DeepSeek’s R1 model matches OpenAI’s performance while using less energy and less advanced chips. The company’s architecture enables a massive parameter count without the massive compute bill. DeepSeek-V3 uses a Mixture-of-Experts (MoE) design with 671 billion total parameters. Only 37 billion parameters activate for any given input token. This sparse activation reduces computational costs because the model does not run every parameter on every token. A dense model with 671 billion parameters would require every token to pass through all 671 billion parameters, which makes the cost per token scale linearly with model size. DeepSeek’s approach breaks this link.
The efficiency of the MoE architecture impacts the training economics. DeepSeek reported its full training run, including pre-training and post-training, at 2.788 million GPU hours on H800 hardware. This training cost roughly $5.576 million. In contrast, typical training for comparable models ranges from $100 million to $1 billion. The sparsity of the routing mechanism allows the model to achieve higher quality at equal compute than a dense architecture.
| Feature | DeepSeek-V3 Specification |
|---|---|
| Total Parameters | 671B |
| Active Parameters per Token | 37B |
| Training Cost | $5.576M |
| Training Hardware | 512 H800 GPUs |
| Training Duration | 2.788M GPU hours |
DeepSeek engineers also developed context caching on disk to lower inference costs. This mechanism enables the system to retrieve cached segments for prefix tokens that share the same identity as recent requests. For cache hits, DeepSeek charges $0.014 per million tokens. This price represents a 90% discount over the normal rate. BytePlus analysis shows this leads to 74% to 75% savings for DeepSeek Chat and DeepSeek Reasoner. A cache hit for DeepSeek Chat costs $0.07 instead of $0.27.
Hardware constraints and software optimizations
DeepSeek faces chip export restrictions that limit access to the most advanced hardware. The company trains and serves its models on NVIDIA H800 GPUs, which are the Chinese variant of the H100. These chips are cheaper and more available for Chinese companies than the H100 or A100. DeepSeek also announced support for Huawei Ascend NPUs and Cambricon accelerators to diversify hardware options. The company uses specialized engineering to squeeze performance out of these accelerators, including open-source tools like FlashMLA and DeepEP.
The company uses quantization to increase inference speed. DeepSeek-R1 variants can use 8-bit or 4-bit numerical representations instead of 16 or 32-bit floats. NeuralMagic reports that INT8 models achieve near-perfect accuracy recovery, while INT4 recovers more than 97% of performance on benchmarks. These quantization strategies allow for up to 4x inference speedups on typical hardware. Faster models reduce the time spent on expensive GPUs, which lowers the cost per inference.
DeepSeek’s use of Multi-Head Latent Attention (MLA) further reduces memory usage. MLA compresses the Key-Value (KV) cache into a latent vector. This compression allows for faster processing and supports longer context lengths. DeepSeek-V2 accommodated up to 128,000 tokens using this method. If you are managing a massive enterprise data migration, you will find that the memory requirements for a 671 billion parameter model remain high regardless of the active parameter count. The MoE architecture shifts the bottleneck from compute to memory because the router needs access to all experts to decide which to activate.
Deployment strategies for enterprise workloads
Enterprise customers can choose between using the DeepSeek API or self-hosting the models. DeepSeek offers open-source versions of its models, including R1 and R2. Customers run these models on generic servers or laptops without paying licensing fees. In this model, the only marginal cost involves hardware usage, power, and labor. This self-hosting option provides a capital expenditure model as an alternative to metered API fees.
The API pricing for DeepSeek remains competitive against Western providers. DeepSeek-V2’s input rate of $0.14 per million tokens matches the $0.15 rate for GPT-4o Mini. DeepSeek’s R1 model performs on par with OpenAI’s o1 in mathematics, coding, and logic. DeepSeek also offers distilled versions of its reasoning models. The DeepSeek-R1-Distill-Qwen-32B model outperforms OpenAI’s o1-mini in dense model benchmarks.
| API Pricing (per 1M tokens) | DeepSeek-V2 Rate | GPT-4o Mini Rate |
|---|---|---|
| Input Rate | $0.14 | $0.15 |
DeepSeek’s models shape responses to fit the limits set by the Chinese government. This regulation addresses safety, privacy, and ethics. While the efficiency of the MoE architecture provides massive savings, the expert skew problem can still cause costs to drift. If more than 40% of tokens land on a single expert, the real cost per token rises above what the active-parameter math predicts. Will the high memory requirements for trillion-parameter MoE models eventually limit the adoption of these models for smaller enterprises? DeepSeek’s technical innovations provide a path for high-performance AI at low costs.
