There is no reliable universal price for training “an AI model” on a supercomputer. Estimate the workload’s runtime on the system you intend to use, convert that time into the facility’s or provider’s billing unit, then add any separately charged storage, networking, data movement, and operations. A cloud bill and a research-facility allocation are different kinds of access, not directly comparable hourly prices.
What determines the cost?
The estimate depends on three things: the work you plan to run, the hardware and efficiency achieved, and the terms under which you access that hardware. Model parameter count by itself does not determine training time or price. Pretraining and fine-tuning can also require very different amounts of compute, data, and evaluation.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
MINISFORUM MS-02 Ultra Workstation Mini PC, Intel Core Ultra 9 285HX (24C/24T, up to 5.5GHz), PCIe... | $1,659.00 | Buy on Amazon |
| 2 |
|
GMKtec EVO-X2 AI Mini PC Ryzen Al Max+ 395 Superchip 128GB LPDDR5X 2TB SSD | $3,649.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
Before estimating, record the architecture and model size, training objective, token or sample volume, sequence length, precision, parallelism plan, number of steps, and number of planned runs. Include evaluation and data-preparation jobs if they will use the same paid or allocated resources.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Estimate compute from a representative benchmark
- Choose the target configuration. Identify the system, node or instance type, accelerator count, and intended parallelism. Confirm whether the published unit is GPU-hours, node-hours, instance time, or a reserved block; “node” can mean different configurations on different systems.
- Benchmark a representative slice. Run a portion of the intended workload on the target system and measure end-to-end throughput and wall-clock time, including data loading and communication. A benchmark on a different configuration may not predict performance at the planned scale.
- Extrapolate workload time. Use the measured throughput and planned training volume to estimate elapsed runtime. Add separately planned time for evaluation, checkpointing, debugging, restarts, and unsuccessful runs rather than hiding those activities in an unexplained multiplier.
- Convert time to the chargeable or allocated unit. For a cloud setup, multiply elapsed instance time by the configured resource count and apply the rate and terms for the chosen region, machine, GPU, and discounts. For a facility quoting node-hours, multiply runtime by the number of nodes, using that facility’s definition of a node.
Neither parameter count nor peak accelerator performance alone gives a dependable runtime. The relevant estimate comes from a workload-specific benchmark. There is no universal runtime conversion factor or standard failure allowance that can safely be applied to every training job.
#1 Best Overall
- High-Performance AI Processor：The MS-02 Ultra features an Intel Core Ultra 9 285HX (24C/24T, up to 5.5 GHz, 13 TOPS NPU), delivering fast and efficient performance for AI inference, algorithm development, and media workloads. A PCIe x16 expansion slot supports desktop-class GPU upgrades for advanced model training and accelerated computing tasks. It's ideal for creators, engineers, and teams handling intensive parallel workloads.
- 4 × M.2 PCIe 4.0 + 4 × DDR5 SODIMM slots:Four DDR5 SODIMM slots support up to 256 GB of memory, while ECC helps maintain data integrity in mission-critical environments. Four PCIe 4.0 M.2 slots support up to 24 TB of storage, supporting RAID 0/1/5/10, combining high-speed performance with data protection. It allows for the creation of independent scratch disks, media libraries, and project drives, providing high-throughput for production workflows.
- PCIe & USB 4.0 v2: Up to three PCIe slots can be equipped, including a dual-slot x16 GPU. The main slot supports PCIe 5.0, meeting the needs of high-bandwidth creative and computing workloads. USB 4.0 v2 (80Gbps) supports high-bandwidth external storage and displays.
- Ultra-fast Networking: Wi-Fi 7 further enhances wireless performance with next-generation speeds and low-latency stability. Intelligent bandwidth switching optimizes throughput in different network environments, ensuring optimal performance for enterprise or local networks. Dual 25GbE ports (providing up to approximately 3.125 GB/s bandwidth, about 25 times faster than traditional 1GbE), enabling seamless large-scale file transfers and parallel computing. 10GbE and 2.5GbE ports, with support for Intel vPro technology, ensure enterprise-grade remote management and deployment flexibility.
- Server-grade thermal architecture: Utilizing a dedicated CPU/GPU airflow design, equipped with a 6-pipe dual-fan cooler, it maintains stable performance even under sustained loads, delivering up to 140W Turbo power while maintaining a 100W TDP, and operating with noise levels as low as 36 dB. An integrated 350W power supply ensures stable and reliable output for demanding computing tasks and fully loaded extended configurations.
Build a complete project estimate
Use the following worksheet. Some services bundle costs; others charge certain items separately, so use the terms for the exact configuration rather than assuming every row appears as a separate bill.
| Cost area | How to estimate it | Inputs to verify |
|---|---|---|
| Training compute | Measured runtime × configured nodes or instances × applicable rate, or the resource units consumed | Representative benchmark, billing unit, current quote or allocation terms |
| Data preparation and evaluation | Estimate or measure these as separate jobs | Workflow schedule and job benchmarks |
| Storage | Capacity × retention duration × applicable rate, when charged separately | Training data, checkpoints, outputs, filesystem or object-store terms |
| Networking and data movement | Apply relevant transfer and network charges, if billed | Data location, transfer plan, provider terms |
| Setup and operations | Include project-specific labor and services | Staffing, support, and service choices |
| Energy, if separately billed | Measured or estimated energy × the applicable billed energy rate | Power measurement or model and the actual billing arrangement |
| Contingency | State a project-specific scenario allowance | Explicit risk assumptions; there is no general percentage established here |
Total project estimate = compute + storage + networking and data movement + setup and operations + separately billed energy, if any + explicitly stated contingency. Label each input as quoted, benchmarked, modeled, or assumed. Google’s cost framework identifies compute, networking, storage, training-data and adapter-layer storage, application and setup, and operational support as cost areas: Google Cloud cost framework.
Cloud prices and supercomputer allocations are not the same
Cloud GPU services
Cloud pricing pages may show a GPU rate without the full configured cost. Google Cloud states that GPU charges are additional to VM machine-type charges; its GPU table excludes disk, networking, and VM instance pricing. Use its GPU pricing page to identify the relevant region and accelerator, then price the configured resources with the provider’s calculator and current terms. Rates, zones, and discounts can change, so a copied rate is not a durable project estimate.
Recommended Free Tools
For any provider, check whether a quoted configuration includes CPU and memory, local or attached storage, networking, licenses, support, taxes, and committed-capacity terms. These depend on the service and contract; do not treat a GPU-only price as the cost of a training run.
Research allocations
An awarded allocation can reduce or eliminate a direct compute charge to an eligible project, but it is not a retail hourly price. Eligibility, award size, resource units, storage, scheduling, and access period affect whether it is useful for a particular workload.
In its March 18, 2026 call for the 2026 allocation year, which runs through January 19, 2027, NERSC said accepted projects could initially receive up to 10,000 Perlmutter GPU node-hours. A Perlmutter GPU node has four A100 GPUs, and the call describes associated filesystem storage quotas. These are award terms, not a public cash rate: NERSC AI for Science call.
Rank #2
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Paid facility access
OLCF’s Lux page says half of the system’s annual 3.5 million node-hours is reserved for the Genesis Mission, with the remaining half available for proprietary paid use under the DOE User Facility rate. It states a minimum commitment of 175,000 node-hours per six-month commitment and says allocated storage access is included. The page describes more than 4,000 MI355X GPUs across more than 500 nodes. Those capacity and commercial terms are time-sensitive; verify the current offer directly with OLCF before building a quote: OLCF Lux system information.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Compare these access options by cash price, eligibility, resource unit, minimum commitment, included storage, available capacity, and scheduling terms—not by an hourly number alone. An allocation can still involve project and opportunity costs, but do not present it as a cash charge per hour unless the facility specifies one.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to treat energy estimates
Energy is measured in units such as joules; it becomes a monetary cost only when multiplied by an applicable electricity or facility billing rate. In a cloud service, energy may be embedded in the service price rather than billed to the customer as a separate meter reading. Add a separate power line only when the actual arrangement charges for it.
An Oak Ridge National Laboratory tutorial slide titled “Energy Budget against Model Size (~P^2)” reports 284 gigajoules for a 22-billion-parameter model, 17.65 terajoules for a 175-billion-parameter model, and 662 terajoules for a 1-trillion-parameter model. The slide’s year is not stated. It says its estimates use iteration time, tokens per iteration, average active power, and total MI250X GPU-card count, with GPU-level energy measured using rocm-smi: OLCF tutorial slide deck. These are method- and hardware-specific estimates, not current prices or universal energy coefficients; they cannot be converted into a training bill without a relevant rate.
Use specifications as context, not a price proxy
A system’s advertised performance or power rating does not tell you how quickly it will complete your workload. For example, OLCF describes a Summit node as having six NVIDIA V100 GPUs and reports 13 MW peak system power consumption: OLCF Summit system page. That is system context, not a quote or a measurement of one model’s training run.
OLCF describes Lux as using MI355X accelerators with high-bandwidth memory, Slurm and Kubernetes scheduling, and included access to the Orion filesystem for Lux allocations. These details can inform a configuration comparison, but they do not replace benchmarking the workload or confirming deployment and access conditions with the facility.
Quick Recap
Make the estimate decision-ready
- Specify the workload and planned number of runs, not just the model size.
- Base runtime on a representative benchmark from the intended hardware configuration.
- State the resource unit clearly and use the current rate or allocation terms for the intended region or facility.
- Include storage, networking, data movement, setup, and operations where applicable.
- List assumptions for evaluation, checkpointing, restarts, and contingency instead of implying a universal overhead.
- Keep cash charges distinct from awarded resources and separately billed energy.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
