Enterprise server buying guide: brands, generations and sizing

Choosing an enterprise server is three decisions dressed up as a hundred spec-sheet rows: which platform family (brand and generation), what size (CPU, memory, storage matched to the workload), and what condition (new with full warranty, or refurbished at a fraction of the price). Everything else, from drive carriers to bezel keys, follows from those three.

This is the hub of our server buying series, written from the vantage point of a trading desk that quotes these machines daily. It links out to model-level comparisons and deeper sizing guides as they publish.

At a glance

• Dell, HPE, Lenovo and Supermicro cover most enterprise purchases; the right one depends on support model and ecosystem, not raw spec
• One generation back often buys most of the performance at a steep discount
• Size by workload: virtualization wants cores and RAM, storage wants bays and controllers, AI wants GPU room
• Refurbished enterprise servers are a mature market with real warranties from credible sellers
• Rack 1U/2U covers most deployments; tower for offices, 4U for GPUs and storage density
• Total cost is chassis plus drives, RAM, licensing, rails and support, never the base price alone

The brand landscape

Four names dominate the machines that cross our dock, and the honest differences are about ecosystem rather than benchmark supremacy.

Practical guidance from the quoting desk: mixed fleets standardize on Dell or HPE for the management tooling; cost-driven and GPU-dense builds look hard at Supermicro; Lenovo wins competitive bids more often than its mindshare suggests. If your team already runs one vendor's management stack, switching brands costs more in habits than the hardware saves.

Generations: where the value hides

Server platforms track CPU generations. The current-versus-previous pattern repeats every cycle: the newest generation (current Xeon Scalable or AMD EPYC platforms; Dell R760/R770, HPE Gen11 class) carries DDR5, PCIe 5.0 and premium pricing, while the previous generation (R740/R750, Gen10 class) delivers most real-world throughput for common workloads at a much lower clearing price, especially refurbished.

The decision rule we give buyers: choose current-generation when the workload is CPU-bound at scale, power efficiency drives the TCO, or the platform must live 5+ years in production. Choose previous-generation when the workload is I/O- or memory-bound, budgets are tight, or the machines back a lab, branch or secondary site. A previous-generation server with more memory routinely beats a newer one with less for virtualization duty.

Sizing by workload

A sizing question we hear constantly is some version of how much server a hundred-person company needs. The honest answer: a single well-specified 2U host (or a two-node pair for failover) virtualizes the entire back office of most 100-person firms, which is why the sizing conversation should start from workloads and growth, not headcount.

New or refurbished

Refurbished enterprise servers are a mature, liquid market fed by datacenter refresh cycles, and for many buyers they are the rational default. What a credible refurbished purchase includes: stated condition grading per unit, burn-in and component test results, cosmetic and functional photographs on request, a real return window and warranty from the seller, and clean provenance. What it usually does not include: transferable manufacturer support contracts, so factor a spares strategy or third-party maintenance into the plan.

When new wins: strict procurement rules, multi-year vendor support requirements, latest-generation-only workloads, or projects where lead time on a specific exact configuration matters more than price. When refurbished wins: virtualization estates, storage boxes, lab and dev fleets, branch sites, and any budget where two refurbished hosts beat one new one for resilience. Our server catalog states condition and grading per listing so both paths are visible side by side.

Form factors in one minute

Rack 1U suits compute density with limited expansion; 2U is the enterprise workhorse with drive bays and GPU room; 4U buys storage density or serious GPU capacity; towers serve offices without racks (quieter, easy to service); blades and multi-node chassis only pay off at scale with uniform workloads. When in doubt, 2U dual-socket is the shape that fits the widest range of futures.

What actually drives the total cost

Base chassis price is the headline, but the invoice grows from parts and terms. As of Q3 2026 the pattern in our quotations is stable: memory and NVMe storage dominate configured cost on virtualization builds; GPUs dominate anything AI-adjacent; and OS or hypervisor licensing (per-core in several ecosystems) can rival hardware on high-core machines. Rails, bezels, second PSUs, network cards and regional power cords are small lines that stall deployments when forgotten. Freight and duties from Asia are a modest, predictable share when quoted with clear Incoterms, which is exactly how our sourcing and import hub guide recommends comparing offers.

A buying checklist that survives contact with procurement

• Workload defined, growth estimated for 24-36 months
• Platform family chosen (brand + generation) with management tooling considered
• Configuration priced with drives, RAM, NICs, rails and licensing included
• Condition decided (new vs refurbished) with warranty terms in writing
• Power, rack depth and cooling confirmed at the destination site
• Quotation on a proforma invoice with Incoterms and lead time stated

FAQ

How much does an enterprise server cost?

The range is too wide for one number: a refurbished previous-generation 2U host costs a small fraction of a fully configured current-generation machine, and configuration (RAM, drives, GPUs, licensing) moves cost more than the chassis does. Get line-item quotations for two or three candidate shapes; the comparison teaches more than any average.

Is it cheaper to build a server than to buy one?

Whitebox builds can undercut brand pricing for homogeneous fleets run by teams comfortable owning integration and spares. For most businesses, enterprise platforms win on management controllers, validated parts, and a single accountable warranty, and refurbished brand-name servers usually beat whitebox economics anyway.

How many servers does a 100-person company need?

Often one to three: a virtualization host (or a pair for failover) plus optionally a storage or backup box. The driver is workloads (ERP, file volumes, build systems), not headcount, which is why sizing conversations start with what runs, not how many people log in.

What are the requirements for an enterprise server room?

Stable power with headroom for the configured load, cooling for the heat actually produced, rack depth matched to the chassis (many 2U servers exceed shallow cabinets), and network drops sized to the NICs. Confirm all four before delivery day.

Do refurbished servers come with a warranty?

From credible sellers, yes: a stated seller warranty and return window, distinct from the original manufacturer contract, which rarely transfers. Ask for the warranty terms in writing on the proforma invoice, the same place price and lead time belong.