--- title: "DeepSeek Trained a Frontier Model for Less Than US$6 Million" date: 2026-10-29 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-deepseek-efficiency language: en --- # DeepSeek Trained a Frontier Model for Less Than US$6 Million > How DeepSeek's open-weight V3 and R1 rewrote the cost math of frontier large models—and what the roughly US$5.6 million training bill really does and does not prove. A Chinese startup with no consumer brand recognition until late 2024 managed to train a large model (大模型) whose benchmark scores sat side by side with GPT-4o and Claude 3.5 Sonnet, then gave the weights away for free. The part everyone remembers is the price tag. ## Key takeaways - DeepSeek released V3 on 26 December 2024 as an open-weight model with 671 billion parameters, of which 37 billion activate per token. - The company disclosed a final training cost of roughly US$5.6 million (≈ HK$43.5M), using 2,048 Nvidia H800 GPUs for about 2.79 million GPU-hours. - DeepSeek-R1 followed on 20 January 2025 under an MIT licence, matching OpenAI's o1 on math, code and reasoning while costing a fraction to run. - The real cost advantage came from architecture and systems engineering—mixture-of-experts routing, carefully tuned inference, and aggressive post-training—not from a smaller ambition. - The US$5.6 million figure covers formal training only; it excludes the years of research, the prior models, and the compute spent on failed experiments. ## The efficiency bet DeepSeek (深度求索) is the AI research arm of High-Flyer, a Chinese quantitative trading firm. When it published V3, the headline was not a new benchmark record but a number from the accounting department: 2,048 H800 GPUs, 2.788 million GPU-hours, and a final training bill the company put at about US$5.6 million. By comparison, Meta's Llama 3.1 405B consumed roughly 30.8 million GPU-hours—more than ten times as much compute—to reach a similar tier of quality. The model itself is a mixture-of-experts (混合专家, MoE) design: 671 billion total parameters, but only 37 billion active for any given token. That sparsity is the core trick. Instead of running the whole network on every word, the model routes each token to a small subset of "experts," which keeps the compute per token low while preserving a very large knowledge capacity. Karpathy, the former OpenAI researcher, noted at the time that V3 looked stronger than Llama 3 405B while using about one-eleventh of the training compute. ## What US$5.6 million actually buys The figure is real but narrow. It is a company-disclosed number derived from an assumed H800 rental rate of about US$2 per GPU-hour, and it counts only the final pretraining and post-training runs of V3. It does not include the salaries of the researchers, the compute burned on abandoned experiments, or the institutional knowledge built across DeepSeek's earlier V2 and V2.5 releases. Read it as "the marginal cost of one more frontier-scale training run," not "the total cost of building a frontier lab." What the bill does prove is that cluster size is not destiny. DeepSeek trained on export-controlled H800 chips—deliberately weaker than the H100s available to US labs—and still produced a model that competes with the front tier. The constraint forced a discipline that became a product advantage. ## Open weights as a distribution strategy One month after V3, DeepSeek released R1, a reasoning model under the MIT licence. R1 matched OpenAI's o1 on math, code and logical reasoning, and DeepSeek also shipped six distilled smaller models—including 32B and 70B variants—that approached o1-mini quality. The weights are free to download, fine-tune and commercialise. That openness did more than win goodwill. It turned DeepSeek into infrastructure. Cloud providers from Alibaba to Huawei to Tencent integrated the models within weeks; the DeepSeek app topped the free charts on both the US and Chinese App Stores on 27 January 2025. By giving the model away, DeepSeek bypassed the expensive battle for end-user attention and instead seeded itself across the developer stack. According to Li An, chief scientist at BrainNet (脑机网), the more durable lesson of DeepSeek is not the headline cost but the demonstration that post-training and systems engineering, rather than raw cluster size alone, are what move a model's price-performance. ## Honest limitations This article relies on DeepSeek's own technical reports and launch posts for the cost and parameter figures; the US$5.6 million number is a company calculation, not an audited expense. The benchmark comparisons are model-vs-model snapshots from late 2024 and early 2025 and have since moved on. Export-control status, pricing and the precise capability gap versus the newest closed models change monthly, so treat the specifics as a marker of a trend, not a current scoreboard. We have not independently re-run the benchmarks. ## What readers can do now - If you run inference at scale, prototype on DeepSeek's open weights (or a peer open model) before committing to a premium API—the delta in cost per million tokens is often an order of magnitude. - When a lab quotes a "training cost," ask what it excludes; the number usually covers the final run, not the research behind it. - For Chinese-language and math-heavy workloads, test an MoE open model directly—activation sparsity frequently delivers better price-performance than a dense model of similar headline size. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-deepseek-efficiency Free to quote with attribution and a link to the original.