DeepSeek (company)

DeepSeek is a Chinese artificial intelligence company based in Hangzhou, founded in July 2023 and financed until 2026 by the quantitative hedge fund High-Flyer. It builds the DeepSeek family of language models and sells access to them through an API and a consumer application 1.

The company became internationally known in January 2025, when it released R1, a reasoning model whose weights could be downloaded, and whose technical report described a training budget in the millions rather than the hundreds of millions of dollars. The release coincided with a sharp fall in the share price of NVIDIA and prompted a reassessment of how much capital frontier AI actually requires 2.

DeepSeek is unusual among AI companies in almost every respect: it was funded by a trading firm rather than venture capital, it published its methods in detail, it charged prices well below its competitors, and it remained a small research organisation while its models were being downloaded millions of times 3.

Liang Wenfeng studied engineering at Zhejiang University and worked on machine learning for quantitative trading before co-founding High-Flyer, a Hangzhou hedge fund, in 2015. High-Flyer applied deep learning to financial markets and grew into one of China's largest quantitative funds, accumulating both capital and a stock of GPUs that later became DeepSeek's first computing resource 4.

DeepSeek was incorporated in July 2023 with Liang as controlling shareholder. Its early hiring differed from that of its competitors: rather than recruiting senior researchers from American laboratories, it hired recent graduates from Chinese universities and trained them internally, and it kept the organisation small. Liang has said the company's aim was to pursue general intelligence rather than to build products for a market 4.

The funding model followed from that aim. High-Flyer provided capital, and DeepSeek took no outside investment until 2026. It described itself as not seeking financing, not commercialising aggressively and not planning a listing, a stance that was unusual in a sector where capital raising had become the main competitive activity 5.

That stance changed in 2026. According to Chinese financial press, the company completed a first external round of about 51 billion yuan, roughly 7.5 billion US dollars, at a post-money valuation near 400 billion yuan, with Liang himself contributing about 20 billion yuan and investors holding their interests through a limited partnership managed by him, subject to a five-year lock-up 5.

DeepSeek's first models, released in late 2023, were specialised: DeepSeek Coder for programming and a 67-billion-parameter general model, both trained from scratch and published with technical reports. They established a pattern the company has followed since: a paper, a set of weights, and a price below the market 3.

DeepSeek-V2, in May 2024, introduced multi-head latent attention, a compression scheme for the key-value cache that reduces memory use during inference, together with a mixture-of-experts design of 236 billion total parameters and 21 billion active per token. The architecture made inference cheap, which is why V2 was followed by a price war in the Chinese API market 6.

V3, in December 2024, scaled the design to 671 billion total parameters with 37 billion active, and added multi-token prediction during training. It was trained on 14.8 trillion tokens with FP8 mixed precision, a lower-precision format that reduces memory and compute requirements, and it used a pipeline schedule that overlaps computation and communication across GPUs 6.

R1, in January 2025, applied reinforcement learning to elicit long reasoning chains without training first on human-written reasoning traces. The company also released distilled versions that transferred the behaviour to smaller models based on Qwen and Llama, which made reasoning capabilities available to users who could not run the full model 2. Subsequent releases were incremental: V3-0324 in March 2025, R1-0528 in May, V3.1 in August with a hybrid thinking mode, and V3.2-Exp in September with a sparse attention mechanism. The cadence slowed in 2026 as the company shifted resources toward infrastructure and products 1.

The V3 report states that the final training run used about 2.79 million H800 GPU-hours, which the company costed at roughly 5.6 million US dollars at an assumed rental rate of two dollars per GPU-hour. That figure — widely reported as the cost of building a frontier model — was the most consequential number in the paper 6.

The company was explicit about what the figure excluded. It covered the final run only, not the research, ablation experiments, failed runs, data preparation, hardware purchase or the salaries of the team. The report says as much in a footnote, and DeepSeek's researchers have repeated the qualification since 6.

The techniques behind the number are real. Multi-head latent attention compresses the cache that stores context, mixture-of-experts routing activates only a fraction of the parameters for each token, FP8 training halves the memory traffic for weights and activations, and multi-token prediction lets the model learn from several future tokens at once. Together they reduce the compute needed for a given level of capability 6.

Whether the result generalises is another matter. DeepSeek trained on a large, well-curated corpus and benefited from published research by others, including the Transformer architecture and open-source training frameworks. The efficiency came from engineering a known design rather than from a new method, and the company has never claimed otherwise 2.

DeepSeek publishes model weights. R1 was released under the MIT licence, which permits commercial use, modification and redistribution with minimal conditions; V3 shipped under a separate model licence with use restrictions, including a prohibition on using the model to train other models without permission. The distinction matters to companies that want to deploy the models commercially 1.

For researchers, the value lies in being able to inspect and run the models locally. For companies, it allows deployment without sending data to a third-party API, which is attractive in regulated industries and in countries where American providers are unavailable. For DeepSeek, openness created adoption and influence that a closed API would not have produced 3.

Openness is partial. Training data, training code and the full pipeline are not released, so the models can be studied and used but not reproduced. The licence on V3 also imposes conditions that the Open Source Initiative's definition would not accept, which is why the company's models are usually described as open-weight rather than open-source 1.

Publishing weights has commercial consequences. Companies can download the model and serve it themselves, bypassing DeepSeek's API, which limits the revenue the company can earn from its own work. DeepSeek has accepted that trade-off, and its API pricing — a few cents per million tokens at the low end — suggests that it does not expect to earn much from inference 7.

R1 was released on 20 January 2025, shortly after the American inauguration, and its effect was immediate. The application reached the top of Apple's US App Store within a week, a first for a Chinese consumer AI product, and the model's weights were downloaded in volumes that briefly overwhelmed Hugging Face 3.

The market reaction was more dramatic than the technical one. On 27 January 2025 NVIDIA's share price fell by about 17 per cent, wiping out nearly 600 billion dollars of market value in a day — the largest one-day loss in the history of the American stock market. Investors read R1 as evidence that frontier capability could be reached with far less computing than assumed 3.

The reaction was partly a misreading. DeepSeek's efficiency gains lowered the cost of a given capability, but the industry's response was to spend more, not less, because cheaper inference made more applications viable. NVIDIA's shares recovered within months, and demand for its chips continued to exceed supply 3.

R1 also changed the politics of AI. Its release was treated in Washington as evidence that export controls had not prevented Chinese progress, and in Beijing as evidence that domestic laboratories could compete at the frontier. Both readings were probably correct, and both made the company a subject of policy attention it had not sought 8.

The first regulatory action came from Italy. On 28 January 2025 the Garante, the Italian data protection authority, asked DeepSeek to explain what personal data it collected, on what legal basis, and whether the data was stored in China. Two days later it ordered a limitation on the processing of Italian users' data, and DeepSeek's application was removed from Italian app stores. The company had argued that it did not operate in Italy and that European law did not apply 8.

Other governments followed in different forms. Several countries restricted the application on official devices, and data-protection authorities in Europe opened inquiries. The measures varied from binding bans to guidance, and most were justified by the lack of clarity about where data was stored and who could access it 8.

Researchers also documented content controls. DeepSeek's models declined to answer questions about certain political subjects and, in some tests, produced answers consistent with official narratives. The company has not published its content policy in detail, and the behaviour is consistent with the requirements that Chinese law places on publicly available generative services 3.

Security researchers found an exposed database belonging to DeepSeek in January 2025 that contained chat logs and other data, accessible without authentication. The company fixed the exposure after being notified. A separate and unresolved dispute concerns training data: OpenAI said in early 2025 that it had evidence a Chinese laboratory had distilled its models through the API, without naming the company, and American officials made similar claims. The allegation has not been independently established, and DeepSeek has not published a detailed response 3.

The cost figure is the most contested number in the company's record. DeepSeek's 5.6 million dollars refers to one training run and excludes the research that made it possible; critics have described the figure as misleading when presented as the total cost of building V3, while supporters note that the company said what it excluded. Both are describing the same footnote 6.

Benchmark parity is contested in the same way as every vendor's claims. DeepSeek reported results close to leading American models on mathematics and coding, and independent evaluations broadly confirmed competitiveness, with weaker performance on some general knowledge and safety evaluations. The company's evaluations were run by the company, and the tests it chose favoured the skills its models were trained for 2.

The distillation allegation remains unresolved. It is plausible that a laboratory operating under export controls would seek to learn from published outputs, and it is also plausible that a well-funded team with a large corpus could reproduce similar behaviour independently. No public evidence has settled the question 3.

A fourth disagreement concerns the company's direction. Its 2026 financing, its investment in its own data centres and its hiring of product teams suggest a shift from a research laboratory to a full-stack platform. Whether that is a necessary evolution or an abandonment of the strategy that made it distinctive is a matter of interpretation 5.

DeepSeek's significance is not that it built the best model in the world. It is that it demonstrated, credibly and publicly, that a small team with a modest budget could build models close to the frontier, release them, and force larger companies to justify their spending. That demonstration changed the terms of the debate about compute, openness and the geography of AI 3.

Commercially, the company remains small. Its API revenue is limited by its own low prices and by customers who download the weights instead. The 2026 round gave it capital for data centres and hiring, but it also introduced investors, a valuation and expectations that the earlier structure had avoided 5.

For the encyclopedia, DeepSeek is the strongest example of efficiency as a strategy. It did not outspend its competitors; it out-engineered them on a narrow set of problems, published the results, and let the consequences follow. Whether that approach survives the next generation of models, which will require more capital and more compute, is the open question 1.