Aquapros
20451 Timberlake Road, Lynchburg, Virginia 24502
Vendor estimate #4593 · 5 September 2026
Prepared by Greg Bentz, MySEODesk
Vendor estimate #4593 · 5 September 2026
Prepared by Greg Bentz, MySEODesk
Two estimates are on the table at $34,800 and $40,960. Both are built on the same single socket entry chassis, and the hardware line alone accounts for almost every dollar of difference between them. Part of that cost is real: DDR5 memory pricing is genuinely elevated right now, and a current generation server is obliged to use it. That is the argument for the previous enterprise generation, which takes DDR4 Registered ECC, carries far more expansion, and prices materially lower without the memory market attached to it.
The quoted machine is Dell's current generation in its entry tier. It is the newest model in the lowest trim. The alternative is the prior generation in the enterprise tier: two processors instead of one, eight times the memory channels, eight times the memory ceiling, every drive bay filled, and memory that is not in the middle of a shortage, for materially less money.
You are being quoted a brand new economy car at the price of a well kept luxury SUV a few years old. The SUV is the bigger, better built of the two, and it is the cheaper one.
Paying more for the newer generation only makes sense if the generation itself is what you needed. Here it is not. What you need is cores, memory capacity and storage flexibility, and the enterprise tier has more of all three even a generation back.
The stronger version of this. For roughly what is being asked for one server, you can have two, running as a Proxmox cluster with your virtual machines replicated between them. To be straight about it: once migration labor is added back, that option lands a little above the cheaper estimate and comfortably below the higher one. It is not the cheapest thing on the table. It is the only thing on the table that keeps the business running when a server dies on a Tuesday morning, and that is what the difference buys.
Warranty is a line item here, and it is already published. The obvious objection to prior generation hardware is support, and on the refurbished channel it is simply a checkbox on the same order form, sold in one, three and five year terms. The tiers run from business hours cover with next business day parts, up to 24x7 cover with a four hour on site response. The quoted new server carries three years of next business day ProSupport, per the estimate. So the cheaper machine can be covered harder than the expensive one, and the proposed figures already include three years of 24x7 cover with next business day response on both nodes. Dell branded warranty is available on some configurations subject to eligibility, which is worth asking about for this exact machine rather than assuming either way.
The short version. The quoted server is a Dell PowerEdge R360, a single socket entry class machine, carrying a price we cannot reconcile without an itemized build sheet. Alongside it sit a Windows Server line billed at quantity two with no explanation given, a RAID controller with no write cache underneath a SQL Server purchase, and storage that nobody has sized against real data. Some of the price is genuinely the DDR5 market, which is exactly the argument for not buying a current generation platform this year. The fix is not to haggle. It is to change what is being bought: a previous generation enterprise chassis with DDR4 Registered ECC memory, a disk controller chosen deliberately for the design, and a hypervisor underneath so one failure does not take the whole company offline.
Both estimates share the same line structure. The only material difference between them is the server hardware line, which moves from $19,650 on estimate #4593 to $25,500 on the five drive variant. Nothing else in the build changes.
| Line item | Quoted | Proposed range | Indicative difference | Note |
|---|---|---|---|---|
| Server hardware, fully configured | $19,650 | $10,400 | $9,300 less | Configured at published list from the refurbished channel and including three years of 24x7 support with next business day on site response. Not an estimate |
| Windows Server Standard x2 | $2,725 | $1,900 to $2,400 | $300 to $800 less | two 16 core licenses. Standard covers two virtual machines each, and this design runs four |
| Windows Server CALs (30) | included | $1,300 to $1,600 | not itemized | not shown separately on the estimate. Confirm whether the vendor's line includes them |
| Remote Desktop CALs (10) | $1,625 | $1,400 to $1,700 | $100 more to $300 less | |
| SQL Server + 15 CALs | $4,350 | $3,700 to $4,300 | $0 to $600 less | confirm edition and channel before purchase |
| UPS 1500VA + battery | $2,324 | $1,600 to $2,000 | $300 to $700 less | |
| Labor and migration | $2,500 | not included | excluded | Deliberately not priced here. Migration scope depends on the current environment, which we have not seen. Note the vendor total does include this line and ours does not |
| Subtotal | $33,174 | $20,300 to $22,400 | $10,800 to $12,900 less | |
| Estimated tax at 4.9% | $1,626 | $1,000 to $1,100 | ||
| Total | $34,800 | $21,300 to $23,500 | $11,300 to $13,500 less |
The R360 is Dell's single socket entry server. It is a capable machine for a light workload. It is not the platform you put Active Directory, SQL Server and ten Remote Desktop users on, and the pricing attached to it here is not explained by anything on the estimate.
Cost falls and capacity rises at the same time. That is the whole argument. Normally you pay more to get more, and the question is whether the extra is worth it. Here the more expensive machine is also the smaller one, so there is no trade to weigh. Cost figures are estimates and the proposed value shown is the midpoint of the range.
A premium is normally defensible because you grow into it. You buy more machine than you need today so that in year three you are not buying again. That argument cannot be made here, because the cheaper option is not the smaller one. It starts with more cores, more memory, more bandwidth and more bays, so there is nothing to grow into.
To be fair to the quoted machine, it does buy something real, and more than we first credited: it is the faster processor outright, on a newer architecture, with the current memory generation. What it does not buy is capacity, and capacity is what four virtual machines and ten Remote Desktop users consume. The one running cost that genuinely favours it is power: the proposed pair of CPUs is 210W against 65W, which across five continuous years is on the order of a few hundred dollars including cooling. That is a real number and it belongs in the comparison. It is also roughly two percent of the purchase gap.
It does not change the recommendation, because the processor was never the problem. Nothing in either estimate suggests this business is short of CPU. What the quoted build is short of is storage: 960GB of usable space, on a controller with no cache, in a chassis with five empty bays. It is also short of memory headroom, short of a second machine, and carrying a licence it only half uses. A faster processor does not fix any of those, and none of them are fixed by spending $34,800.
What the older platform buys is width and headroom rather than speed: eight times the memory channels, eight times the memory ceiling, a real host bus adapter, eight populated drive bays, and two populated processor sockets rather than one. For four virtual machines, a database and ten desktop sessions running at once, that shape matters more than peak clock. But if raw processor speed is the priority, the honest answer is that the quoted machine has it.
The list runs from eight cores to twenty four, and the obvious move is to take the most cores per dollar. That would be a twenty four core part. It is the wrong answer here for two separate reasons.
The first is Microsoft. Windows Server Standard is licensed in sixteen core increments. Anything above sixteen physical cores needs additional core licences, on every host, and twice over on a host running four virtual machines. A modest processor upgrade quietly costs several times its own price once licensing is counted. Note the same rule cuts the other way against the quoted build: an eight core chip wastes half of the base licence it is already paying for.
The second is that clock cannot be bought back on this platform. The obvious counter to a sixteen core part is to spend more on a higher clocked one instead, since database latency is largely bound to single thread speed. We priced that out and rejected it. The faster options in this range gain a few hundred megahertz and cost considerably more, and they still do not approach the quoted processor, which is two generations newer. Paying a premium to lose a single thread comparison by slightly less is not a good use of the budget.
So the money goes where it is not already lost: width, inside one licence. Sixteen cores and thirty two threads is double what the quoted machine offers, it consumes exactly the licence being bought rather than half of it, and it is the shape that four virtual machines, a database and ten desktop sessions actually consume.
Those sixteen cores are delivered as two processors rather than one, and that detail is worth money. Windows counts physical cores, not sockets, so two eight core processors licence identically to one sixteen core processor. But two populated sockets double the memory channels, which is where this workload is genuinely constrained, and the pair costs slightly less than the single larger chip. It also leaves no empty socket, which is the usual reason a server cannot be expanded later without being replaced.
The trap to avoid is filling both sockets with the large processors. Two sixteen core parts is thirty two physical cores, and Windows would then need twice the licences on every host, on both nodes. The processor upgrade would be the cheapest part of that decision by a wide margin.
One honest cost of that choice. The Silver parts cap memory at DDR4-2666 where the pricier Gold parts reach 2933, so the proposed build gives up a step of memory speed. It still ends up ahead on the number that matters, because eight populated channels at 2666 move more than double what two channels of faster memory manage on the quoted platform.
The same care applies to how the memory is arranged, not just its rating. Each processor reads from eight memory channels, so this machine has sixteen, and bandwidth follows the number actually populated. The same 128GB can be supplied as a few large modules or several smaller ones at broadly similar cost. A small number of large modules is the cheapest looking option and the slowest, because it leaves most of the channels idle. The proposed build uses eight modules spread across both processors, which takes the bandwidth benefit and still leaves eight slots free, so capacity can be doubled later without discarding anything.
Finally the SQL Server line here is licensed server plus user CALs rather than per core, so core count does not change the SQL cost at all. If it ever moves to per core licensing, this whole calculation changes sharply and must be redone before a processor is chosen.
| Component | Quoted build | Proposed build | Why it matters |
|---|---|---|---|
| Chassis | Dell PowerEdge R360, single socket 1U entry | Dell PowerEdge R450, dual socket capable 1U | Enterprise platform built for sustained multi tenant load. |
| CPU | Xeon E-2468, 8 cores / 16 threads, 2.6GHz | Two Xeon Silver 4309Y, 16 cores / 32 threads total, 2.8GHz | Twice the cores and threads of the quoted chip, split across two sockets, and exactly the sixteen cores one Windows Server licence covers. Two processors also double the memory channels. The quoted chip is still the faster individual core; this is the wider machine. |
| Memory | 64GB DDR5 ECC UDIMM, unbuffered | 128GB DDR4 2666 ECC RDIMM, both sockets populated | Both are ECC. Registered memory is what lets this platform reach eight channels and a far higher ceiling, which is where the real difference sits. The larger point is price: this generation sidesteps DDR5. |
| Disk controller | PERC H355 RAID controller, no write cache | HBA355i, a true host bus adapter | ZFS must see raw disks. The H355 can present them via its non-RAID mode, so this is workable rather than fatal, but a true HBA keeps RAID firmware out of the path entirely and is the cleaner part to specify at order time. |
| Write cache | None. Cacheless card under a SQL Server workload | ZFS ARC in host RAM, optional PCIe NVMe write log | ZFS caches reads in RAM, which is far faster than a controller cache. The cache size is a tunable with a conservative default, so size host memory with the cache in mind rather than assuming it takes what it needs. |
| Storage | 3 or 5 of 8 bays filled. 960GB usable on the cheaper estimate, 1.9TB on the other | All 8 bays filled with 800GB SAS SSD, 3.2TB usable, plus a separate mirrored boot card | SAS rather than SATA for the database pool: higher write endurance, dual ported, and better behaved under the deep concurrent queues a database and ten desktop sessions produce. A dedicated boot pair keeps every front bay free for data. |
| Power | Redundant PSUs | Dual hot plug Platinum PSUs | Higher efficiency with thermal monitoring under continuous load. |
The quoted platform is a current generation server, which means DDR5, which means buying memory at a moment when DDR5 supply is tight and pricing reflects it. The previous enterprise generation takes DDR4 Registered ECC instead, which is cheaper right now and is also the more resilient part: registered memory buffers the address lines so the host stays stable under sustained multi tenant load, which is exactly what a virtualized server does all day.
Both estimates list Windows Server 2025 Standard Software qty 2. Whether that is right depends entirely on a decision the estimate never states.
A Standard license covers up to 16 physical cores, and the quoted CPU has 8, so if this machine is being built the way the migration notes describe, as one Windows install on bare metal, one license covers it and the second is doing nothing. But Standard also grants two virtual machines per license. The moment the server is virtualized and Active Directory, SQL, Remote Desktop and file services become separate VMs, four Windows instances need two licenses, and quantity 2 is exactly right.
So this is a question, not an error: ask which one they are building. If they are planning to virtualize, that is the right instinct and it should be written into the proposal rather than left implied in a licensing quantity. For the avoidance of doubt, the build proposed in this document runs four Windows VMs and therefore carries two licenses of its own. The vendor's combined licensing line is competitively priced. The savings argued for here are in hardware, not in Microsoft.
Two further licensing items need confirming before any purchase. The SKUs read as Server 2025 and SQL 2025, which should be pinned to exact editions and channels in writing. And the CAL count sits at a round 30, which is worth checking against actual headcount rather than accepting as a default.
The quoted disk controller is a PERC H355: a RAID controller with no write cache. Under database write load it has nothing to absorb bursts with, so writes queue against the drives directly. That is a real bottleneck, and it is specified in the same estimate that buys SQL Server Standard with 15 user CALs. The software says this machine runs a real database. The controller says it does not.
The storage layout compounds it. Three to five SATA SSDs carry the host OS, the databases and the file shares together, so OS activity and database I/O contend for the same drives. A dedicated mirrored boot pair moves the operating system onto its own devices and leaves every front bay for data.
Worth separating two words that get used as if they were alternatives. SAS is an interface. SSD is the media. A vendor offering "SAS or SSD" is really offering SAS spinning disks, SATA SSDs at 6Gb/s, or SAS SSDs at 12Gb/s. For a database the answer is SAS SSD, specified as mixed use rather than read intensive, because a database with ten concurrent users rewrites the same pages continuously and read intensive drives are rated for a fraction of the daily writes.
Two things about the 12Gb/s number, so it is not oversold. It is per drive sequential headroom, and a database is mostly small random operations that never approach it. And drives are also sold at 24Gb/s for roughly two and a half times the price, which on this generation of backplane buys nothing at all, because the backplane runs at 12Gb/s and the drive simply negotiates down.
This is the part of both estimates that deserves the most attention, and it costs the least to fix.
| Build | Drives | Bays used | Usable | Layout |
|---|---|---|---|---|
| Estimate #4593 | 3 | 3 of 8 | 960GB | One mirrored pair plus a hot spare, as written on the estimate |
| Estimate, 5 drive variant | 5 | 5 of 8 | 1.9TB | Two mirrored pairs plus a shared spare, assumed to match the 3 drive layout. Worth confirming |
| Proposed | 8 | 8 of 8 | 3.2TB | Four mirrored pairs, with the operating system on its own boot card |
Bay counts assume the eight bay 2.5in backplane, which the Dell configuration ID requested in the plan below will confirm. On that basis the cheaper estimate puts 960GB of usable space into a $34,800 server. That is the working capacity available to a domain controller, a SQL database, the file shares and ten Remote Desktop users, and it is why the vendor's own note says the storage "may or may not be enough long term".
Smaller drives are the better answer here, not larger ones. Eight 800GB drives cost less than five larger ones and give four mirrored pairs instead of two, and in a mirrored pool the number of pairs is what determines how much random database work the array can do at once. More spindles, more usable space, lower cost. Putting the operating system on a separate mirrored boot card rather than on the data drives is what makes all eight front bays available in the first place.
Mirrors rather than a parity layout, because parity punishes exactly the small random writes a database produces. If capacity sizing later comes back larger than expected, the same eight bays take bigger drives without changing anything else about the design.
This is worth getting right at order time, because changing it later means opening the server. ZFS is not a filesystem that sits politely on top of a RAID array. It expects to own the physical disks: it checksums every block, and when it finds a bad one it repairs from a good copy. It can only do that if it can see the individual drives.
So the controller wants to be a plain host bus adapter, which hands the raw disks over untouched. The quoted H355 is not a dead end: it supports non-RAID passthrough natively and ZFS will work with it. But it remains a RAID controller running the megaraid_sas driver, so the disks still reach ZFS through a RAID firmware layer. The correct part is a true host bus adapter such as the HBA355i, which runs the mpt3sas driver and puts nothing between ZFS and the drives. That is what keeps drive health reporting, error handling and hot swap behaving the way ZFS expects.
One controller to avoid on this build: the H755, which costs several times more than the host bus adapter for a large write cache this design deliberately does not use.
The cache question then answers itself. ZFS uses host RAM as its read cache, which is far faster than a controller cache, and a PCIe NVMe write log can absorb the synchronous writes SQL Server cares about. One caveat worth stating plainly: that cache does not simply take whatever memory is free. It ships with a conservative default limit and is tuned deliberately, so host memory should be sized with the cache budgeted in rather than assumed. That is a further argument for 128GB on a platform where DDR4 is inexpensive, which is what the proposed build carries.
The migration plan as quoted installs Windows Server on bare metal and runs the domain controller, the databases, the file shares and Remote Desktop from that single install. Every one of those is a single point of failure for all the others. Installing Proxmox VE first, then running those roles as separate virtual machines changes what a bad day looks like. Proxmox itself is free, though note the Windows licensing does change: Standard covers two virtual machines per license, so four Windows VMs need two licenses. That is already carried in the proposed figures above.
Everything above compares one server against one server, and on that comparison the proposed build saves money outright. But the money involved opens a door neither estimate mentions. A second machine, same generation, lighter specification, turns a single point of failure into a pair. It does not save money against the cheaper estimate. It lands between the two, and buys resilience with the difference.
Proxmox includes clustering and scheduled replication at no licensing cost. Virtual machines run on node one and replicate to node two on a schedule measured in minutes. If node one fails, the machines can be started on node two in minutes. Not a restore, not a rebuild, not a day of downtime with everyone sent home.
One detail worth stating rather than glossing: a two node cluster has no majority, so fully automatic failover needs a third vote. That is a quorum device, which is a tiny service running on any existing always on machine, not a third server. Without it, failover is a deliberate manual start, which is still minutes rather than days. With it, it is automatic.
Both machines are quoted complete: chassis, processor, memory, controller, boot card, network, redundant power, rails, out of band management, drives and three years of 24x7 support with next business day on site response. Nothing is left to add later.
Both are 1U rack mount, so this fits the existing rack rather than putting a tower on the floor. Both are available through the same refurbished channel, alongside the R650, R750xs and R550 if more expansion is wanted later.
| Added on top of Option 1 | Estimated range | Note |
|---|---|---|
| Second server, replica target | $9,800 | Dell PowerEdge R450, configured at published list including its own three year 24x7 support. Holds replicas and runs the workload only when node one is down |
| Windows licensing, second host | $1,900 to $2,400 | Standard has no free failover right. Budgeted here; confirm the exact requirement before purchase |
| Option 1 total, one server | $21,300 to $23,500 | |
| Option 2 total, two servers with failover | $33,500 to $36,300 | Between the two vendor estimates, for two machines instead of one |
The server is listed at $19,650 on one estimate and $25,500 on the other, for what is described as the same chassis with a different drive count. That is $5,850 for two additional 960GB SSDs. Two separate things sit inside that number and the estimate separates neither: the vendor's margin, and the genuinely elevated cost of a current generation platform right now, which is largely DDR5 memory pricing. Both are real, and we cannot reconcile the total without an itemized build sheet. Ask for the Dell configuration ID and a line by line breakdown, with the memory on its own line, because that is the part that changes the buying decision.
Estimate #4593 specifies three 960GB drives as a RAID 1 mirror with a hot spare. A mirror of two drives plus a spare is 960GB usable. That is the whole storage capacity of a $34,800 server, for a domain controller, a SQL database, file shares and ten Remote Desktop users. The five drive variant reaches 1.9TB. Both sit in a chassis with eight bays, so the cheaper build leaves five bays empty and the other leaves three. Filling those bays with smaller drives costs very little and is the single cheapest performance decision available here, because every additional mirrored pair adds usable space and another set of drives to spread the database load across.
One Standard license covers up to 16 physical cores, and the quoted CPU has 8, so on the bare metal install the estimate describes, the second license is not doing anything. It becomes necessary the moment you virtualize, because Standard grants two virtual machines per license and the architecture below runs four. So the quantity may well be right, for a reason the estimate does not give. Ask which it is. If the vendor is planning to virtualize, that changes the conversation and should be on the page. Note that our own proposed build carries two licenses for exactly this reason, and the vendor's combined licensing line is competitively priced: the savings in this proposal are in hardware, not in Microsoft.
No speed, no module count, no rank. On this platform that is not a detail: the processor has two memory channels, and the same 64GB behaves differently depending on how it is delivered. Two modules run at the platform's full rated speed; four modules drop a step. Rank matters again on top of that. We have assumed the best case for the quoted build throughout this document, which means assuming they have specified the faster arrangement. It is worth asking, because on an entry platform bought during a memory shortage the cheapest parts that satisfy the line are the likely ones, and the line as written would accept them.
There are two defensible ways to build this. Bare metal Windows on hardware RAID, which wants a controller with a battery backed write cache. Or Proxmox with ZFS, which wants the opposite: raw disks handed straight to the operating system. The quoted PERC H355 is a RAID controller with no write cache. For the first design it is the slow option. For the second it can be made to work through its non-RAID mode, but a true host bus adapter is the cleaner part. Either way this is a configurator default rather than a choice, on the component that decides database performance.
The entry platform locks the purchase to DDR5 UDIMMs while that is the most expensive memory on the market. The previous enterprise generation takes DDR4 Registered ECC, which is both cheaper today and the more resilient memory type for a virtualization host.
Estimate #4593 reads: "this may or may not be enough storage long term, but will definitely be enough for now." That is an honest sentence and it points at a real gap, but it applies to this proposal too, in the sense that nobody has measured anything yet. The proposed pool is 3.2TB usable, roughly three times the cheaper estimate, and it fills every bay so there is headroom in drive size rather than drive count. But the answer is still not a bigger guess, it is a number. Measure current data volume and annual growth, then specify. The same eight bays take larger drives if the number comes back higher.
With 3 to 5 SATA SSDs carrying host OS, databases and file shares together, OS activity competes with database I/O for the same drives and the same controller queue.
"Server 2025" and "SQL 2025" appear as line items. Confirm exact edition, SKU and licensing channel in writing before any money moves, because the license is the part you cannot swap later.
Active Directory, SQL Server, Remote Desktop Services and file shares on a single bare metal Windows install means one reboot, one failed update or one crash takes down authentication, the database and everyone's desktop at the same time.
In order, cheapest and fastest first. The first two items cost nothing and can happen before any decision about who supplies the hardware.
Nothing here says the vendor is acting in bad faith. It says one line item carries most of the cost, has no itemization behind it, and cannot be reconciled from what the estimate shows. That is a fair question to put to them, and their answer will tell you a lot. If the answer is not satisfying, the alternative build in this document is a real one and we can source and price it properly.