A quick note before you read: In light of the ongoing AI bubble, I have provided a full cost breakdown of all parts of when I bought them. I am currently in the process of building a new AI cluster — a big upgrade . When it’s done, I will post about it. I hope this can be a guide for you, to do it right the first time, learn from my mistakes, and save you literally dozens of hours of your life and hair loss because I didn’t have something like this.

Building a Local AI Cluster from Scratch

I built an AI supercomputer. A purpose-built machine for running LLMs, generating images, fine-tuning open source models, and conducting the kind of AI research that you can’t do on a MacBook Pro. It’s been a long, incredible AI journey for me. I am writing this as a series because I can’t fit everything in one blog post. I built my first AI cluster in the fall of 2024. When I first started, I knew pretty much nothing about hardware and infrastructure. Since then, I have caught the hardware bug and now I can’t stop. It is now April 2026 as I write this. In a way, this is a post mortem, writing 1.5 years after using my GPU cluster. I learned a ton along the way and made a ton of mistakes. The goal of this blog post isn’t just to talk about my build – it will also be a guide for anyone out there who is brave enough to enter the world of hardware and real AI.

Why Build?

I’m a Senior Data Scientist who has spent the last two years moving deeper into applied AI research and local AI systems. The work I’m doing — fine-tuning LLMs, running inference on 13B+ parameter models, generating images with Stable Diffusion, and experimenting with quantization techniques — requires hardware that most employers do not hand you and no cloud provider gives you for free.

When I first thought about doing this project, I didn’t really think about renting vs owning in terms of cost. My soul just wanted to do it. But the math does math and it matters now more than ever. The trend is clear: ALL compute is getting more expensive and harder to get, combined with the fact that the large frontier labs are no longer going to subsidize their agentic coding platforms. This is one of the many reasons why local AI is more important today than ever before. AI overall is going to get A LOT more expensive on every front.

The Math

A single RTX 3090 on RunPod costs roughly $0.46/hour USD on their Secure Cloud tier (their cheaper Community Cloud tier runs $0.22/hour, but it’s peer-to-peer spot capacity — fine for experiments, not production). At $0.46/hr, that’s $331/month per card, running 24/7. Two of them, so you can run bigger models? $662/month. In under 2 months you’ve already surpassed the cost of a single used RTX 3090: ~$1,050 USD per Best Value GPU (April 2026) and in under 7 months, you’ve spent more on cloud compute than it costs to own the hardware outright. When you own it – there is no meter running. You can experiment at 3 AM without worrying about a bill or if you can even access a GPU, or if AWS or Cloudflare has yet another outage. You can leave training to run over the weekend, you can iterate on ideas without the psychological overhead of watching dollars tick up. Once paid for, the only cost is your time and electricity bill. The price of all consumer GPUs has rapidly climbed up, and it will likely continue to do so until the AI bubble pops.

Path	Cost for 48 GB VRAM	Break-Even vs This Build
This build	$4,283 (hardware, pre-tax)	Forever — you own it
RunPod 2x 3090 (Community Cloud)	~$317/month	~14 months
RunPod 2x 3090 (Secure Cloud)	~$662/month	~7 months
Lambda Labs A100 80GB (1x card)	~$1,793/month	~3 months

I wanted to think freely – that requires owning the compute.

The Build

Looking back on this build, as I transition to the new one, I will do this part with some hindsight thinking. I will blend my current mindset at the time based on what I was thinking when I got these parts, as well as tidbits of warnings and recommendations after making minor and not so minor mistakes. I will not give too much feedback post build, as this post is the first part of an ongoing series.

The GPUs — RTX 3090 Founders Edition (2x)

The RTX 3090 is by far the best value proposition in AI compute right now and it’s not even close. That said, it has its caveats but price/compute is what matters for local AI, and not everyone has Jeff Bezos money. 3090s run on the Ampere architecture — solid and beyond heavily battle-tested.

The specs: 10,496 CUDA cores, 328 third-gen Tensor cores, 936.2 GB/s memory bandwidth, 35.6 TFLOPS FP32 boost, 142 TFLOPS FP16 Tensor dense (FP16 accumulate; ~71 TFLOPS with FP32 accumulate due to NVIDIA’s deliberate consumer-card throttle), a TGP of 350W, and 24 GB GDDR6X VRAM per card.

Two cards = 48 GB combined VRAM. That much VRAM lets you run models most individual GPUs can’t touch, although that’s changing fast thanks to quantization. A 13B parameter model at FP16 needs roughly 26 GB of VRAM. That won’t fit on a single 3090 unless quantized, but it fits comfortably across two cards with model parallelism. A single RTX A6000 (Ampere) has the same 48 GB, but it’s a professional card at $3,500–4,500 used (eBay, May 2026) — you get the same VRAM with dual 3090s for a fraction of the price.

The 3090’s other killer feature: 3rd-gen NVLink at 112.5 GB/s bidirectional (~1.8x PCIe 4.0 x16, apples-to-apples). The 3090 is the last affordable consumer card with NVLink support. NVIDIA killed it on the consumer line starting with the 4090 (Ada Lovelace), and didn’t bring it back for the 5090 (Blackwell). NVLink primarily accelerates training, not inference. Training requires constant gradient synchronization between GPUs at every backward pass, and that’s where NVLink’s bandwidth shines — 30–70% faster training vs PCIe alone. For inference, modern PCIe Gen 4 has plenty of bandwidth. I don’t have an NVLink installed on this build.

The Founders Edition card specifically was an intentional choice at the time. I did not understand the GPU market and what consitutes a GPU when I first embarked on this build, and I have many regrets. The FE (Founders Edition) cards are THICKKKK — they take up 3 PCIe slot spaces, which makes a dual setup very tight. Ideally you want at least one empty PCIe slot between non-blower GPUs because, by reference design, the cards take air from the bottom and blow it up above the card. So when you stack two of them, the top card gets a shit ton of heat. If the space between them is too little, you get serious thermal throttling on the top card, which not only cooks it but reduces its lifespan.

On top of the FE thickness, the 3090 itself is notoriously hot. Ampere was built on Samsung’s 8nm process — less efficient than the TSMC nodes the 4090 and 5090 use — and it runs at a 350W TGP with GDDR6X memory that already runs hot. The FE design makes it worse: memory chips on the back of the PCB regularly hit 100°C+ junction temps. The 3090 is essentially a space heater by reputation.

When I built this rig in fall of 2024, you could still buy an RTX 3090 at Micro Center. That was ideal. They checked cards for issues, you could add a warranty, and the price was the best compared to other channels. Now the price of RTX 3090s has blown up because of local AI demand. At the time of this post, they sit at ~$1,050 USD (Best Value GPU, April 2026). Looking back, the FE was the wrong type of 3090 to buy. Don’t make the same mistake. If you’re going dual 3090, get a 2-slot blower card instead — they exhaust heat out the rear bracket rather than dumping it onto the card above, and they still support NVLink. Best options: Gigabyte RTX 3090 TURBO 24G, ASUS RTX 3090 TURBO, or MSI RTX 3090 AERO 24G — all 2-slot, all NVLink-compatible.

One last thing: the 4090. Architecturally it’s the better card — Ada Lovelace, more AI TOPS, more tensor cores, better local AI kernel performance. But the price kills you at ~$2,000–2,500 used (eBay, May 2026), and the 4090 has no NVLink. A single 4090 is capped at 24 GB forever, while two NVLinked 3090s give you 48 GB for less money. There are Chinese-modded 4090s on the gray market with 48 GB GDDR6 swapped in, typically $4,000–$5,000 USD +, popular in the Chinese AI server scene.

GPU Cables

This is a very important topic to discuss. For reasons I do not understand, Nvidia has chosen the power pin connectors that they use. There are many issues with this – the dreaded pin burn aka your GPU either catches fire or starts smoking. This is a real issue – 5090s have it as well. This is why GPU cables are extremely important – it’s one thing to lose your GPU, but another to lose your house/lab. When the 3090s came out for FE cards, they shipped with a 2x 8-pin to 1x 12 pin adapter. You’ll see reports of melted cables online — it’s statistically rare, but it does happen. The 3090s had fewer issues compared to 4090s because 350 Watts vs the 450 Watts on the 4090 is a lot less wattage. The 12 pin was the precursor to the 12VHPWR melting crisis on the RTX 4090s (hundreds of documented cases — NVIDIA cited ~0.04% officially, while Gamers Nexus’s reporting from AIB sources put it closer to 0.05–0.1%). The stock adapter cable that came with the 3090 FE was weak. With 2x 8 pins and a 1x 12 pin, you now have three failure points to worry about. Rated at 30 insertion cycles (that’s what she said) – pins stretch and lose tension over time, creating air gaps. Loose fit is really the major issue with this. I cannot stress this enough: 1) always check for debris on the cables prior to inserting, as well as the holes where the pins go for debris 2) when you insert the cables/pins, make sure you hear the click noise – and then push in slightly further.

Also very important – your cable must be compatible with your PSU. This is the safest way to go: if you get a Corsair PSU, then you get Corsair GPU cables for the GPU and PSU to connect to each other. By using the Corsair cables to match my Corsair PSU, I have brand compatibility, and I don’t have to worry about the Nvidia FE adapter cable.

The rules, summarized:

• Never use a cable from a different PSU brand — pinouts aren’t universal, wrong brand = possible disaster.
• Avoid cheap Amazon/eBay cables.
• The stock adapter is the weak link, not the connector itself — buy the direct cable from your PSU manufacturer.
• Never daisy chain — each 8-pin connector on the PSU side needs its OWN dedicated, separate PCIe power cable from the PSU.
• Ensure full insertion.
• Route cables with slack.
• You’re running sustained 350W loads for hours during inference/training, not just gaming spikes.
• The 12-pin was NVIDIA’s first proprietary power connector experiment — it worked, but it was a warning sign of the melting disaster that hit the RTX 4090 a generation later.

CPU — Ryzen 9 7950X3D

16 cores, 32 threads, and the 3D V-Cache that makes this chip special. For AI workloads, the CPU isn’t the bottleneck – the GPUs are doing the heavy lifting. CPUs do generalized computation, while GPUs do specific computation really well. Because of this, I opted for a 16 core CPU. I didn’t feel the need to go crazy and get anything with more cores. This CPU gives me enough cores to handle data preprocessing, run a web server for OpenWebUI, manage multiple Docker containers, and still have headroom for multitasking. The 3D V-Cache does not help that much for AI training, but it’s nice to have for everything else. AMD makes the best CPUs. However, in the last 6 months, I have been blown away by Intel’s comeback, as well as the stock’s price $$$. They’re making serious headway with their newer CPUs and especially their progress on the Arc GPUs. Ultimately, I knew that I wanted an AMD CPU – which then drove home the motherboard purchase, because the motherboard uses the AM5 socket.

RAM — 96 GB of DDR5

I can write an entire essay alone on RAM. All you have to know is that we are in a RAM crisis, mainly due to Sam Altman signing letters of intent with Samsung and SK Hynix for 900,000 DRAM wafer starts per month — roughly 40% of global DRAM output TrendForce, October 2025 — for OpenAI’s Stargate project. It’s gotten so bad, that even my grandma is asking what RAM means. At the time of the build in fall of 2024 – we were not in a RAM crisis. The big MoE explosion (DeepSeek-V3, then R1) hadn’t hit yet, Llama 3 models were still rolling out (3.2 had just released, 3.3 wasn’t out yet), and Sam Altman did not ‘ram’ the market into max pain. I have 96 GB of DDR5 with a RAM speed of 6800 MT/s. Currently I am using 6000 MT/s of RAM speed. This is critical to understand – people will try to convince you that you need fast RAM. This is NOT true. Prior to MoEs, RAM speed did not matter for AI workloads that much. When I bought this RAM, my motherboard could not support RAM speed being stable above 5200 MT/s. I recently updated my BIOS, and tested it out going from 5200 MT/s to 6000 MT/s and it’s safe to say that it’s stable. You must check your motherboard’s QVL and determine via the past BIOS for it, what RAM and RAM speed it can run. My current motherboard cannot support RAM speed higher than 6000 MT/s being stable – keyword being stable here.

96 GB is more than most workstations have – but I knew that at a price point of $346.00 USD, it was well worth the price. That same RAM is now worth $1,300 USD. This puts into perspective how bad the RAM crisis has actually gotten. This amount of RAM is enough to hold a full 7B model in system RAM for CPU-only inference, manage large datasets during preprocessing, and keep everything running smoothly when I have training jobs, a web UI and a notebook running simultaneously.

PSU — Corsair AX1600i (1600W)

The Corsair AX1600i is overkill for the current configuration – and that’s by design. Two 3090s pull up to 350W each under full load. A single 3090 can actually spike well past 500W (up to 550-600W for milliseconds before power management kicks in), so I wanted extra headroom to be extra safe. Having 1600W of headroom means clean, stable power delivery with room to grow. I knew before this build, that I wanted extra wattage headroom to grow either with bigger, badder cards, or more 3090s. You can limit GPU power with a single command – and the performance loss is almost nothing. Limited at 280 watts/card instead of 350, I lose 4% of my tokens per second, but save 20% on power, run 15°C cooler (for you Americans 25-27°F cooler) – and my cards will last years longer. The exact nvidia-smi script I run to apply this lives in my repo on GitHub. You need to think like this. Everything about local AI is optimization – not just software either, for hardware as well. Here is a general wattage pull of the system (numbers calculated with the help of Claude Code).

Component	Idle	AI Inference	Max Load
CPU (Ryzen 9 7950X3D)	15 W	75 W	150 W
Motherboard (ProArt X870E)	30 W	45 W	55 W
GPU #1 (RTX 3090 FE)	20 W	220 W	350 W
GPU #2 (RTX 3090 FE)	20 W	220 W	350 W
AIO Cooler (pump + LCD screen)	4 W	5 W	7 W
Fans (10x SL Infinity 120mm)	18 W	30 W	52 W
RAM (2x 48GB DDR5-6800)	5 W	12 W	18 W
Storage (990 PRO 2TB NVMe)	0.05 W	5.5 W	8.5 W
System Total (DC)	~112 W	~613 W	~991 W

Case — Lian Li O11 Dynamic EVO XL

A full tower eATX case with tempered glass panels and excellent airflow. This case has room for dual 3-slot GPUs, a 360mm radiator up top, and enough fan mounts to keep everything cool under sustained AI workloads. It’s a solid unit of a case. When you’re running a dual 3090 setup at full load for hours at a time, you want volume and airflow. This case is gorgeous, an absolute beauty, I love it very much.

One important caveat with this case, and this is real insider knowledge. Hypothetically, if you were to run 3 GPUs (even the slimmest 3090s) using the 3rd PCIe lane — you’d hit a wall, literally. In this specific case, in a 3-card scenario, the last card (3rd GPU) will not be easy to place into the case. You will have to mod the case – meaning file off where the connecting metal pieces of the GPU insert into the case. This is due to the case and motherboard, more so the case. I’ve seen one person do this – and they did it temporarily for their third card – and eventually mounted the 3rd GPU in the back case vertically because of this issue. This is one of the best cases to get for this setup despite this one issue.

Storage — Samsung 990 PRO 2TB NVMe (PCIe Gen 4)

I went with a single 2TB NVMe stick at the time because most local AI work was not how it is now for local models. The local model landscape was simpler back then. Llama 3.1 had just dropped, Qwen 2.5 had just been released, and Mistral/Mixtral and Gemma 2 were the other serious options. Nowhere near the explosion of strong local models we have today. On top of that, pre RAM crisis, SSDs were cheap compared to now. Because it was my first time building a rig of any kind, I figured 2TB was enough. Looking back, 2TB is tight – I’m constantly bumping up against the ceiling between local model weights, training data, Docker images and OS overhead. It works, but barely. If I had to do it again, I’d start with 4TB minimum.

Heads up: SSDs are the next to spike in hardware prices – spoiler: they already have. This is sadly the world we live in. Western Digital is sold out on their HDDs, and all NVMe makers will continue to do price hikes. If you must, to save money, then go with HDDs, not SSDs. You will get more storage per dollar – but the tradeoff is that loading data from the drive to the CPU will take a lot longer. Not a bad tradeoff if you have patience. Also keep in mind that modern-day storage is subject to bitrot.

Motherboard — ASUS ProArt X870E-CREATOR WiFi

This is an AM5 system motherboard. At the time – it was brand new when I purchased it. It has a total of 3 PCIe slots, DDR5 RAM slots, PCIe Gen 5, WiFi 7 Marvell Chip, 10G and 2.5G ethernet, a crap ton of ports for a workstation vibe, and looks gorgeous. Out of all of the parts that I bought – this is the one that I regret the most and where I learned the saying: “Do your research first, and after you feel like you are confident – do it again.” In hindsight, this was a big mistake I made. I trusted the guy from Micro Center way too much – shoutout to this guy who complicated my life – but it’s really my fault since I didn’t do enough research.

The flaws:

1) The newness of the motherboard. When it came out, I was pitched at Micro Center that this could run a dual GPU setup. The board has 3 PCIe slots – seems fine right? 3 slots = 3 GPUs was my mindset at the time. Turns out, the two top slots are PCIe 5.0 x16 connected directly to the CPU, while the third slot is PCIe 4.0 x4 connected to the AMD chipset, not the CPU. What this means is that because consumer AM5 CPUs only have 24 usable PCIe lanes (vs. 128 on Threadripper PRO platforms), when you populate both top slots at once (dual 3090s) – the 16 CPU lanes bifurcate, splitting 8x/8x. Instead of each card getting dedicated 16x lanes – they each drop to 8x. The chipset slot only ever runs at PCIe 4.0 x4. There’s also a sneaky gotcha: the second top slot shares lanes with the M.2_2 NVMe slot — populating M.2_2 drops slot 1 to x8 and slot 2 to x4. Read the manual. Twice.

This translates into my bandwidth getting cut by 50% for data transfer speed. During inference and training – this matters a lot, but not as much as one thinks. Tim Dettmers writes a lot about this, his work made me feel better – but as time went on, local models got better and bigger. This meant that even with 4-bit quantization, every t/s actually matters now. As I write this post, coding agents are HOT. t/s matter more than ever. We are well past the days of running a Llama 3 model. The new local models are badass and to do agentic work – it’s like training for the Olympics now.

2) PCIe lanes – generation and space. Going from 16x to 8x isn’t such a big loss in the big picture. Out of all of the issues – this was the least concerning. However, if I did my homework – I would have saved myself a lot of brain cell usage and anxiety. The RTX 3090 uses PCIe generation 4, or what is called Gen 4. Each PCIe gen roughly doubles the bandwidth of the previous one. The problem: I have a Gen 5 PCIe mobo, but my cards use Gen 4. I wasted money in this aspect, because my cards will never use the Gen 5 perks of this board, one of the coveted things about it. Yes, the NVMe M.2 SSD can use a Gen 5 PCIe bus – but I have a Gen 4 NVMe. But worst of all was the slot spacing. This issue haunts me to this day, and I continue to lose sleep over it.

Because I have the Founders Edition cards – they are 3-slot spaces wide (VERY THICK). On this mobo, when you have dual FE 3090s, it leaves about 0.25 inches between each card. This is nowhere even close to the amount of space needed between a GPU. The 3090, due to its Ampere architecture and design, gets HOT. The space between them right now, is the absolute bare minimum level which one can push. The top card, GPU 0, gets absolutely cooked. For most people, this is acceptable – but when your GPUs are running for days, this gives you the Nvidia sweats at night. It’s one thing to have your card die, it’s another to worry about it catching on fire. The top card suffers from a massive thermal throttling issue. You can limit the power consumption of the cards, but it still doesn’t fix the issue entirely. At sustained peak load, there is a 10-15 °C difference between the top card and bottom card. Thank the Lord that I have strong cooling and a good setup, it’s just enough to allow me to do what I need – but please DO NOT make this same mistake that I made.

AIO — Lian Li Galahad II LCD 360mm

For cooling the CPU, I went with a liquid cooler. It’s does the job but has one caveat: Lian Li software BLOWS. What I have learned – avoid flashy AIOs for all local AI. Don’t get suckered like I did to get a cool LCD display. The L-Connect 3 Lian Li software on Windows that manages the LCD screen is always buggy and consumes way too much of my CPU. I regret this one – just get a simple AIO if you do liquid cooling. FYI – this model is known to be extremely buggy.

Fans — 10x Lian Li SL Infinity 120mm

I have 10 Lian Li SL Infinity fans. They are called Infinity because when they spin, it looks like a scene from the movie Interstellar. The layout is what is most important here. We have a single fan by the CPU area, 3 fans daisy chained together on the bottom, 3 fans daisy chained together on the radiator of the AIO, and 3 fans daisy chained together in the back of the case. The airflow is shown in the image: positive pressure. Positive pressure means – fresh cold air is pulled into the AI server case at all angles, and it flows upwards through the GPUs, up through the case, exiting the hot air via the radiator. Cooling is just as important as anything else, so make sure that you do this right the first time. The last thing you want is to cook your GPUs. To date, the best fans to get are the Phanteks T30 fans. They will run on any operating system. Fancy fans most of the time will require Windows, or if you want a Linux system with fancy RGB fans, you must check open source places to make sure you can run them on a Linux machine.

Operating System — Windows 11 Pro

Screw Windows OS. I paid $199.99 to Microsoft for Windows 11 Pro. I could write an essay ranting about how shitty Windows is as an operating system. When I first did this build, I didn’t thoroughly understand the choice I made. All AI essentially runs on Linux. Windows is NOT a good AI operating system. They claim to support AI. Windows is like communism – it looks really good on paper, but in reality it never works. The craziest part is how bad Windows has gotten since I installed it. Towards the end of 2024 is when I did a fresh install. Since then it has gotten noticeably worse. Running Windows is not good for AI – as the WDDM drivers for Windows seriously conflict with GPUs sometimes, as well as having a VRAM overhead of roughly 0.5-1.5 GB (depending on resolution and monitor count) required to maintain the monitor display via the GPU. One workaround is if your CPU has an iGPU (integrated graphics built into the processor itself) – use the motherboard’s HDMI/DisplayPort output to render the monitor screen through the iGPU instead of the discrete GPU – which allows full VRAM dedicated for AI work. So why am I still on Windows? Migration risk.

Re-doing my entire CUDA stack, NVIDIA drivers, Docker configs, model weights, and OpenWebUI setup would mean taking the rig offline for a couple of weeks assuming no major issues occur. The pain of moving outweighs the pain of staying. The V2 cluster I’m finishing right now is being built on Ubuntu 22.04 LTS from day one. No Windows tax, no WDDM driver weirdness, no auto-reboot at 3 AM during a training run. Why Ubuntu 22.04 LTS specifically and not 24.04? Stability — 22.04 has the most mature CUDA, cuDNN, and PyTorch ecosystem support in 2026. Most AI tooling targets it as the reference platform. It’s the safe pick.

There are many other legitimate and serious reasons to use Linux for AI servers, which I won’t get into in this post, but trust me bro on this one — do not make this mistake. Also, Linux is free — Windows is $199.99 USD one time — @BillGates, what are you even doing dude?

Connecting to the Server (Remote Access)

Once your AI rig is built, you don’t want to be physically sitting in front of it every time you want to run a model. You want to ssh in from your laptop, kick off a training job, close your screen, and walk away. Here’s how I do it.

SSH

On Windows 11, OpenSSH Server is built in but disabled by default. Enable it under Settings → Apps → Optional Features → OpenSSH Server. From your laptop, you `ssh user@local-ip` and you’re in.

Tailscale

This is the move. Tailscale is a mesh VPN built on WireGuard that creates a private, encrypted network between your devices — your laptop, your AI rig, your phone — without opening a single port on your router or exposing anything to the public internet. You install the client on each device, log in with your Google or GitHub account, and your devices get private 100.x.x.x IPs that only you can see. From a coffee shop, on cellular, on a friend’s WiFi, `ssh user@my-rig` just works. No port forwarding. No dynamic DNS. No public IP exposure. It’s free for personal use. If you take one piece of advice from this post — install Tailscale before you do anything else. Tailscale is an amazing company. I cannot emphasize this enough – this is the gold standard for remote AI server access.

Disable Windows Auto-Updates — This one is mission-critical for an AI server. By default, Windows 11 sucks. Windows 11 will not let you disable updates permanently (shocker). Because of this, you can auto-disable updates for a certain period of time, up to 5 weeks max I believe. There is talk that Windows 11 may change this, but until they do, assume they will not. For example, I was recently running a Hermes agent locally, and I woke up the next day to find out that my machine crapped out. Because it’s a remote server, I had to travel there to find out what happened. Because the machine went down, and I could not SSH into it, I had to physically be there to see what the problem was. After going crazy, it turns out Windows updated literally at 3 AM and set me back big time. Here I was thinking I had a GPU hang issue, or an issue with the machine. This is another reason why Linux is the better choice for AI servers — `apt update` only runs when I tell it to.

The Full Parts List

Here’s every component in the build, with costs at the time of purchase in 2024. All prices are pre-tax. I’m including this so you can see exactly what a 48 GB VRAM AI workstation costs in real dollars. It took me about 3.5 weeks to source/acquire all the parts, and roughly another two to three weeks after that to assemble, troubleshoot, and get the machine fully operational. About six weeks from the first purchase to a working system.

I bought protection plans on every component I couldn’t easily replace out of pocket. GPUs, CPU, motherboard, and the AIO cooler. No protection on the PSU or RAM — the Corsair AX1600i has a 10-year warranty from the manufacturer, and DDR5 rarely fails. This is up to you to decide if you want these. Since it was my first time building a PC, let alone doing anything this serious with hardware, I was very paranoid and played it safe. Some parts do break, some not so much with care and good maintenance practice.

#	Category	Part	Price
1	GPU	NVIDIA RTX 3090 Founders Edition — 24 GB GDDR6X VRAM	$629.96
2	GPU	NVIDIA RTX 3090 Founders Edition — 24 GB GDDR6X VRAM	$699.99
3	CPU	AMD Ryzen 9 7950X3D — 16 cores, 32 threads, Zen 4, 3D V-Cache, AM5	$579.99
4	Motherboard	ASUS ProArt X870E-CREATOR WiFi — AMD AM5 ATX	$479.99
5	RAM	TeamGroup T-Create Expert 96GB Kit (2x48GB) DDR5-6800 CL36	$345.99
6	Storage	Samsung 990 PRO 2TB NVMe — PCIe Gen 4	$144.99
7	CPU Cooler	Lian Li Galahad II LCD-SL Infinity 360mm AIO — 2.88″ IPS LCD, includes 3x 120mm fans + fan controller	$289.99
8	PSU	Corsair AX1600i — 1600W, 80+ Titanium, fully modular	$609.99
9	Case	Lian Li O11 Dynamic EVO XL — eATX full tower, tempered glass	$234.99
10	Fans (3-pack)	Lian Li UNI Fan SL Infinity 120mm — 3-pack	$89.99
11	Fans (4 singles)	Lian Li UNI Fan Reverse SL Infinity 120mm x4	$119.96
12	GPU Cable	Corsair 12-Pin GPU Power Cable (PSU to GPU)	$29.98
13	GPU Cable	Corsair 12-Pin GPU Power Cable (PSU to GPU)	$26.98
Hardware Subtotal			$4,282.79
14	Protection Plan	GPU #1 (RTX 3090 FE) — 2 Year	$79.99
15	Protection Plan	GPU #2 (RTX 3090 FE) — 2 Year	$79.99
16	Protection Plan	CPU (Ryzen 9 7950X3D) — 2 Year	$69.99
17	Protection Plan	Motherboard (ProArt X870E) — 2 Year	$74.99
18	Protection Plan	AIO Cooler (Galahad II LCD) — 2 Year	$34.99
Protection Plans Subtotal			$339.95
19	Software	Windows 11 Pro (one-time license)	$199.99
20	Software	Microsoft 365 Personal (annual subscription)	$99.99
Software Subtotal			$299.98
Grand Total (pre-tax)			$4,922.72
Tax			$412.58
Grand Total (with tax)			$5,335.30

$5,335 all-in. That’s everything — every cable, every fan, every protection plan, every software license. No hidden costs. The machine you see in the specs above is the machine these receipts paid for.

Further Reading

These are some of the best resources I used to start, to determine how to structure my rig, and to learn. If you are just starting, read these, and you will be just fine.

• Tim Dettmers — Which GPU for Deep Learning — The canonical reference for GPU selection. Tensor cores, memory bandwidth, perf-per-dollar. Read this first.
• Tim Dettmers — Deep Learning Hardware Guide — Older but still relevant. PSU, cooling, motherboard reasoning all in one place.
• James Betker — My Deep Learning Rig (archived) — OpenAI’s DALL-E 3 lead documenting his 8x 3090 home rig. Real-world multi-GPU pain you don’t get from product pages. (Original site nonint.com is currently down — linked via Wayback Machine.)
• Chan Kha Vu — Meet Iva, my DL rig — Another dual-3090 build, open-air mining-style approach. Useful contrast to the closed-case route I took.
• LlamaBox — LLM AI Workstation (PCPartPicker) — A 3-GPU build (2x RTX 3090 + 1x RTX 3090 Ti, 72GB VRAM total) in the same Lian Li O11 Dynamic EVO XL case I used. The builder confirms what I describe in the Case section above: “Had to bend one of the tabs of a 3090 where it slots into the case. The case ran out of slots under the 3rd PCI lane.”
• Curtis Northcutt — The Best 4 GPU Deep Learning Rig – A 2020 build documenting four blower-style RTX 2080 Tis stacked in a single case. Useful historical reference for what 4-GPU rigs cost in 2020 and a real-world example of why blower cards matter in dense multi-GPU configurations.
• Best Value GPU — eBay-tracked GPU pricing history. I used this to sanity-check used market pricing throughout this post.
• TrendForce — Source for the RAM crisis data cited above. Authoritative for memory market reporting.

What I’ve Learned in 9 Points

• You learn the best by doing. Building and running this machine has taught me things that no book, YouTube video, research paper or blog could ever do.
• Building the machine is part of the education. Understanding the hardware stack makes you a better AI engineer.
• Learning by necessity always wins. You will never have enough compute, EVER, and neither do the biggest labs in the world. Being in a constrained environment pushes you to learn more and think differently. Don’t fear it — accept it and learn to use it as an advantage.
• You will make a ton of mistakes. That’s part of the process. In fact, you will make so many mistakes, that you will actually learn to be okay with it as your fear/pain threshold quickly rises. This is what makes you stronger.
• Most people in AI don’t know about hardware or don’t do hardware at all. Building a rig alone gives you a major confidence boost, different avenues for job opportunities, and teaches you a lot about what it truly means to be a master of your work.
• DO your research. When you think you have done enough — do more. I could have saved myself a lot of sleepless nights doing this if I spent more time doing the right research.
• You can’t beat physics. Be creative within it, but don’t fight it
• Simple is always best. Occam’s Razor.
• Until the AI bubble pops, assume ALL hardware prices will continue to rise – buy the parts you need NOW and NOT LATER.

Closing Thoughts

And there it is, the V1 cluster. 48 GB of VRAM, a dream, and $5,335 of receipts that turned a Senior Data Scientist into someone who thinks about PCIe lanes at 2 AM. Eighteen months in, this machine has run every late-night experiment, every fine-tune, every weekend deep dive and it’s the reason V2 is even possible.

This is just the build. Part 2 will compare every part, then vs. now, what I paid in fall 2024 versus what these same parts cost today. Spoiler: not in your favor, and the gap is widening. Part 3 goes into the actual AI research I’m doing on this machine — the Hermes agent that crashed at 3 AM, the experiments that worked, the ones that didn’t, and what local AI looks like when you can finally just do it.

If you’ve made it this far — thank you. Build the rig. Make the mistakes. The learning curve is the point.

*This is part of Project Arrival — my journey from Senior Data Scientist to AI research engineer. I’m documenting the build, the research, and the path to joining a top AI lab. Follow along at blog.zacharycangemi.com or at my GitHub https://github.com/zacangemi/local-llm-infrastructure