Custom AI Workstations Built for Machine Learning and Deep Learning

The honest through-line across tiers: on Ryzen and on Threadripper PRO Max-Q there is no NVLink, so multi-card VRAM sits in separate 96 GB islands rather than one large pool. The only way to pool VRAM across cards is the H200 NVL 4-way NVLink bridge (564 GB) in the enterprise tier. Choose the tier by GPU count and pooling need, then get a current quote — we will not post a stale number here while RAM and GPU pricing keep moving.

Component pricing has moved fast enough that a parts-total comparison against last year's numbers is already out of date. 16 GB DDR5 chip prices rose about 298 percent from September to December 2025. Consumer DRAM contract prices climbed roughly 89 percent in a single 2026 quarter. NAND flash pricing is up about 115 percent over the same stretch, because memory makers redirected wafer capacity to HBM for AI accelerators instead of consumer DRAM and SSDs. Analysts do not expect a meaningful correction until late 2027. A warrantied, hardened, benchmarked, and shipped machine from Petronella is a different product than a pile of boxes you assemble yourself and hope run stable, and the parts you would buy to assemble it are getting more expensive by the month.

For a single dedicated GPU, the AM5 platform built around the Ryzen 9 9950X is the right choice. AM5 exposes roughly 24 PCIe lanes from the CPU, which is enough for one full-bandwidth GPU slot. Add a second GPU on this platform and both slots drop to x8/x8, splitting bandwidth neither card was designed to share. A single 600-watt-class GPU workstation runs comfortably on a dedicated 20-amp circuit.

Once you need two or more GPUs, move to a platform built for it. Threadripper PRO provides up to 128 PCIe 5.0 lanes, enough for multiple GPUs at full bandwidth without a lane-sharing bottleneck. Dual EPYC platforms extend that further, to roughly 160 lanes. Both change the electrical picture: two or more GPUs typically push a build to a 30-amp, 208 to 240-volt circuit, and dual EPYC servers commonly require 208V power to begin with. Never spec a 2+ GPU Threadripper PRO or EPYC build on a standard 20-amp, 120-volt circuit.

One more reality check before you plan around multiple RTX PRO 6000 Max-Q cards: Blackwell has no NVLink. Two 96 GB cards are two separate 96 GB islands, not a pooled 192 GB card. A model that does not fit on one card needs tensor or model parallelism across PCIe, with real overhead, not a seamless larger pool. For the largest single models, one large-VRAM card beats splitting the load across two smaller ones. The RTX 5090 remains the strongest single-GPU option below the Max-VRAM tier.

A custom AI workstation is a desktop-class computer built specifically for artificial intelligence workloads. The hardware configuration prioritizes GPU compute, memory bandwidth, storage throughput, and cooling capacity over the metrics that matter for general-purpose PCs. While a standard office workstation might have a single consumer-grade GPU, a custom AI workstation can house up to four high-end GPUs connected via NVLink, with sufficient power delivery to run all of them at full utilization simultaneously. The CPU, motherboard chipset, RAM capacity, NVMe storage configuration, power supply wattage, and chassis airflow are all selected to support the specific demands of machine learning and deep learning training.

The distinction between a custom AI workstation and an OEM workstation from Dell, HP, or Lenovo comes down to control and optimization. OEM workstations are designed for broad markets. Their cooling systems are tuned for quiet office operation, their BIOS settings are often locked or restricted, and their component options are limited to whatever the manufacturer offers in their current catalog. A custom build gives you full control over every component, unrestricted firmware access, and the ability to upgrade any part of the system at any time without voiding warranties or hitting proprietary limitations. Petronella Technology Group is not a Dell or HP reseller. We assemble and configure every AI workstation in-house at our Raleigh, NC facility, selecting each component by hand based on your specific workload requirements.

For AI teams doing daily training runs, fine-tuning large language models, running computer vision pipelines, or serving inference workloads, the performance difference is significant. A properly configured custom AI workstation with an NVIDIA RTX 5090 can deliver training throughput comparable to a cloud A100 instance at a fraction of the ongoing cost. If you would rather build one yourself, we currently have two used, tested RTX 5090 32GB cards for sale. Scaling up to an RTX PRO 6000 Blackwell with 96 GB of GDDR7 memory allows single-GPU training of the largest open-weight language models without the complexity and expense of multi-node distributed training setups. Petronella has been building custom hardware for more than 24 years and has served clients since our founding. Our CEO, Craig Petronella, is the author of 8+ published books on cybersecurity and technology and hosts the Encrypted Ambition podcast, where he regularly covers AI hardware, security, and infrastructure topics. Our AI workstation configurations reflect real-world testing against production AI workflows, not theoretical benchmarks.

The economics of ownership are straightforward. A comparable cloud GPU instance runs roughly $32.77 per hour, about $23,594 per month at sustained utilization. A Petronella workstation with comparable or better performance is a one-time purchase across any of the three tiers above, and it runs unlimited hours for the life of the hardware. Because component costs move week to week right now, we do not post a stale from-price here; request a quote for the current number. Most teams recoup the full cost of a custom build within 4 to 8 months of daily use. For organizations that need dedicated compute available around the clock, the financial case for custom hardware is overwhelming. Petronella also offers GPU server hosting for teams that want the economics of owned hardware without managing physical infrastructure.

GPU Selection. The GPU is the most important component in an AI workstation. Your choice depends on model size, training batch size, and whether you need single-GPU or multi-GPU capability. For smaller models, an RTX 5080 or RTX 5090 provides excellent performance at a reasonable cost. For models above 30B parameters, the RTX PRO 6000 Blackwell with 96 GB memory is the right choice. If you plan to scale to multiple GPUs, confirm that your motherboard and chassis support the physical and power requirements. Petronella handles all of this analysis during our workload assessment phase and recommends the GPU configuration that matches your current needs while leaving room for future upgrades.

CPU Platform. The CPU matters more than many AI practitioners realize, particularly for data preprocessing, tokenization, and pipeline orchestration. AMD Ryzen 9 9950X excels at data pipeline operations with strong single-thread performance and a large L3 cache, and it is a single-GPU platform (see "Which Platform" above). AMD Threadripper PRO is the choice for multi-GPU builds because it provides 128+ PCIe lanes, ensuring full bandwidth to each GPU without PCIe switching bottlenecks. Intel Xeon W provides ECC memory support for teams that need error correction for mission-critical training stability. Petronella matches the CPU platform to your GPU count and workflow requirements.

Memory (RAM). AI workloads are memory-hungry. Data preprocessing, tokenization, and batch assembly all happen in system RAM before data moves to GPU memory. We recommend 64 GB as a minimum for single-GPU builds, 128 GB for the Pro tier above, and 192 GB or more for multi-GPU Threadripper PRO or EPYC configurations and large-dataset workflows. ECC RAM is recommended for training runs that last multiple days, where a single memory error can corrupt a checkpoint and waste hours of compute time.

Storage Architecture. AI workstations need fast storage for dataset loading and large capacity for storing datasets, checkpoints, and model artifacts. Petronella configures tiered storage: a primary NVMe SSD for the OS and active datasets (Gen4 or Gen5 for maximum throughput), secondary NVMe for checkpoints, and optional high-capacity SATA SSDs or HDDs for cold dataset storage. The storage configuration prevents GPU idle time caused by slow data loading, which is one of the most common performance killers in poorly configured AI workstations.

Cooling and Power Delivery. This is where custom builds differ most from OEM systems. Consumer-oriented workstations optimize for acoustic comfort, which means thermal throttling under sustained load. AI training runs 100% GPU utilization for hours, days, or weeks. Petronella engineers the cooling solution for sustained thermal performance: high-static-pressure case fans, direct-contact GPU coolers, optimized airflow paths, and chassis designs that exhaust heat efficiently. Power delivery is equally important. Multi-GPU builds can draw 1,600W or more under full load, so Petronella specifies server-grade or redundant power supplies with sufficient headroom to avoid stability issues.

Operating System. Petronella is a Linux-first shop. Most of our AI workstations ship with Ubuntu or Rocky Linux, but we also offer NixOS builds for teams that want fully reproducible, declarative system configurations. NixOS is particularly valuable for AI workstations because it allows you to define your entire software environment, including CUDA versions, Python packages, and system libraries, in a single configuration file that can be version-controlled and rolled back instantly. We also support Windows builds for teams that require it, though we recommend Linux for maximum AI framework compatibility and performance.

A GPU workstation for machine learning needs to handle two distinct phases of the AI development lifecycle: training and inference. During training, the GPU processes massive amounts of data through neural network layers, adjusting millions or billions of parameters over thousands of iterations. This phase demands sustained memory bandwidth, high FLOPS throughput, and thermal stability over extended periods. During inference, the trained model processes new inputs to generate predictions or outputs. Inference workloads prioritize latency and tokens-per-second over raw throughput. The ideal machine learning workstation balances both requirements.

Deep learning workstations face even more demanding requirements because deep neural networks are larger, train for longer, and consume more GPU memory than traditional machine learning models. A convolutional neural network for medical imaging might train for days on a single GPU. A transformer model for natural language processing might require 96 GB of GPU memory just to load the model weights. Petronella configures deep learning workstations with the GPU memory, storage throughput, and cooling capacity needed to sustain these extended training runs without performance degradation.

The choice between a single-GPU and multi-GPU deep learning workstation depends on your model sizes and training timelines. Single-GPU builds with an RTX 5090 or RTX PRO 6000 are sufficient for most fine-tuning and medium-scale training jobs. Multi-GPU builds with two to four GPUs connected via NVLink reduce training time linearly for workloads that parallelize well. When the training run grows beyond what a workstation chassis can sustain, customers usually graduate to an NVIDIA HGX 8-GPU baseboard server with NVLink fabric and pooled HBM3e memory - the rack-scale platform we configure for serious training workloads. Petronella helps you determine the right configuration based on your actual models and datasets, not theoretical benchmarks. We also explain the tradeoffs clearly so you can make an informed decision about where to invest your hardware budget for the greatest impact on your AI development velocity.