Where to Buy GPU Servers

New GPU servers come from OEMs like Dell, HPE, Supermicro, and Lenovo, or from Value-Added Resellers (VARs) and brokers. Used and refurbished hardware comes from brokers and resellers on the secondary market. OEM hardware costs more and takes longer to arrive, but you get full warranties and validated configurations. Used hardware is cheaper and ships faster, but the risk is higher.

Where to buy

Buying servers directly from NVIDIA is unrealistic. NVIDIA sells GPU chips andHGX baseboards to OEMs and hyperscalers in volume.

Two main channels serve buyers:

1. OEMs like Dell, HPE, Supermicro, and Lenovo sell new, factory-sealed servers with full warranties.
2. Brokers source new and used hardware across the spectrum: OEM relationships, other data center operators, and other brokers. Some have warehouses and test labs. Others broker deals off spreadsheets without touching the hardware.  ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)} The range is wide.

For new hardware the path is NVIDIA → OEM → VAR → you. A VAR (value-added reseller) is a middleman with OEM relationships who can bundle financing or deployment services.

For used hardware it's longer: NVIDIA → OEM → VAR → original buyer → reseller → you. More hops mean more middlemen, and more places where pricing, condition, and history get murky.

Buying new

OEMs (Original Equipment Manufacturers) build complete servers around NVIDIA's HGX baseboard. ^{[1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/} The HGX is a pre-built board with 8 GPU cards integrated through NVIDIA's proprietary interconnect.

Dell, HPE, Supermicro, and Lenovo are the main four. They add their own chassis, power supplies, cooling, and management firmware, then sell you a branded, warranted system.

What you get: a factory-sealed, tested server with a 3-year warranty, a support contract (next-business-day or 4-hour parts replacement depending on tier), validated firmware, and a known-good configuration. A GPU fails, you call Dell, they send a replacement.

The reason OEMs matter is integration. GPU servers are complete systems where everything has to work together. The baseboard has to match the GPU generation. Firmware across GPUs, NVSwitches, and the server's management controller all need compatible versions.

Get any of this wrong and the NVLink mesh won't initialize. You end up with 8 expensive standalone GPUs instead of a unified system. OEMs handle all of this for you. You get a complete, working system.

Among OEMs, SuperMicro ^{[6]Supermicro, "GPU Systems" (accessed March 2026)https://www.supermicro.com/en/accelerators/nvidia} and Dell ^{[5]Dell Technologies, "PowerEdge Servers for AI" (accessed March 2026)https://www.dell.com/en-us/shop/dell-poweredge-servers/sc/servers} have the strongest reputations.

HGX and DGX

These server nodes come in two main forms:

Most small cluster buyers end up with HGX-based OEM servers. DGX makes sense if you want turnkey and can absorb the premium. Buying loose SXM GPUs only makes sense if you have in-house hardware engineering talent.

The buying process

This is enterprise sales. You contact a sales rep or go through a VAR. VARs often hold pre-negotiated allocations and are the most realistic path to a quote for small buyers who OEMs deprioritize.

You spec the configuration, get a quote, and negotiate. Lead times range from 2 weeks if the config is in stock to 3-6 months if GPUs are allocation-constrained. ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)}

Once you agree on price, you place a purchase order and pay on net-30 or net-60 terms, meaning the invoice is due in 30 or 60 days.

Pricing is not transparent. OEMs don't publish GPU server list prices the way Dell publishes laptop prices. An 8xH100 SXM server runs $250-400K ^{[2]NVIDIA, "H100 Tensor Core GPU" (accessed March 2026)https://www.nvidia.com/en-us/data-center/h100/} depending on NIC count, drive count, and memory. There is room to negotiate, especially on multi-unit orders.

The catch for small buyers: OEM sales teams are incentivized on deal size. Buying 2 servers is not exciting for a Dell rep whose quota is measured in millions. Configs may be limited to the standard catalog, and you may get pushed toward a VAR.

Supermicro is often the most accessible OEM for smaller orders. ^{[6]Supermicro, "GPU Systems" (accessed March 2026)https://www.supermicro.com/en/accelerators/nvidia} Broader catalog, lower minimum quantities, more direct sales process. They're also the most common brand on the secondary market, which matters if you plan to resell.

Servers today, racks tomorrow

Today you will likely buy whole servers. But with Blackwell, NVIDIA expects you to buy whole racks: the GB200 NVL72 is 72 GPUs in a single liquid-cooled rack connected by NVLink across nodes. ^{[4]NVIDIA, "GB200 NVL72" (2024)https://www.nvidia.com/en-us/data-center/gb200-nvl72/}

As systems get more complex and integrated, the next step is pre-assembled clusters, which will command even higher premiums. You should expect procurement to move in this direction: either learn to integrate and build the systems yourself, or expect OEMs to charge higher premiums for integrated systems in the future.

Buying used

New vs. used vs. refurbished is mainly about risk tradeoffs

• New means factory-sealed, never deployed, full OEM warranty. Highest price. Lowest risk.
• Used means previously deployed, sold as-is. Limited verification, limited documentation, little to no warranty. If you rack it and a GPU throws ECC errors (error-correcting code memory faults that indicate degrading hardware), or the NVLink mesh won't fully initialize, that's on you.
• Refurbished is a used system with process applied: completeness check, firmware updates, health diagnostics, 24-72 hours of burn-in testing under load, wear part replacement, a test report with serial numbers, and a seller-backed warranty (typically 30-365 days).

The refurbished premium over used, often 10-25% more, ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)} pays for the validation work and the warranty. You're paying the seller to take on risk instead of you.

Used conditions varies wildly. A server that ran inference for 6 months in a well-cooled Tier III facility, ^{[7]Uptime Institute, "Tier Classification System" (accessed March 2026)https://uptimeinstitute.com/tiers} one with redundant power and cooling paths, is practically new. One that ran training 24/7 at thermal limits for 3 years will have wear. You often don't know which you're getting.

Where the hardware comes from

A secondary market for GPU servers has existed for years, but it was fragmented and mixed in with consumer and standard enterprise gear. That's changing. As of early 2026, there's is more enterprise-grade used GPU server inventory in the market than ever before.

Hyperscalers like AWS, Azure, and Google refresh hardware on 2-3 year cycles. ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)} As Blackwell deployments ramp up, thousands of Hopper servers will eventually be decommissioned and enter the secondary market. Most hyperscalers are still holding onto Hopper hardware as of early 2026, but the wave is coming.

Other than typical used equipment, the rest of the supply comes from three main sources:

• Canceled orders: A company orders 20 servers, financing falls through, and the hardware ships back without ever being racked.
• Surplus: Overprovisioned buyers now offloading excess.
• Failed startups: When neoclouds run out of runway, their hardware gets liquidated.

Vetting sellers

Four questions that separate good brokers from bad ones:

1. Where did this hardware come from, and how was it used?
   
   A reputable broker names a source. "Decommissioned from a Tier III colocation facility in Virginia." or "Canceled order from a neocloud." Refusal to answer is a red flag. A lot of secondary-market hardware was abused in crypto mining or sustained training workloads.
   
   
2. Has it been tested under load, or just powered on? Can you provide a test report?
3. What warranty do you offer? 30 days, 90, a year, or none?
4. What's the resale policy if it doesn't match what we need?

Pricing is negotiable. Get multiple quotes (at least three) to establish a realistic price band.

What actually wears out

GPUs are durable. The silicon doesn't degrade under normal operating temperatures. Transistors don't wear out.

What fails is everything around the die. Fan bearings wear. PSU capacitors age. Thermal interface material between the GPU die and heatsink dries out over 2-3 years, causing thermal throttling. ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)}

These are all replaceable parts, and they're exactly what a good refurbisher checks and swaps.

Due diligence

Buying is exciting, but the first piece of diligence is confirming that buying is the right decision. The GPU cloud market is opening up and getting competitive: more providers, lower prices, more flexibility. Owning hardware means capital outlay, colocation contracts, maintenance, and depreciation risk. If you can't commit the time, money, and operational overhead, don't force it. Renting lets you scale without the procurement risk this article describes.

Warranties

A warranty is a promise to fix or replace failed hardware. What matters is who's making that promise and how fast they move.

OEM warranty comes standard with new hardware: typically 3 years from Dell, HPE, or Supermicro, covering manufacturer defects. Response time depends on the support tier. Basic is next-business-day, premium is 4-hour on-site.

OEMs have the parts supply chain, which makes response times predictable.

When used hardware changes hands, OEM warranty coverage gets complicated. Some OEMs allow warranty transfer to the new owner if the original warranty is active. Others require the original purchaser to file claims. Some void coverage entirely.

Always verify with the OEM using the serial number before counting on it.

Third-party maintenance (TPM) providers sell contracts on used or out-of-warranty hardware. They stock replacement parts and dispatch technicians, usually next-business-day, with 4-hour on-site available at a premium. The risk: they may not have the specific part in stock when you need it, like a particular NVSwitch baseboard or GPU SKU.

ITADs, refurbishers, and brokers offer their own seller warranties, typically 30-365 days, covering dead-on-arrival and early failures. The quality depends entirely on the seller's reputation and financial stability. A warranty from a company that folds six months later is worthless.

The other option is self-insuring: skipping warranties and keeping spares. Buy 50 servers, set aside a spare GPU, a few PSUs, some fans. For smaller clusters under 50 servers, this is especially risky because one GPU failure could be $150K+. You also need dedicated technicians on standby to perform repairs, which adds ongoing cost.

At hundreds of servers the economics could shift. Failure rates are low enough that spares cost less than warranty premiums across the fleet. Depends on your use case and reliability requirements.

What fails

What to check

If you're buying new from an OEM, most of this is handled for you. For used or refurbished hardware, every item here matters.

Configuration. Get the full spec sheet from the seller and match it against the OEM's published configuration for that model:

• GPU model and count (ex: confirm 8xH100 SXM5 80GB specifically; PCIe is a different, lower-performance socket)
• HBM memory per GPU
• NIC model and count (ConnectX-6 vs ConnectX-7 determines NDR 400 Gb/s vs HDR 200 Gb/s InfiniBand)
• NVMe drive count and capacity
• DDR5 DIMM count
• PSU count and wattage

Many ITADs and less experienced brokers don't fully understand GPU server specifications themselves. Cross-reference the quote against the OEM's published configuration. Upload both documents to an LLM and ask it to flag discrepancies.

Missing or swapped components are common in secondary-market hardware, ^{[8]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)}especially systems that were parted out and reassembled.

Burn-in testing. Any reputable seller runs hardware under load for 24-72 hours before shipping. That means GPU stress tests under sustained load. It means running multi-GPU bandwidth tests across all 8 GPUs to confirm the full NVLink mesh is working.

Ask for the test report. If they can't produce one, that tells you something.

Firmware. Outdated firmware causes real problems: performance drops, CUDA compatibility issues, NVLink initialization failures. Check the GPU firmware version, NVSwitch firmware, and server BIOS against the OEM's recommended versions.

ECC error history. GPUs track correctable and uncorrectable memory errors internally. A small number of correctable errors is normal. A high count, or any uncorrectable errors, means memory degradation. Ask for nvidia-smi output or GPU diagnostic reports showing error counters.

Physical inspection. Scratches on the chassis don't matter. Corrosion on connectors does. Check for thick dust buildups (signals a poor previous environment), bent CPU socket pins, and cable condition. Keep in mind that the refurbishment process itself can introduce damage. Even servers that were working perfectly in their previous data center can suffer from mishandling during decommissioning, transport, and reassembly.

Serial numbers. Confirm them with the OEM. Check remaining warranty status. Check whether the hardware has been reported stolen. A legitimate seller provides serials upfront without hesitation.

References

1. NVIDIA, "HGX Platform" (accessed March 2026)
2. NVIDIA, "H100 Tensor Core GPU" (accessed March 2026)
3. NVIDIA, "DGX Systems" (accessed March 2026)
4. NVIDIA, "GB200 NVL72" (2024)
5. Dell Technologies, "PowerEdge Servers for AI" (accessed March 2026)
6. Supermicro, "GPU Systems" (accessed March 2026)
7. Uptime Institute, "Tier Classification System" (accessed March 2026)
8. Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)