Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThere is no universal cost or electricity figure for an AI prompt. Estimate API charges and energy separately: API prices depend on the provider’s billable categories and current rates, while energy depends on the model, workload, hardware, utilization and what parts of the serving system you count.
Estimate hosted API charges from the provider’s current rates
For a hosted API, the provider’s bill is not a direct measurement of electricity. Calculate each billable category using the applicable published rate:
As an Amazon Associate I earn from qualifying purchases.
API cost = (input tokens ÷ billing unit × input rate) + (output tokens ÷ billing unit × output rate) + other applicable charges.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Check the selected model or service tier’s official pricing page on the day you estimate. Include any distinct prices for cached input, reasoning tokens, tools, images, audio or batch processing when they apply. Record the provider, model or tier, currency, billing unit, relevant region and rate-check date. Rates can change; a remembered price is not a current quote. The calculation needs the chosen provider’s current rate schedule, so no single price can be supplied here.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
Turn a workload into an estimate
Estimate the number of input and output tokens for the workload, then apply the matching rates to each category. For a workload of many requests, calculate the expected totals across all requests. Agentic or reasoning workflows can use substantially more tokens than a short answer, so estimate the whole workflow rather than counting only the user’s prompt.
Estimate self-hosted compute cost using throughput
For self-hosting, an hourly accelerator price alone does not tell you the cost per token. A practical compute-cost approximation is:
Compute cost per million tokens = effective infrastructure cost per hour ÷ delivered tokens per hour × 1,000,000.
Rank #2
- Delivers 25 TFLOPS of AI inference performance using four MemryX MX3 accelerators in a compact M.2 2280 M-key module.
- Connects through PCI Express Gen 3 and above for straightforward integration into embedded and edge systems with an M.2 M-key slot.
- Includes 42 MB of distributed on-chip weight storage to support low-latency, batch-1 pipelined inference.
- Designed for low-power deployment with a 1 W–12 W operating range and host-managed airflow requirements.
- Supports ONNX, TensorFlow, Keras, and TFLite models through the MemryX software stack on Linux and Windows hosts.
Use delivered throughput measured or benchmarked with the same model, precision, prompt and output lengths, batch size or concurrency, serving stack, and latency target you expect to use. Include the host and other infrastructure costs relevant to your estimate. For a fuller total cost of ownership, account for hardware purchase amortization or lease, host CPU and memory, networking, storage, software, facility and power costs, and operations.
Account for utilization and task size
Include utilization in the calculation: paid-for capacity that sits idle lowers delivered tokens per hour and raises cost per token. Compare cost per completed task as well as cost per token when tasks use different numbers of tokens; the cheaper token may not produce the cheaper completed task if one workflow needs more tokens.
NVIDIA’s token-economics guide recommends looking at both throughput per megawatt and cost per token. Its H100/B200 example assumes $3.50 per H100-hour and $6.00 per B200-hour, then compares illustrative workloads. Those are assumptions used to demonstrate the calculation, not current universal prices or an independent cross-vendor benchmark. Read NVIDIA’s token economics guide.
Rank #3
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
The same NVIDIA page reports a vendor-published GB300 NVL72/Hopper comparison of $4.20 versus $0.12 per million tokens and 54,000 versus 2.8 million tokens per second per megawatt. It also cites a SemiAnalysis InferenceX result of $0.123 per million tokens at 116 tokens per second per user interactivity, as of April 2026. These are benchmark- and workload-specific claims, not a typical cost a reader should expect to pay.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Calculate electricity use and cost from power
If you have average power for the system being measured and runtime, calculate energy in kilowatt-hours:
Energy (kWh) = average power (W) × runtime (hours) ÷ 1,000.
Then apply the electricity tariff for the relevant location and billing period:
Electricity cost = energy (kWh) × applicable tariff ($/kWh).
Use your actual tariff rather than a generic price. If estimating from inference throughput instead of metered runtime, first estimate energy per token from measured power and delivered token rate, or use an explicit hardware-and-workload model. Keep prompt processing (prefill) separate from text generation (decode) where possible; they place different demands on the hardware.
Best Value
- Memory Size: 16 GB GDDR6 ECC.
- Memory Bus Width: 128-bit.
- Memory Bandwidth: 200 GB/s.
- CUDA Cores: 1280.
- Peak Single Precision floating point performance: 18 Tflops (GPU Boost Clocks).
Be explicit about the system boundary
Say what the power measurement or model includes. A GPU-only figure is not the same as total serving energy: a broader estimate may include idle provisioned machines, host CPU and RAM, and data-center power distribution and cooling. Google Cloud’s serving methodology accounts for achieved accelerator utilization, idle machines, host CPU and RAM, and data-center overhead. Google’s post explains why a chip-only estimate leaves out operational load: Google Cloud: Measuring the environmental impact of AI inference.
One way to model GPU-level inference energy is to account for model parameters and memory traffic, separating computation, parameter access, KV-cache writes and attention reads. The authors of this analytical approach describe it as an approximation, not a replacement for physical power measurement. Read the GPU-level energy analysis.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why published per-query energy estimates differ
Published estimates can use different models, workloads and system boundaries, so they should not be treated as direct comparisons unless those conditions match.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Estimate | What it covers | How to interpret it |
|---|---|---|
| 0.24 Wh per median Gemini Apps text prompt; 0.03 gCO2e and 0.26 mL water, reported by Google in 2025 | Google’s broader serving methodology, including production utilization, idle machines, host CPU and RAM, and data-center overhead | Google’s own estimate for Gemini Apps text prompts; it does not establish energy use for other models or providers. Google Cloud, August 21, 2025 |
| 0.10 Wh per median Gemini Apps text prompt, reported by Google in 2025 | Google’s active-TPU/GPU-only calculation | A narrower chip-energy boundary than Google’s 0.24 Wh figure; not a general estimate for other systems. Google Cloud, August 21, 2025 |
| 0.34 Wh median per query, with an interquartile range of 0.18–0.67 Wh, reported by Oviedo et al. in 2025 | Modeled realistic workload assumptions for frontier-scale models over 200 billion parameters on an H100 node | A modeled estimate, not a direct comparison with Google’s Gemini disclosure. The populations, serving systems and boundaries differ. Oviedo et al., 2025 |
| 4.32 Wh per query, reported by Oviedo et al. in 2025 | A test-time-scaling scenario using 15 times more tokens per typical query | A scenario estimate under the paper’s assumptions, not a typical per-prompt figure. Oviedo et al., 2025 |
These examples show why a per-prompt energy number needs its workload and boundary beside it. Google Cloud’s August 21, 2025 post, co-authored by Amin Vahdat and Jeff Dean, states: “Measuring the footprint of AI serving workloads isn’t simple.”
Identify the inputs that can change your estimate
When comparing estimates or services, match the conditions that materially affect cost and energy:
- Model and workload: model size and architecture, quantization or precision, prompt length, generated-token count, and whether the work is prefill or decode.
- Serving pattern: batch size, concurrency, context length and KV-cache behavior, serving software, utilization, and latency target.
- System boundary: accelerator power alone versus a system or facility estimate that includes host equipment, idle capacity, cooling and power delivery.
- Local costs: the applicable electricity tariff and, for self-hosting, infrastructure and operational costs.
For a meaningful comparison, align cost per million input and output tokens or cost per task, energy per request or token, throughput per watt or megawatt, latency at your intended concurrency, model quality, and included infrastructure. An hourly compute price by itself cannot establish which option is cheapest.
Quick Recap
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.
