For most of January 2025 DeepSeek was a name known to people who read model cards. By the end of the month it had been the subject of a letter from the Irish Data Protection Commission, a block in Italy and a record one-day loss on the world's most valuable chipmaker. This briefing is an attempt to separate the three things that happened, which were a model release, a market reaction and a regulatory reaction, and to say what each of them meant. It is written at the start of February with the sources available then.
What was released
On 26 December 2024 DeepSeek, a Hangzhou lab spun out of the quantitative hedge fund High-Flyer, published DeepSeek-V3: a mixture-of-experts model with 671 billion parameters in total, 37 billion of them active for any given token, trained on 14.8 trillion tokens using 2.788 million hours of Nvidia H800 time.[1] The technical report priced that at $5.576m at an assumed $2 per GPU-hour, and said in the same passage that the figure covered the final training run only, not the research and ablation experiments that preceded it.[1]
On 20 January it released DeepSeek-R1, a reasoning model built on V3, under an MIT licence, alongside six smaller distilled models from 1.5 to 70 billion parameters.[2] The paper set R1 beside OpenAI's o1: 79.8% against 79.2% on the AIME 2024 maths competition, 97.3% against 96.4% on MATH-500, 49.2% against 48.9% on SWE-bench Verified, and 71.5% against 75.7% on GPQA Diamond, where o1 was still clearly ahead.[3] The API price was $0.55 per million input tokens, $0.14 on a cache hit, and $2.19 per million output tokens. OpenAI's o1 listed at $15 and $60.[2] That last line is what moved the market.
The four techniques underneath
The interesting thing about V3 is not that it is large but that it is cheap to run, and that comes from four choices, each published in enough detail to copy.
- Compressed attention. Multi-head Latent Attention projects the keys and values a model must hold in memory into a small latent vector before caching them, instead of storing them per attention head. DeepSeek measured the effect on its previous model: a 93.3% reduction in the size of that cache and 5.76 times the maximum generation throughput. Long context stops being the expensive part.[4]
- Sparse experts, balanced without a penalty. Each layer holds 256 routed experts plus one that is always on, and eight of the 256 run for any given token. Most systems keep experts evenly loaded with an auxiliary loss that fights the main training objective. DeepSeek instead nudges a per-expert bias up when an expert is idle and down when it is busy, and switches the nudging off for the last 500 billion tokens of training.[1]
- Predicting two tokens at once. A second small module predicts one token further ahead during training, so each example teaches more, and the same module can be reused at inference to draft ahead and run faster.[1]
- Eight-bit arithmetic. The heavy matrix multiplications run in FP8 rather than the usual 16-bit, with fine-grained scaling and periodic promotion to higher precision so errors do not accumulate. To make the chips talk to each other efficiently at that precision the team wrote its own low-level GPU instructions, said so in the paper, and recommended that future hardware make the workaround unnecessary.[1]
None of these was invented in December. What was new was stacking them at frontier scale and publishing the recipe, including the parts that are ordinarily a trade secret.
Learning to reason without being shown how
The claim that mattered to researchers was about training rather than architecture. The usual way to teach a model to reason is to show it worked examples written by people. DeepSeek's method, Group Relative Policy Optimization, had been published a year earlier in its mathematics paper: for each question the model produces a group of answers, each is scored by a checker rather than a human, and each answer's reward is judged against the average of its own group. Because the group supplies the baseline, there is no second network estimating value, which is what makes it cheap to run.[5]
An intermediate model, R1-Zero, was trained that way directly on the base model with no supervised examples at all, and developed behaviours nobody scripted: longer chains of reasoning, and pauses to re-check its own earlier steps. The paper prints a transcript in which the model interrupts itself: "That's an aha moment I can flag here."[3] It is equally direct about what was wrong with it. R1-Zero, the authors write, "struggles with challenges like poor readability, and language mixing": with nothing in the reward for staying in one language, it drifted between Chinese and English mid-answer.[3] The released R1 fixed that with a few thousand hand-curated starting examples and a fixed output format before the reinforcement learning began. That is a fair summary of what these methods give you and what they do not: capability from the reward, legibility from the humans.
The week the market noticed
- $589bn
- Fall in Nvidia's market value on Monday 27 January, the largest single-day loss in US market history [6]
- −3.1%
- The Nasdaq Composite the same day. The Dow, light on AI infrastructure, closed up [7]
- $0.55
- R1's price per million input tokens, against $15 for OpenAI's o1 [2]
Over the weekend of 25 and 26 January DeepSeek's free assistant reached the top of Apple's app chart in the United States, and the venture capitalist Marc Andreessen called R1 "AI's Sputnik moment".[8][9] On the Monday, Nvidia fell 17% and lost $589bn of market value, the largest one-day loss for any company on record.[6] The sharpest falls were not among the chipmakers but among the companies that would have built and powered the data centres: Vertiv, which makes cooling and power equipment, fell more than 30%; Vistra 28%; Oklo 25%; Constellation Energy about 20%; Siemens Energy 19.95% in Frankfurt; GE Vernova 19%.[10] The logic was straightforward. If frontier capability could be trained for millions rather than billions, the forecast demand for chips, buildings and electricity was too high.
The most prominent argument against that logic came from a buyer of the chips. On the day of the fall Satya Nadella invoked the Jevons paradox, the observation that making a resource cheaper to use tends to increase total consumption rather than reduce it, and argued that cheaper AI would mean more of it, not less.[11] Two days later, on Meta's results call, Mark Zuckerberg reaffirmed capital expenditure of $60bn to $65bn for 2025, up from $39bn, and said heavy infrastructure spending would be a long-term advantage.[12] Neither man was a disinterested observer. On the evidence of the eighteen months since, both were describing the world correctly.
What the $6m figure means
The number that carried the story was the wrong one to argue about, and by the end of the week the better analysts had said so. SemiAnalysis estimated on 31 January that DeepSeek had access to around 50,000 Nvidia Hopper-generation GPUs, that its server capital expenditure was about $1.6bn, and that the $6m covered the GPU rental for one pre-training run and nothing else: no research, no failed experiments, no data, no salaries.[13] That is also what DeepSeek's own report said, in a sentence most coverage did not reach.[1] The hardware did not appear overnight either. High-Flyer had been accumulating GPUs since 2015, reaching about a thousand by 2019 and roughly ten thousand A100s by 2021, ahead of the export controls that would have stopped it.[14]
Anthropic's chief executive Dario Amodei published a long essay on 29 January making the other half of the argument. Algorithmic progress, he wrote, shifts the cost of a given capability by roughly four times a year, so a model matching American models seven to ten months old at a lower cost was to be expected. "DeepSeek-V3 is not a unique breakthrough," he wrote, but "an expected point on an ongoing cost reduction curve."[15] His concern was not the model. It was that a Chinese lab had reached the point first, and what that implied for export controls.
Both readings are right and they are not in tension. The training run was cheap and the organisation was not. The trend line said this was coming; the surprise was who published it, how openly, and with what price tag. A business does not need to adjudicate the geopolitics to draw the practical conclusion: the cost of a given level of capability was falling faster than most procurement cycles, and would keep doing so.
The week the regulators noticed
Three separate problems arrived together, and the coverage tended to blur them. The first was availability: on 27 January DeepSeek suspended new registrations, citing large-scale malicious attacks on its service, at the moment it topped the app charts.[16] The second was security. On 29 January the research firm Wiz disclosed a DeepSeek database exposed to the open internet with no authentication at all, holding more than a million log entries including chat history, backend API secrets and operational metadata, and allowing arbitrary queries; DeepSeek secured it after being told.[17] The third was data protection, and it is the one Irish businesses asked us about.
On 28 January Italy's data protection authority asked DeepSeek what personal data it collected, where it was stored and on what legal basis. On 29 January the Irish Data Protection Commission wrote to the company requesting information on the data processing conducted in relation to data subjects in Ireland; because DeepSeek has no European establishment here, the Commission is not its lead supervisory authority in the way it is for the American platforms.[18][19] On 30 January the Italian regulator, having found the reply insufficient, imposed an immediate limitation on the processing of Italian users' data and opened an investigation; DeepSeek's position was that it did not operate in Italy and that European law did not apply to it. Belgium, Portugal, Spain and France's CNIL opened their own inquiries in the same week.[20] Government bans followed elsewhere: the US Navy on 24 January, NASA on 31 January.[21]
The same week, OpenAI told reporters it had evidence that DeepSeek had used its models' outputs to train its own, which its terms prohibit, and the White House AI adviser David Sacks said publicly that there was substantial evidence of it without producing any.[22][23] The allegation was never adjudicated. It sits oddly beside the fact that every part of DeepSeek's method was published.
What it changed
- Open weights reached the reasoning frontier. Before R1 the best open models trailed the best closed ones by something like a year. R1 put an MIT-licensed model level with o1 on most of the benchmarks that mattered, and its distilled versions put usable reasoning on a single consumer GPU.[2][3]
- The price anchor moved. OpenAI released o3-mini on 31 January at $1.10 per million input tokens and $4.40 per million output tokens, a fraction of o1's price, and every lab's pricing conversation changed with it.[24]
- The recipe was public. Reinforcement learning against checkable answers, using the method DeepSeek had published a year earlier, became the default way to build a reasoning model, and academic groups reproduced it within weeks.[3][5]
- Where the data goes became a board question. The regulatory letters were about the consumer app and the API hosted in China, not about the weights. The distinction matters, and most of the coverage missed it.[18][20]
What it did not change
The closed labs were still ahead, and stayed ahead: o3-mini arrived within eleven days, and o1 still led on graduate-level science. The full R1 needs a rack of data-centre GPUs to serve at a useful speed, and the distilled versions that run on a laptop are noticeably weaker than the headline numbers suggest. A reasoning model is still a language model: it produces confident wrong answers, it is not a source of record, and it is no safer to paste client data into than any other service, wherever the servers are. And for the vast majority of Irish businesses the question in January 2025 was not which frontier model to bet on. It was which of their own processes were worth putting any model on at all, and that question was unchanged by the week's events.