Server hardware procurement in 2026 is governed by the same set of constraints it has been for two decades, only the numbers have moved: minimum RAM has climbed from 1 GB to 8 GB across mainstream database and grid infrastructure checklists, and entry CPUs now sit at 1 GHz minimum while production workloads run on multi-core silicon an order of magnitude faster [S4][S5].
Procurement engineers who treat the OEM minimum as the target get bitten; those who treat it as the floor and oversize for RAID, polling, and concurrency are the ones whose fleets stay out of the incident queue [S1][S3]. The point of this guide is to walk the BOM line by line, with the figures the vendors actually publish, so the RFQ goes out with the right tolerances the first time.
Workload tier defines the minimum bill of materials
Every server hardware sizing exercise starts with a tier, and the tier drives the published numbers. Atlassian's Confluence 7.9 hardware guide is explicit that peak visitors (concurrent browsers making requests) and content size are the two scaling axes, and that live load testing is the only reliable way to predict production demand [S1]. The guide anchors its small/medium/large tiers to user count and page volume rather than to a fixed CPU model, which is the methodology procurement should copy.
Oracle Enterprise Management gives a more concrete template: Small Systems cover 1 to 100 sites, Medium Systems 100 to 250 sites, and Large Systems 250 to 1,000 sites, with polling or non-polling data retention extending out to 400 days [S3]. Dual processors are not used by the EM software, so buying a second socket delivers no headroom, only heat; that kind of vendor-stated limitation should always be checked before paying for it [S3]. RAM on the EM reference build moves from 2 GB at the small/non-polling tier to 4 GB at medium-to-large/non-polling, with 2× 80 GB SATA disks stepping up to 3× 80 GB RAID 5 when the workload justifies it [S3].
CPU, RAM, and disk floors from three vendor checklists
Sun N1 Service Provisioning System 5.1, a documented legacy benchmark, lists Master Server requirements of 450 MHz on SPARC/Solaris, 1 GHz on Red Hat Linux, and 1 GHz on Windows, with a flat 1 GB RAM floor and 2 GB free disk across all three [S4]. Local Distributor, Remote Agent, and CLI Client roles drop to 400 MHz on most Unix variants, 1 GHz on Linux/Windows, and 256 MB RAM on AIX, HP-UX, and older Solaris, again with 1 GB disk [S4]. Those numbers are entry-points, not production targets, but they define the minimum costed configuration in any RFQ where a vendor proposes a sub-1 GHz part.
Oracle Grid Infrastructure 19c, by contrast, demands at least 8 GB RAM and a 1,024 × 768 capable display card for the Universal Installer, and requires the Intelligent Platform Management Interface (IPMI) on the server for out-of-band management [S5]. Runlevel 3 or 5 is the supported Linux console state, and the checklist explicitly puts IPMI alongside RAM as a preinstall gate, not a nice-to-have [S5]. If the procurement spec does not call out IPMI and the 8 GB floor, the install will fail before the database binaries are laid down.
RAID, network, and power as decision criteria, not afterthoughts
Three decision criteria separate a workable server from a reliable one: disk resiliency, network throughput, and power profile. Oracle's EM checklist tells procurement to implement RAID 5 in the hardware controller rather than the operating system, and to size for 3× 80 GB drives when a single drive failure must not take a site offline [S3]. Network card tier moves from 10/100 in the small configuration to 10/100/1000 in medium-to-large, and that jump is the cheapest insurance against polling backlogs when 400 days of data are in flight [S3].
Microsoft's Windows 10 hardware dev guidance frames power as a balance, not a binary: high performance and low energy are conflicting goals, and component-level power characteristics (CPU C-states, PSU 80 Plus rating, VRM efficiency) must be tuned together to hit the desired perf/watt point [S6]. For procurement, this means an RFQ that asks only for "low power" or "high performance" is unanswerable; the spec has to call out target utilization, idle power, and peak power, plus the operating-temperature envelope. Articles like Clamp meter compatibility with temperature limit requirements: 0 to 40 °C envelope, CAT handle a similar envelope question for instrumentation, and the same discipline applies to server inlet temperature and rack airflow.
Comparison table: vendor minimums side by side
The three vendor-published checklists give procurement a defensible comparison anchor. Oracle EM Small/Medium (non-polling) ships a 3.2 GHz Pentium 4 reference processor, 2 GB RAM, and 2× 80 GB SATA, with a 10/100 NIC; the Medium/Large tier keeps the same CPU, jumps to 4 GB RAM, 3× 80 GB RAID 5, and a 10/100/1000 NIC [S3]. Sun N1 SPS 5.1 Master Server floors at 1 GHz CPU, 1 GB RAM, 2 GB disk, with no RAID requirement called out in the legacy spec [S4]. Oracle Grid Infrastructure 19c on Linux requires 8 GB RAM minimum, IPMI, runlevel 3 or 5, and a 1,024 × 768 capable display, with no public minimum CPU clock but an implicit modern-x86 assumption [S5].
Reading these three side by side, the procurement line that matters is not the CPU model; it is the RAM floor (1 GB → 2 GB → 4 GB → 8 GB as workload climbs) and whether RAID is mandatory or optional. If the BOM does not include RAID 5 in hardware for any production database tier, the failure-mode cost is a single-drive event taking the service down, which the OEM docs warn against explicitly [S3][S5].
Procurement workflow: from checklist to RFQ to acceptance
A spec-first procurement workflow runs in five steps. First, classify the workload into a tier (small / medium / large, or polling / non-polling) and write that tier into the RFQ so vendors quote against the same envelope [S1][S3]. Second, pull the OEM hardware checklist for the exact software version being deployed (Oracle Grid Infrastructure 19c, Confluence 7.9, z/OS, etc.) and treat the listed minimums as gate conditions, not as targets [S1][S5]. Third, map each RFQ line to a decision criterion: CPU to tier and per-instance concurrency, RAM to peak user count, disk to retention window and RAID level, NIC to polling volume, and IPMI to out-of-band management requirement [S3][S5].
Fourth, require vendors to populate a BOM against that matrix, and reject any quote that lists sub-floor specs or that omits RAID, IPMI, or NIC speed. Fifth, on delivery, run a hardware acceptance test that exercises the same peak-visitor or polling scenario the sizing was built around; the Atlassian guidance is direct on this: live testing is the only reliable way to validate production readiness [S1]. For broader industrial equipment procurement patterns that follow the same tier-and-checklist logic, the VFD Suppliers in 2026: Specs, Brands, and Sourcing Map walkthrough is a useful cross-reference.
Where server hardware sizing goes wrong
Three failure modes account for most post-deployment pain. The first is buying to the OEM minimum: 1 GB RAM meets the Sun N1 floor but does not survive a real polling workload with 400 days of data retention, so production collapses the first time a long poll cycle overlaps with a software update [S3][S4]. The second is implementing RAID in the OS rather than the hardware controller, which both Oracle EM and general server-storage guidance flag as a resilience regression because the controller is what survives a drive pull with the array still online [S3].
The third is omitting IPMI from the BOM; without it, every firmware update or hung kernel recovery requires a physical visit, and the Oracle Grid Infrastructure checklist is unambiguous that IPMI is a preinstall requirement, not an option [S5]. Skipping it turns a remote-management feature gap into a truck-roll cost on day one. The fix in all three cases is the same: write the criterion into the RFQ, gate acceptance on it, and treat the OEM checklist as the floor not the ceiling.
Signals worth tracking into Q4 2026
Two procurement signals are worth watching over the next two quarters. First, Oracle's hardware checklist for Grid Infrastructure 19c lists 8 GB RAM as the current floor; any future major version bump on the 19c track, or a 23ai release line, will likely move that number upward, and RFQs written against 8 GB should include a clause allowing the buyer to requote at the new floor when the next revision publishes [S5]. Second, IBM's z/OS cryptographic hardware feature support, with a maximum of 16 features per IBM z13 or zEnterprise EC12 frame, is the kind of capacity ceiling that mainframe procurement teams hit before they expect to, and any RFQ for crypto-accelerator-bound workloads should reserve headroom against that 16-feature cap [S2]. Server hardware platforms also continue to span architectures from Netra X1-class small servers to multi-socket enterprise frames [S7], so the BOM conversation has to start with architecture, not with a brand.
For component-level specifications, see serial server, architectural hardware, and linear guide.