The useful version
- DeepSeek builds large language and reasoning models, a free chatbot, downloadable model weights and a paid developer API.
- Its famous $5.576 million figure describes estimated GPU cost for one V3 training run - not the total cost of DeepSeek or R1.
- Its first pure-reinforcement-learning model reasoned well but wrote badly. Cold-start examples and staged training made the research usable.
- The copyable lesson is to design around the bottleneck, then use openness as distribution. It works less well when privacy, political neutrality or turnkey safety matter most.
On January 27, 2025, a blue cartoon whale frightened Wall Street. The whale belonged to DeepSeek, a two-year-old research company in Hangzhou. Its new reasoning model had become the talking point of the weekend; its free assistant was sitting above ChatGPT in the US App Store. By the closing bell, Nvidia had fallen 16.9 percent and lost roughly $593 billion in market value. This was a peculiar amount of financial drama for software anyone could download.
The panic rested on one seductive proposition: perhaps frontier AI did not require a frontier-sized bill. DeepSeek had reported that training V3 consumed 2.788 million Nvidia H800 GPU-hours. At an assumed rental price of $2 an hour, that was $5.576 million. Headlines rounded it to six. Traders compared it with the enormous budgets of American labs. A technical footnote became an economic referendum.
The number was small. The footnote was large.
The figure was real within its stated boundary, and misleading outside it. It did not include the cost of earlier experiments, failed runs, salaries, data preparation, or the infrastructure accumulated by High-Flyer, the quantitative hedge fund founded by DeepSeek chief Liang Wenfeng. It was not the cost of training R1. It was certainly not the price of creating a company capable of training R1. DeepSeek was constrained, but it was not a garage operation improvising on a gaming laptop.
IN: the reported V3 pre-training, context extension and post-training compute. OUT: prior research, ablations, people, data, owned hardware, facilities and R1's later post-training.
Yet the caveat does not erase the accomplishment. DeepSeek's engineers built V3 so that only 37 billion of its 671 billion parameters activate for each token. They compressed the memory burden of attention with Multi-head Latent Attention. They trained in lower-precision FP8 arithmetic and worked on the ugly, consequential layer where chips talk to one another. The company did not discover a coupon for GPUs. It stacked many savings until efficiency became visible on an invoice.
“Money has never been the problem for us; bans on shipments of advanced chips are the problem.”Liang Wenfeng, in a translated interview
First, the model learned to think badly
R1's most useful story begins with an awkward prototype. DeepSeek-R1-Zero was trained with large-scale reinforcement learning and no supervised fine-tuning at the start. Give the system a problem, reward the right answer and let it discover a route. Reasoning behavior emerged. The model paused, reconsidered and corrected itself - the paper even records an “aha moment” marked by a sudden rise in the word “wait.”
It was also an unruly colleague. Its reasoning could be repetitive, difficult to read and prone to mixing Chinese and English. A benchmark can reward the destination while ignoring the manners of the journey. DeepSeek had produced evidence that pure reinforcement learning could cultivate reasoning, but not yet a product people would enjoy using.
That failure changed the recipe. For R1, the team supplied thousands of “cold-start” examples showing a cleaner, human-readable reasoning format before applying reinforcement learning. It added a language-consistency reward, generated more training examples, filtered them and fine-tuned again. The breakthrough was not stubborn allegiance to a pristine research idea. It was accepting that a reasoning engine also needed an editor.
Give away the engine. Charge for the ignition.
DeepSeek has three overlapping constituencies. Consumers get a free web and mobile assistant for writing, search, code and problem solving. Developers can call an OpenAI-compatible API and pay by usage. Researchers and companies can download model weights, modify them and, for R1, use them commercially under an MIT license, while observing the licenses attached to distilled base models.
That looks generous because it is generous. It is also distribution. Open weights allowed R1 to appear in local tools, Hugging Face workflows, Amazon Bedrock, SageMaker and Microsoft Azure AI Foundry. In March 2025, AWS said thousands of customers had already deployed R1 through Bedrock. DeepSeek did not need to win every enterprise security review itself; cloud platforms could wrap the model in their own controls and sell the managed experience.
What travels freely
Model weights, research ideas, community modifications and smaller distilled variants.
What earns money
Convenient hosted inference through DeepSeek's API, with no cluster to provision or model to operate.
The distinction matters because “DeepSeek” is not one risk. Running open weights inside a controlled environment is different from typing confidential material into the official app. DeepSeek's hosted privacy policy places collected data on servers in China. Regulators challenged the service over disclosure and data handling; Italy blocked the app during an investigation. Researchers and public agencies also raised concerns about security and refusals around politically sensitive Chinese topics. Open weights offer deployment choice, but they do not perform compliance work on the buyer's behalf.
The hedge fund under the laboratory
Liang's route into AI ran through quantitative trading. High-Flyer had money, mathematicians, machine-learning habits and a large stockpile of Nvidia A100 chips before export controls tightened. In 2023, Liang created DeepSeek to pursue foundational models beyond finance. Public accounts describe a flat, research-heavy organization that recruits promising graduates from Chinese universities and lets small project teams follow technical questions.
This explains the apparent contradiction in DeepSeek's strategy. The company publishes the artifact but guards the capacity to make the next artifact. Liang has called the team the moat. An open release lets the world test, adapt and distribute today's model; the lab accumulates the tacit knowledge to build tomorrow's. The product can be copied byte for byte. The sequence of judgments that produced it is harder to download.
What another company can actually copy
Most companies cannot assemble thousands of GPUs or hire a concentrated group of model researchers. They can copy the operating logic. DeepSeek treated a constraint as a specification, published enough to attract an ecosystem, and turned an embarrassing prototype into a source of training design.
Name the bottleneck
Optimize the scarce resource, not the fashionable metric. For DeepSeek, memory, communication and chip access mattered more than parameter bragging rights.
Reward outcomes
Where answers can be checked - math, code, logic - objective reinforcement signals can reduce dependence on expensive hand labeling.
Productize the failure
R1-Zero proved the idea and exposed the flaw. Cold-start examples fixed presentation without discarding the underlying learning method.
Separate artifact from service
Make the core portable, then sell convenience, uptime and integration to customers who do not want to operate it.
The approach has conditions. Objective rewards are easier to design for code and mathematics than for taste, strategy or ambiguous human judgment. Sparse architectures demand deep systems skill. Open distribution can accelerate adoption, but it also weakens control over misuse and makes differentiation migrate toward operations, data and the next release. And a low sticker price cannot compensate for privacy rules, safety requirements or political constraints that exclude the hosted service.
DeepSeek now competes with OpenAI, Anthropic, Google, Meta, xAI, Alibaba's Qwen, Moonshot's Kimi and a crowded field of model makers. Its V4 family has pushed into million-token context and agentic work. Features will converge; benchmark leads will rotate. The durable contribution is a change in accounting. After DeepSeek, a giant compute budget is no longer self-justifying. Someone in the room will ask what each activated parameter is doing. The blue whale made that question unavoidable.