A cheap rack-mounted AI server - a used PowerEdge/ProLiant + Tesla, or why "free" cooling in a rack is not free at all

A short answer before I expand on it: if you have a rack cabinet (your own server room or space in a colocation facility), a used 2U server (Dell PowerEdge R730/R740 or HPE ProLiant DL380 Gen9/Gen10) with a used Tesla P40 or V100 is the most predictable way to get a "server-grade" host for local AI. But I am about to demolish a myth that I used to repeat myself: that in a rack "you get cooling for free, because the chassis blows air through the card". The truth is less convenient, and that is what this post is about.

This is a companion piece to "How much does a cheap AI server cost in Poland" (workstation). That one was about putting a passive Tesla on a desk, where you have to bolt a fan onto the card, because otherwise it will suffocate. This one is about the same question ("how much does it really cost and what do you need to add besides the card"), only in a cabinet. The bill of materials is different and, more importantly, so are the main pain points.

I am writing as someone who sells these cards and sees what customers buy them for, so two caveats to start with. First: yes, we trade in used Teslas (full disclaimer at the end). Second: a rack is not "a better version" of a workstation. It is a different compromise. I will show when it pays off and when I advise against it, including arguments against our own range.

Why a passive Tesla "wants" a rack, and where the catch is

Let us start with what actually works better in a rack. The Tesla P40 and V100 PCIe cards are passive: they have no fan of their own. The P40 has a TDP of 250 W, so at full load almost all the energy the GPU draws ends up as heat that has to be removed, and the card relies on the chassis to do that (NVIDIA Tesla P40 Product Brief). That is not a flaw; it is a design decision: the card was designed to work in a server in which a row of high-static-pressure fans pushes air through the heatsink from the front to the back of the chassis.

A typical workstation, however, does not have the kind of forced airflow through the GPU heatsink that the design of a passive Tesla assumes, so you have to attach a shroud with a fan to the card. In a rack server the airflow is built into the design, and this is where the myth is born that "you put the card in and do nothing". Well, not quite.

Take the HPE ProLiant DL380 Gen9, one of the most popular used servers for GPUs. To fit a dual-slot Tesla into it at all, you need the Primary GPU riser (719076-B21) and the Graphics Enablement Kit (719082-B21): mechanical parts and cabling required to install supported dual-slot cards. But that is not all: passive cards such as the P40 or P100 additionally require the High Performance Fan Kit (719079-B21), a set of six higher-performance fans in place of the standard ones. Without it a passive Tesla will not get enough air. And one more condition from the QuickSpecs that can come as a surprise when buying a used server: for the P40 and P100 the QuickSpecs allow only E5-2600 v4 CPUs (the V100 entry says v3 only, which looks like an error in the document, so confirm with the seller), and the P40 is limited to a 25 °C inlet in the 12LFF chassis (HPE QuickSpecs, confirmed with parts distributors).

And that is the heart of this section: in a rack you pay for cooling too; only, instead of bolting on a PLN 60 fan, you buy a factory fan kit and a riser. The difference is qualitative (the solution is server-grade, predictable and integrated with the chassis), but it is not zero in cost. Whoever tells you that "in a rack cooling comes free" has not read the parts list to the end.

The question you are probably already asking yourself: if I pay for cooling either way, why a rack at all? We will come back to that; first, the base.

The base: which used 2U server

You do not need the latest generation. You need a chassis with the power, the space and the firmware ready for a 250 W card. I will split this into two tiers, because they answer different needs.

The "entry" tier: one, maybe two cards. These rule here:

• Dell PowerEdge R730 (2U, 13th-generation PowerEdge, Haswell/Broadwell): the cheapest sensible choice on the secondary market, a sensible base for one or two P40/V100 cards provided the requirements of the specific configuration are met. The Owner's Manual sets firm conditions (more on them below): 1100 W power supplies (each) in a redundant configuration for supported GPU configurations, a GPU enablement kit, a maximum of two dual-slot cards.
• HPE ProLiant DL380 Gen9 (2U): HPE's equivalent. It requires the Primary GPU riser (719076-B21) + the Graphics Enablement Kit (719082-B21) plus the High Performance Fan Kit (719079-B21) for passive cards; a second card also needs the Secondary GPU riser (719073-B21) and both processors. HPE requires at least one 1400 W power supply (720620-B21) per card over 150 W, and the QuickSpecs list the P40 and P100 as supported only with E5-2600 v4 processors.

The "scalable" tier: three or more cards. If you are thinking of more than two accelerators, or of headroom for the future:

• Dell PowerEdge R740 (2U, Xeon Scalable / Skylake–Cascade Lake): a newer platform than the R730, supporting up to three 300 W dual-slot cards (or six 150 W cards). For a GPU configuration Dell requires power supplies of 1100 W or more set to non-redundant mode, plus the GPU kit with six high-performance fans and low-profile heatsinks; 2000 W and 2400 W supplies are also available, and PSU capacity must be sized for the complete configuration (Dell R740 Installation & Service Manual / Technical Guide). A more expensive used server, but it gives you room for a third card. An example of such a platform is the Dell R740 prepared for up to three V100 cards.
• Supermicro SYS-2029GP-TR (2U, the same Xeon Scalable as the R740): Supermicro's equivalent from the same tier, and in fact a "beefier" one: six PCIe 3.0 x16 slots, up to six GPUs, redundant 2000 W 80 Plus Platinum PSUs (Supermicro specification). That is the choice when you want the maximum number of cards in 2U and do not mind a less "corporate" service ecosystem than Dell/HPE.

A simple rule for choosing: one or two cards → R730/DL380 Gen9 (cheaper), three or more → R740 or Supermicro 2029GP-TR (more expensive, but there is somewhere to plug them in and something to power them with). Do not buy an R740 "for later" if you will realistically install one P40: you are paying extra for slots and 2000 W that you will not use.

A note that saves you from the most common mistake: you cannot settle compatibility from the model name alone. Two "R730" servers can arrive with different PSUs, with or without a GPU riser, with standard or high-performance fans. The R730 Owner's Manual lists explicitly what has to be met for dual-slot GPUs:

• 1100 W power supplies (each) in redundant mode (with the most power-hungry cards, e.g. the K80, Dell even tells you to set the PSUs to non-redundant mode to deliver the full power);
• a GPU enablement kit: low-profile heatsinks, power cables for the GPU cards, EMI blanks for empty slots;
• a maximum of two dual-slot cards, and a two-card configuration requires riser 3;
• both processors installed, with a TDP of ≤135 W with the low-profile heatsinks from the kit;
• inlet air temperature ≤30 °C.

This is the same problem that I described for a workstation as "Above 4G and the BIOS": here it comes back in the form of "PSU, riser, kit, inlet temperature". When you buy a used server for a GPU, you look at the specific configuration of the unit, not at the name.

Price ranges for the base (used, a complete 2U server without a GPU, Polish secondary market, Buy Now, read on 2026-09-21): in the entry tier, an R730 from ~PLN 1,300 for a complete 2× CPU + 32 GB configuration (with 16 GB even from ~PLN 900, but the cheapest units usually come with 2× 750 W, so for a GPU add the cost of replacing the PSUs with 1100 W ones), a DL380 Gen9 from ~PLN 2,200 (2× E5 v3 + 32 GB; with the v4 processors required for a P40/P100 from ~PLN 2,500); in the scalable tier, an R740 from ~PLN 2,200–2,600 (2× Xeon Gold/Silver, 16–64 GB); better-equipped configurations (Xeon Gold/Platinum, 128–256 GB of RAM, SSDs) reach PLN 10,000–13,000. The Supermicro 2029GP-TR practically does not appear on Allegro in Poland: get prices from distributors/eBay. The amounts are indicative and volatile: they depend on the CPUs, RAM and PSUs; check them just before buying.

If, instead of assembling the platform yourself, you need a ready-made solution, we also prepare Dell PowerEdge R740 AI servers with 1–3× Tesla V100 32 GB.

Power: here the rack really does win

In a workstation, power needs attention because you have to watch the connector: a Tesla takes power through a CPU 8-pin (EPS) connector, not a PCIe 8-pin one, and in an ordinary PC it is easy to mix them up (P40 Product Brief).

In a server the problem is solved at the factory. In GPU configurations the appropriate enablement kit (Dell/HPE) supplies the required parts, including the correct power cabling routed from the motherboard/PSU to the card. Redundant hot-swap power supplies give you spare power and, in a correctly configured R730 or a DL380 Gen9 with one card, resilience to the failure of one PSU; in an R740 with GPUs Dell requires non-redundant mode, and in a DL380 Gen9 with two cards HPE requires a 1400 W supply per card with at most two supplies, so in those configurations the second supply adds power, not protection. You do not fiddle with adapters: you take the cable from the kit.

A commercial aside, to be fair: we sell our cards with the correct power adapter included. In a workstation that is a real saving. In a rack with the original kit you will most often use the cable from the kit, so the adapter in the box is a spare, not a necessity, but it is better to have it in the box than to go looking for one on a Wednesday at 23:00.

The only thing you have to check on the power side: whether the PSU has headroom. The P40 and V100 PCIe are 250 W each, plus two Xeons, drives and fans. That is why Dell requires redundant 1100 W power supplies in supported R730 GPU configurations other than the K80, which requires non-redundant mode, not the base 495 W that used servers are sometimes sold with, and HPE requires one 1400 W power supply per card in a DL380 Gen9, so a used unit with 500 W or 800 W supplies needs new ones.

Bill of materials: rack vs workstation

I will set it out side by side, because this is the heart of it. The same card, two contexts. Card prices from the shop's website (read on 2026-10-27); the secondary-market items are indicative ranges from 2026-09-21 (sources in the Sources section): volatile, check them before buying. The noise, power-draw and colocation values in the table and below are estimates for an example configuration (2× CPU + 1–2 cards under load), not a characteristic of every R730/R740 unit: the real figures depend on the CPUs, the number of cards, the load and the cooling.

Item	Workstation (desk)	2U rack server
Base (chassis + motherboard + CPU + RAM)	used Dell/HP/Lenovo workstation ~PLN 1,500–2,500	R730: from ~PLN 1,300; DL380 Gen9: from ~PLN 2,200; R740: from ~PLN 2,200–2,600 (Polish secondary market, Buy Now, 2026-09-21)
P40 24GB card	PLN 1,695 net / 2,085 gross	the same
V100 16GB / 32GB card	PLN 1,711 / 4,415 net	the same
Card power	CPU 8-pin adapter: included with our cards (PLN 0)	cable from the GPU kit / redundant 1100 W PSUs
Card cooling	shroud + fan, ~PLN 60–130	High Performance Fan Kit (HPE) / GPU kit with high-performance fans (Dell R740; the R730 keeps its standard fans); cost depends on the model and kit contents: check a current quotation
Riser / enablement	not applicable	extra (riser + GPU cables): ~PLN 200–400
Noise	manageable on a desk	very high (roughly ~60–75 dB(A) under load): not something you can put in an office
Power draw (idle→load)	moderate (~80 → ~350 W)	~180–220 W → ~600–750 W* (whole platform)
Space	under the desk	2U in a cabinet / illustrative colocation cost ~PLN 480–525 net/month if the provider agrees to bill for an actual draw of 650–750 W

* An estimate for an example 2× CPU + 1–2 GPU configuration (the whole platform, not the card alone): not a measurement of a specific server. The actual power draw depends on, among other things, the CPU model, the number of GPUs, the load and the PSU efficiency.

Note the "card cooling" row. In a workstation it is a one-off ~PLN 60–130 for a shroud. In a rack it is not zero: it is a factory fan kit and a riser, in a typical used-server configuration usually more expensive than a shroud, just in a different form. A rack does not remove the cost of cooling. It swaps a do-it-yourself fan for a factory (paid) kit, and adds noise, electricity and space on top.

Three costs nobody talks about

Noise. The High Performance Fan Kit is not a random name: under GPU load these fans sound like a vacuum cleaner that never stops. A 2U server with a Tesla under full load cannot be placed in a room where someone is working. It needs a separate room, a soundproofed cabinet or colocation. That is the most common reason I advise against a rack for someone who "has a free spot in the office".

Electricity. Two server processors, a row of fans, redundant PSUs: a 2U server at idle draws clearly more than a workstation before you even count the card. With continuous operation (and a local model usually means continuous operation), the difference on the bill is real. At an assumed energy-and-distribution price of ~PLN 1.2–1.3 gross/kWh (a calculation input rather than a uniform 2026 G11 rate; a business using, for example, C11 should substitute its contracted energy and distribution prices and account for fixed charges), a 2U server under continuous load (~650 W drawn by the whole server, not the card alone) comes to around PLN 560–610 gross/month, and even at idle (~200 W) ~PLN 170–190/month. This is an estimate (we did not measure the power draw), but the order of magnitude is solid: you pay for it around the clock, seven days a week.

Space / colocation. If you have a server room, the cost is a slot in the cabinet. If not, you pay monthly for a U in colocation plus power. That turns a one-off expense into a recurring cost that has to be included in the TCO just like the cloud. In Polish data centres this is usually charged as space + power, e.g. from ~PLN 90 net for 1U plus ~PLN 46 net for every 100 W of power supply; for a 2U server drawing ~650–750 W this gives roughly PLN 480–525 net/month, assuming twice the 1U space rate and a proportional charge for agreed actual draw. This is a calculation scenario, not a confirmed metered-power offer. Billing on the sum of the PSU ratings (2× 1100 W) would instead give about PLN 1192 net/month, so agree the billing basis before signing (Sprint Data Center, post of 2026-06-23).

This is the moment when "cooling solved for free" shows the full bill.

The card: P40, V100, and the SXM2 trap

The cards do not change compared with a workstation. P40 24GB (PLN 1,695 net): budget and a lot of VRAM, without Tensor Cores. V100 16/32 GB (PLN 1,711 / 4,415 net): faster inference, long context, Tensor Cores. The choice is set out in the companion article, so I will not repeat it.

One trap specific to the server world. The V100 also exists as SXM2, and that is not a PCIe card. SXM2 is a module that plugs into a dedicated socket on the motherboard, with NVLink, a higher power budget (300 W against 250 W for PCIe) and intended for DGX/HGX-type chassis (NVIDIA Volta Architecture Whitepaper). You cannot fit it into an ordinary R730 with a PCIe riser: it needs a motherboard with an SXM2 socket (servers such as the Dell C4140, HPE Apollo, Supermicro SXM2). The tempting "cheaper V100 SXM2" at auction is a road to nowhere if you do not have a chassis for it. I cover this trap in a separate "SXM2 vs PCIe" post (I will add the link once it is published); for now, a warning: for a 2U server you take the V100 PCIe, not the SXM2.

When a rack, and when a workstation

Honestly, because this is a decision, not a ranking:

• A rack makes sense if: you already have a cabinet/server room or colocation, you need continuous 24/7 operation, you want "server-grade" predictability and, where the GPU configuration supports it, redundant power (the R740 requires non-redundant mode with GPUs), you are planning several cards (two in an R730/DL380, three in an R740, up to six in a Supermicro 2029GP-TR), and noise is not a problem (a separate room).
• A workstation wins if: you are putting it in an office/study/home, one card is enough for you, you care about quiet and low idle draw, and you do not want to pay a recurring cost for space. Then the workstation route is cheaper to run, despite the shroud with a fan.

Choosing a workstation precisely to avoid the noise and the colocation is a defensible choice, not a compromise "because a rack is unaffordable". A rack is not a reward for a bigger budget. It is the answer to a specific requirement: continuity, redundancy, several cards.

What you maintain afterwards

A rack is not "set it and forget it". Keep the firmware (iDRAC/iLO) current and follow the server vendor's guidance. On Dell PowerEdge 13G servers, including the R730, a card recognised as third-party PCIe receives the default additional cooling response, which can increase fan speeds. Lack of validation, such as for a V100 in an R730, does not establish how firmware will classify that card. With a passive Tesla, leave that response enabled unless the required airflow and temperatures have been verified, because the card depends on server-supplied cooling. You monitor the card's temperatures (nvidia-smi) and the fan speeds. And you remember that redundant PSUs protect against a power supply failure, not a card failure: data backup is a separate topic. You also keep an eye on the condition from the Owner's Manual: inlet air temperature ≤30 °C; a server room without air conditioning can exceed that limit in summer.

Summary: cooling does not disappear, it just changes the invoice

To finish, as simply as possible. In a workstation you fight the heat with your bare hands: you bolt a fan costing a few tens of PLN onto the card so that the passive Tesla does not suffocate on the desk. In a rack you do not fight that battle yourself: a row of server fans does it, just as NVIDIA's passive design assumes. But that does not mean you do not pay for it: you pay with a factory fan kit, a riser, redundant PSUs, and then with electricity, noise and space.

There is no free cooling. There is only the choice of what you want to pay for: once, for a fan on the desk, or later, for a factory kit, higher power draw and space in a quiet, separate room. Choose consciously, not because "a rack sounds more serious".

---

Disclaimer: Agile Store sells used NVIDIA Tesla accelerators (P40, V100 and others): the cards discussed are in our range, and we sell them with the correct power adapter included. We write from a seller's perspective, but we give the figures and the limitations honestly, including those that argue for the cheaper route (a workstation) instead of a rack. Prices are indicative, card prices per store.agilesoftware.pl (as of 2026-10-27): check the current ones before deciding.

Sources

1. NVIDIA Tesla P40 GPU Accelerator: Product Brief (PB-08338-001): TDP 250 W, passive cooling, power via a CPU 8-pin connector. https://images.nvidia.com/content/pdf/tesla/Tesla-P40-Product-Brief.pdf
2. NVIDIA Tesla V100 PCIe: Product Brief (PB-08744-001): V100 PCIe 250 W, bidirectional passive heatsink. https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-product-literature/Tesla-V100-PCIe-Product-Brief.pdf · NVIDIA Volta Architecture Whitepaper (WP-08608-001): SXM2 300 W, NVLink, DGX/HGX form factor. https://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf
3. Dell PowerEdge R730 Owner's Manual: GPU card installation guidelines: redundant 1100 W PSUs, GPU enablement kit (low-profile heatsinks, power cables, EMI blanks), ≤2 dual-slot cards, riser 3, CPU ≤135 W, inlet ≤30 °C. https://www.dell.com/support/manuals/en-us/poweredge-r730/r730_ompublication/gpu-card-installation-guidelines
4. HPE ProLiant DL380 Gen9: QuickSpecs: Primary GPU riser 719076-B21 (slot 1), Graphics Enablement Kit 719082-B21 (heatsinks + cables for dual-slot GPUs), P40/P100 support only with E5-2600 v4 CPUs; at least 1× 1400 W PSU (720620-B21) per card over 150 W; Secondary GPU riser 719073-B21 for slot 2. https://www.hpe.com/psnow/doc/c04346247
5. HPE ProLiant DL380 Gen9: High Performance Fan Kit (719079-B21): a set of 6 fans replacing the standard ones; required for "certain graphics accelerators" (passive Tesla P40/P100): QuickSpecs as above. Part page: https://expresscomputersystems.com/products/719079-b21
6. Dell PowerEdge R740: Installation & Service Manual / Technical Guide: up to 3 dual-slot 300 W cards (or 6× 150 W); PSUs ≥1100 W in non-redundant mode for GPUs; PSU options up to 2400 W; Xeon Scalable platform; GPU enablement (high-performance fans + cables). https://www.dell.com/support/manuals/en-us/poweredge-r740/per740_ism_pub/installing-a-gpu · https://i.dell.com/sites/csdocuments/shared-content_data-sheets_documents/en/aa/poweredge_r740_r740xd_technical_guide.pdf
7. Supermicro SuperServer SYS-2029GP-TR: 2U, dual Xeon Scalable (Skylake/Cascade Lake), 6× PCIe 3.0 x16 (up to 6 GPUs), redundant 2000 W 80 Plus Platinum PSUs. https://www.supermicro.com/en/products/system/2u/2029/sys-2029gp-tr.php
8. Card prices: Agile Store shop (read on 2026-10-27): Tesla P40 24GB PLN 1,695 net / 2,085 gross; V100 16GB 1,711 / 2,105; V100 32GB 4,415 / 5,430. https://store.agilesoftware.pl/en/products
9. Secondary-market and running-cost ranges (read on 2026-09-21, volatile):
   • Used servers: Allegro (Buy Now, sorted by price): a complete R730 (2× E5 v3/v4, 16 GB) from ~PLN 900, with 32 GB from ~PLN 1,300 (the cheapest usually with 2× 750 W PSUs); DL380 Gen9 (2× E5-2620 v3, 32 GB) from ~PLN 2,200, with v4 CPUs (2× E5-2680 v4, 32 GB) from ~PLN 2,500; R740 (2× Xeon Gold/Silver, 16–64 GB) from ~PLN 2,200–2,600, 128–256 GB configurations ~PLN 10,000–13,000; Supermicro 2029GP-TR: no listings. https://allegro.pl/listing?string=serwer+dell+r730 · https://allegro.pl/listing?string=serwer+hpe+proliant+dl380+gen9 · https://allegro.pl/listing?string=serwer+dell+r740
   • Energy: G11 all-in ~PLN 1.2–1.3 gross/kWh in 2026 (active energy ~0.64 + distribution ~0.62; per distribution system operators 1.19–1.32), enerad.pl summary (updated 2026-01-12). https://enerad.pl/ceny-pradu/
   • Colocation: 1U from ~PLN 90 net + ~PLN 46 net per 100 W of power (Sprint Data Center, 2026). https://www.sprintdatacenter.pl/blog/ile-kosztuje-kolokacja-serwera-w-2026-model-1u-op-ata-za/
   • GPU enablement kit / fan kit / risers: parts market (eBay/distributors), ~PLN 300–800 for a set.