DeepSeek V3 Serving Cost Calculator: Cost per Million Tokens On-Prem
This free DeepSeek V3 serving cost calculator converts GPU capital cost into an effective cost per million tokens, so you can compare self-hosting V3 against API pricing on the same basis instead of guessing. Enter precision, context length, concurrency, and your expected monthly token volume, and the tool returns hardware footprint, GPU count, capital cost, and amortized cost per million tokens over a 3-year hardware life. V3 is the base model DeepSeek later refined into R1 through reinforcement learning, and it remains the more economical choice for enterprises that need strong general-purpose quality without R1's longer reasoning traces and higher effective token cost per response.
Your numbers
DeepSeek V3 was natively trained in FP8; that is the recommended production precision.
V3 supports up to 128K context via its multi-head latent attention design.
Simultaneous in-flight requests during peak hours.
VRAM per GPU; enter a matching price below.
H100 80GB runs roughly $25,000-$32,000; H200 runs $32,000-$40,000; B200 runs $45,000-$60,000 in 2026.
Combined input and output tokens processed per month across all applications served by this cluster.
Your results
Planning estimates only, excluding power, cooling, and engineering labor. Real cost per token also depends on achieved GPU utilization; idle capacity raises your effective rate substantially.
Get your full DeepSeek V3 cost model
We will email you a personalized cost-per-token breakdown against your projected traffic, plus a utilization checklist, and a Netray AI specialist will follow up with a deployment plan.
No spam. Your results stay private. Unsubscribe anytime.
How DeepSeek V3 runs on hardware you own
V3 shares R1's core architecture: 671B total parameters, 37B active per token, mixture-of-experts routing across 256 experts, and multi-head latent attention for compressed KV cache. Because V3 was not further trained to produce extended reasoning traces, its output lengths are shorter and more predictable than R1's for equivalent tasks, which makes serving capacity easier to plan and generally cheaper per completed response. The full 671B parameter set still has to reside in GPU memory regardless of the 37B active subset, so at FP8 you need roughly 671GB for weights before KV cache, putting a full-precision deployment at 9 or more H100-class GPUs.
- 671B total parameters at FP8 need roughly 705GB of VRAM including overhead, before KV cache.
- Shorter, more predictable output lengths than R1 make V3 easier to capacity-plan for high-volume workloads.
- MLA keeps KV cache small relative to the model's enormous weight footprint.
- A realistic minimum deployment is 9 H100-class GPUs at FP8, or 5-6 H200/B200 GPUs.
Quantization tradeoffs for DeepSeek V3
FP8 is V3's native training precision and the default for production serving, offering the best accuracy-to-memory ratio available for this model. INT4/GGUF quantization roughly halves memory again, bringing weight footprint to around 335GB, which can bring a self-hosted deployment within a single 8-GPU 80GB node once KV cache and overhead are accounted for. The accuracy cost of INT4 on V3 is generally milder than on R1, since V3's outputs are more direct and less dependent on long chains of intermediate reasoning that compound small per-step errors, but validate against your own tasks before committing to production.
- FP8 is V3's native format and the default for production serving.
- INT4/GGUF roughly halves memory, potentially fitting a single 8-GPU 80GB node.
- V3 tends to tolerate quantization better than R1 because its outputs are shorter and more direct.
- Cost per million tokens improves substantially at INT4 due to the lower GPU count required.
Serving DeepSeek V3 with vLLM and SGLang for cost efficiency
Sustained GPU utilization is what determines whether self-hosting V3 actually beats API pricing at your volume, and continuous batching in vLLM or SGLang is the mechanism that makes high utilization achievable. A cluster sized for peak concurrency but running at 20% average utilization has an effective cost per token several times higher than the same cluster running at 70-80%. Before committing capital, model your realistic utilization curve across a full week including nights and weekends, not just peak business hours, since that gap is usually where a self-hosting cost projection quietly falls apart in practice.
- Continuous batching is what makes high GPU utilization, and therefore low cost per token, achievable.
- Model utilization across a full week, not just peak hours, before committing to a hardware purchase.
- Multiple applications sharing one V3 deployment improves utilization and lowers effective cost per token.
- Request routing and queuing strategy matters as much as raw GPU count for realized throughput.
When to fine-tune DeepSeek V3 instead of prompting
V3 is a strong fine-tuning base for enterprises that have already committed to the hardware footprint and want to specialize the model for a house style, structured output format, or domain vocabulary, since LoRA on a model this size still trains meaningfully faster and cheaper than full fine-tuning. For most enterprises, though, the more common pattern is not fine-tuning V3 itself but using it as a teacher model: generating high-quality synthetic training data or distillation targets for a much smaller model that will actually run in production at a fraction of the cost. Reserve direct fine-tuning of V3 for cases where its scale is genuinely load-bearing for quality.
How Netray helps you get the real cost of DeepSeek V3
Netray builds the utilization discipline into V3 deployments that most in-house teams underestimate: workload consolidation across applications, request routing to keep GPUs busy, and honest capacity planning against real traffic patterns rather than peak-hour assumptions. For manufacturers with ITAR or CMMC constraints where sending data to a public API is not an option, we size V3 or a smaller alternative against your actual token volume and produce a real cost-per-token figure you can defend to finance. Engagements start with a two-week traffic and cost modeling phase using your own usage data.
Frequently Asked Questions
At what token volume does self-hosting DeepSeek V3 beat API pricing?
It depends heavily on achieved GPU utilization, but as a rough guide, self-hosting typically becomes competitive once monthly volume exceeds several tens of billions of tokens and you can sustain utilization above 50-60%. Below that volume, the fixed capital cost of a multi-GPU cluster rarely pays back against API pricing within a reasonable timeframe. Run this calculator against your real projected volume and compare the result to your current API spend before committing capital.
Why does GPU utilization matter so much for cost per token?
Because hardware cost is fixed once purchased, but the tokens it produces are not. A cluster running at 20% average utilization produces a fifth of the tokens the same cluster could produce at full utilization, while the amortized hardware bill stays identical. That means effective cost per token is inversely proportional to realized utilization, which is why consolidating workloads and smoothing traffic across time zones matters as much as raw GPU count.
Is DeepSeek V3 cheaper to serve than R1 for the same task?
Usually yes, when the task does not require R1's extended reasoning. V3 produces shorter, more direct outputs for equivalent questions, which means lower decode time and lower effective cost per completed response even though both models share the same 671B-parameter memory footprint and hardware requirements. If your workload does not need step-by-step reasoning traces, V3 is generally the more economical choice between the two.
Does this calculator include power and cooling costs?
No, deliberately, to keep the comparison focused on capital cost per token. Power and cooling for a multi-GPU cluster of this size typically add 15-25% to total cost of ownership over a 3-year period. Use this tool alongside a dedicated power and cooling calculator, and factor in engineering and operations labor, before presenting a final total cost of ownership figure to finance.
Get a real cost-per-token model for DeepSeek V3 on-prem, benchmarked against your actual traffic volume.
Related Tools
DeepSeek R1 GPU Requirements Calculator
Estimate the GPU count and capital cost required to self-host DeepSeek R1, a 671B-parameter mixture-of-experts reasoning model with only 37B active per token.
On-Prem AIKimi K2 Deployment Cost Calculator
Estimate the multi-GPU cluster cost required to self-host Kimi K2, a roughly 1 trillion parameter mixture-of-experts model with only 32B active per token.
On-Prem AILLM API vs Self-Hosted Cost Calculator
Model your API bill from requests and token mix, compare it against an all-in self-hosted monthly cost, and see monthly and annual savings.
Go Deeper
DeepSeek V3 Enterprise Deployment Guide
Deploy DeepSeek V3 on-prem: 671B MoE architecture, MIT license, GPU sizing by quantization, vLLM serving setup, fine-tuning options, and when to use it.
DeepSeek R1 On-Prem Deployment: An Enterprise Guide
Deploy DeepSeek R1 on-prem: 671B MoE architecture, MIT license, distilled model sizing, GPU requirements, serving stack, and when reasoning models fit.
Enterprise GPU Cluster Planning for AI Workloads
Plan an enterprise GPU cluster for AI workloads: H100 vs L40S sizing, networking, power, cooling, and cost models for on-prem LLM inference and training.