourself, and what you would rather pay someone to own.

Where to buy an AI server: the five channels

OEM direct and online configurators

Dell, HPE and Lenovo let you configure many servers on their own websites. For smaller GPU servers you can often see a price and order online. For 8-GPU systems most makers send you to a sales team for a quote, because the price depends on volume, support level and current GPU supply. OEMs also run financing arms, such as Dell Financial Services and HPE Financial Services, if you would rather lease than pay up front.

VARs and distributors

A VAR (value-added reseller) sells a maker's hardware and adds services: sizing, installation, or a support contract. Distributors such as TD SYNNEX and Ingram Micro supply the VARs. NVIDIA lists its approved resellers in the NVIDIA Partner Network. A VAR is often the only practical route for a small business, because the OEM's direct team focuses on large accounts.

System integrators

A system integrator builds, tests and installs the system for you. For a multi-server cluster this matters more than the server brand. Someone must cable the high-speed network, run burn-in tests and prove that every GPU works under load. Ask who will do that, and in writing.

The refurbished market

Previous-generation GPU servers come back on sale when cloud providers and labs upgrade. They can be good value for inference. Check these points before you buy:

• How long the warranty lasts, and whether the refurbisher or the original maker honours it.
• Whether the GPUs are SXM modules or PCIe cards. SXM modules are mounted on the HGX board, so a failed one needs a matching board part, not any GPU card.
• Whether the current NVIDIA driver branch still supports that GPU generation. NVIDIA drops old generations from new drivers, which means you can end up stuck on an old kernel. A kernel update can also break the driver build, as when the NVIDIA DKMS build fails on a new kernel explains.
• What input voltage the power supplies need. High-wattage server power supplies often reach full output only on 200 to 240 V.

Renting instead of buying

You can rent the same GPUs by the hour or the month. A VPS with a GPU gives you a full Linux server with no hardware to own, and you can stop paying the day the project ends. If you only call a model, a hosted API may cost less still. The point where owning wins depends on how many hours a day the GPU stays busy, and the break-even between a GPU VPS and API tokens works through that maths. For local models in particular, compare Ollama Cloud with running Ollama on your own server.

What does an AI server cost?

Prices change fast and most 8-GPU systems have no public list price. Treat any figure you find online as an estimate unless it is a written quote with your configuration on it.

One sourced example: NVIDIA put DGX Spark on sale on 15 October 2025 at $3,999, before local taxes and tariffs. That is NVIDIA's announced price for a desktop unit. It says nothing about the price of an 8-GPU server, which is in a different class entirely.

If you plan to build your own instead, the parts list and the trade-offs are in what an AI server costs to build.

Lead times: why the ship date matters

New GPU generations are supplied in limited batches at first, so a quote for a current-generation system can carry a ship date many weeks out. Ask for an estimated ship date in writing. Also ask which part is holding the date back, and what happens if it slips. The previous generation is often in stock, and an integrator sometimes holds built systems. Both can be faster than waiting for the newest GPU.

Power and cooling decide where it can live

Power is the limit most first-time buyers find too late. NVIDIA publishes maximum power figures for its DGX systems, and they show how wide the range is.

[
  {
    "label": "DGX Spark (power supply rating)",
    "max_power_kw": 0.24
  },
  {
    "label": "DGX H100",
    "max_power_kw": 10.2
  },
  {
    "label": "DGX B200",
    "max_power_kw": 14.3
  }
]

NVIDIA lists the DGX B200 at about 14.3 kW maximum, in 10 rack units of space. The older DGX H100 is listed at 10.2 kW. DGX Spark ships with a 0.24 kW power supply. A North American 120 V, 15 A office circuit delivers 1.8 kW in total, so one 8-GPU server needs dedicated high-voltage circuits and rack PDUs (power distribution units) sized for it.

All the power a server draws becomes heat in the room. An 8-GPU air-cooled server is also very loud, so it belongs in a server room or a colocation rack, not an office. NVIDIA's rack-scale GB200 NVL72 design is liquid cooled, which means the building needs a coolant system, not only air conditioning. Before you sign a quote, ask your data centre or colocation provider how many kW one rack can draw and remove as heat. If the answer is lower than the server's maximum, the server cannot run at full load there.

When the server arrives, check it before you sign off. Run nvidia-smi and confirm it lists every GPU you paid for. Run nvidia-smi topo -m and confirm the GPUs connect the way the quote described. Then keep the GPUs under full load for a day. A fault that appears under load is far easier to return in the first week than in the tenth month.

Buy, rent or build: a decision rule

Use this order. Stop at the first rule that fits.

1. The data cannot leave your building. Buy, and plan the network isolation first. What an air-gapped AI server takes covers updates, model transfer and licensing with no internet link.
2. The GPUs would sit idle most of the day. Rent. A GPU VPS or an API costs nothing while you are not using it.
3. You run one or two models for inference, for a team. Buy a 1 to 4 GPU inference server from an OEM or a VAR, or a workstation if it will sit in an office.
4. You train large models across many GPUs, every day. Buy 8-GPU systems from an OEM or through an integrator, but only after the facility confirms the power and cooling.
5. You have no one to run Linux and GPU drivers. Buy a turnkey appliance or an integrator support contract, and price that support as part of the server.

The maker is the last choice, not the first. Pick the class of system and the channel. Then compare the OEMs that sell that class through that channel on support terms and ship date.

FAQ

Does NVIDIA make AI servers itself?

Yes, under the DGX brand, from the desktop DGX Spark up to 8-GPU DGX servers and whole DGX SuperPOD clusters. Most NVIDIA-based servers, though, are built by OEMs such as Dell, HPE, Lenovo, Supermicro, Gigabyte and ASUS on NVIDIA's HGX and MGX reference designs. The OEM, not NVIDIA, provides the warranty on those systems.

Who makes AI servers in India?

Netweb Technologies builds AI systems in India under its Tyrone brand, including the Tyrone Camarero GB200 system and Tyrone Camarero Spark launched in February 2026. VVDN Technologies in Manesar designed an 8-GPU AI server called Adipoli, shown in April 2025. HPE also manufactures some servers in India through VVDN. Ask the seller which exact models are built locally.

Can I buy an 8-GPU AI server online with a list price?

Usually not. Most makers sell 8-GPU systems by quote, because the price depends on GPU supply, volume and support level. Smaller 1 to 4 GPU servers and workstations can often be configured and priced on OEM websites. Treat any 8-GPU price you read online as an estimate until you hold a written quote.

Should I buy an AI server or rent a GPU VPS?

Rent if the GPU would sit idle for much of the day, or if the project has an end date. Buy if the GPU stays busy most hours, or if your data must stay on hardware you control. The break-even depends on daily GPU hours, the rental rate and the power cost where the server would live.