DeepSeek-R1 and the Reset of the AI Race, January 2025

---

The Setup: Reasoning Models and the OpenAI o1 Era

To understand why DeepSeek-R1 mattered so much, you have to understand what a “reasoning model” is, and why it was, in early 2025, the cutting edge of AI capability.

Large language models like GPT-4, Claude, and Gemini are, at their core, next-token predictors. Given a sequence of text, they predict what comes next. This simple mechanism, scaled up to enormous amounts of data and computation, produces systems that can write essays, answer questions, translate languages, and write code. But these systems have a limitation: they generate their answers in a single pass, without spending much time “thinking” about hard problems.

A reasoning model is different. When you ask a reasoning model a hard question — a mathematics problem, a coding challenge, a logical puzzle — it does not produce its answer immediately. Instead, it generates an explicit chain of reasoning first. It breaks the problem into steps, considers different approaches, checks its work, corrects its mistakes, and only then produces a final answer. This process takes more computation — the model is generating more text, and each token requires computation — but it produces dramatically better results on tasks that require careful reasoning.

The first major reasoning model was OpenAI’s o1, released in September 2024. o1 was a significant advance. On AIME 2024 (the American Invitational Mathematics Examination — the standard benchmark for evaluating reasoning models), it scored 79.2%, far higher than any previous model. On Codeforces (the competitive programming platform — the standard coding benchmark), it achieved an Elo rating (a numerical competitive-skill measure) that placed it in the top 1.5% of human competitors. o1 was widely understood to be a glimpse of the next generation of AI capability — systems that could reason, not just pattern-match.

But o1 was closed. It was available only through OpenAI’s API, at a substantial cost per query. OpenAI did not release the model weights, the technical details of how it was trained, or the architecture that made it work. The broader research community could use o1 through the API, but could not study it, modify it, or build on it directly.

This was the state of the art when, on December 26, 2024, DeepSeek released V3 — the base model on which R1 would be built. And it was the state of the art when, on January 20, 2025, DeepSeek released R1 itself.

---

January 20, 2025: The Release

DeepSeek-R1 was released on January 20, 2025. The release included the model weights, a technical paper published on arXiv, and six smaller “distilled” versions of the model, ranging from 1.5 billion to 70 billion parameters. Everything was released under the MIT license — one of the most permissive open-source licenses. Anyone could download the model, modify it, build on it, and use it commercially, for free.

The technical paper, published as arXiv:2501.12948, described what DeepSeek had built. arXiv:2501.12948 — the DeepSeek-R1 technical paper — described the pure-RL R1-Zero result and the multi-stage training pipeline. R1 was a reasoning model, built on top of the V3 base model. It had been trained using a multi-stage pipeline that included reinforcement learning — a technique where the model learns by trial and error, receiving rewards for producing correct reasoning and penalties for producing incorrect reasoning.

The most striking result in the paper was about a variant called R1-Zero.

R1-Zero had been trained using pure reinforcement learning, without first being trained on human-generated reasoning examples. This was a significant departure from the dominant approach, which involved collecting large amounts of human-generated reasoning data — examples of people solving math problems step by step, writing code, and so on. This data was expensive to collect and limited in supply. R1-Zero demonstrated that it was possible to train a reasoning model without this data, using only reinforcement learning.

The paper’s abstract stated: “We present DeepSeek-R1-Zero and DeepSeek-R1, which rely on large-scale RL to incentivize model reasoning behaviours. Our results demonstrate that the reasoning abilities of LLMs can be incentivized through pure RL without the need for SFT [supervised fine-tuning on human data].” Supervised fine-tuning (SFT) — training on human-generated reasoning examples — was the dominant approach that R1-Zero showed could be skipped.

This was a technical result with significant implications. If it was correct, it meant that reasoning models could be built more cheaply and more easily than the dominant approach assumed. It also meant that the dependence on human-labelled reasoning data — a bottleneck for the field — could be reduced or eliminated.

R1’s performance was remarkable. On AIME 2024, it scored 79.8%, slightly above OpenAI’s o1-1217 at 79.2%. On MATH-500 (a mathematics benchmark), it scored 97.3%. On MMLU (a broad knowledge benchmark), it scored 90.8%. On GPQA Diamond (a graduate-level science benchmark), it scored 71.5%. On Codeforces, it achieved an Elo rating of 2029, outperforming 96.3% of human competitors. These were frontier-level results — comparable to or exceeding the best closed-source models in the world.

The release of R1 on January 20 was, in the AI research community, immediately recognised as a major event. But it did not immediately cause a market panic. For the first few days, the model was discussed primarily in technical circles. The mainstream financial press had not yet picked up the story.

---

January 21, 2025: Stargate

The day after R1’s release, January 21, 2025, President Trump announced a $500 billion AI infrastructure project called Stargate, a joint venture between OpenAI, Oracle, and SoftBank. Stargate — the $500 billion AI infrastructure project announced by Trump on January 21, 2025 — was the American response to R1’s release. The announcement was made at the White House, with OpenAI’s Sam Altman, Oracle’s Larry Ellison, and SoftBank’s Masayoshi Son in attendance. Stargate was billed as the largest AI infrastructure project in history, with the goal of building massive data centres across the United States to train and run AI models.

The timing was striking. On January 20, a Chinese startup had released a frontier reasoning model for free, claiming it had been trained for about $5.6 million. On January 21, the American response was a $500 billion infrastructure project. The contrast was not lost on observers. It suggested two very different theories of how to win the AI race: the American theory, which held that winning required enormous amounts of capital and computing power, and the Chinese theory, which held that winning required clever engineering and efficient use of resources.

At the time, the American theory was dominant. The major American AI companies — OpenAI, Google, Anthropic, Meta — were all investing billions of dollars in compute, and the major American investors were betting that the demand for compute would continue to grow. NVIDIA, which made the chips that powered most of this compute, had become one of the most valuable companies in the world, with a market capitalisation of over $3 trillion. The Stargate announcement was consistent with this theory: more compute, more capital, more infrastructure.

But the contrast with DeepSeek was uncomfortable. If DeepSeek could build a frontier model for $5.6 million — even if the figure was misleading in important ways — then maybe the American theory was wrong. Maybe the path to more capable AI was not simply more compute, but better engineering. And if that was the case, then NVIDIA’s valuation, and the valuations of the American frontier labs, might be based on assumptions that were too optimistic.

This reassessment had not yet happened on January 21. But the ingredients were in place.

---

January 27, 2025: DeepSeek Monday

The reassessment happened on Monday, January 27. Over the weekend, the DeepSeek app — which provided access to R1 through a chat interface — had become the most-downloaded free app on Apple’s US App Store, displacing ChatGPT. The app’s popularity, combined with the technical achievement of R1 and the widely-cited $5.6 million training cost figure, triggered a reassessment of the assumptions that had been driving the AI investment boom.

The market reaction was severe and immediate. NVIDIA’s stock fell about 17% in a single day, erasing approximately $589 billion in market capitalisation. This was the largest single-day market-cap loss in the history of the US stock market. The Wall Street Journal reported the loss as $592.7 billion in paper losses; Bloomberg reported it as $589 billion. The variance reflects different snapshots of the trading day, but the broad picture is the same: this was an unprecedented loss.

The broader market also fell. The NASDAQ Composite dropped 3.1%. The S&P 500 dropped 1.5% (88.96 points, to 6,012.28). The Dow Jones Industrial Average, which is less weighted toward technology stocks, actually rose 0.7%. Other AI-related stocks also declined sharply. ASML, the Dutch company that makes the machines used to manufacture advanced chips, fell about 7%. TSMC, the Taiwanese company that manufactures most of the world’s advanced chips, fell about 3%. Broadcom and AMD, other major chip companies, also declined.

Bloomberg estimated that the total value wiped out across the AI sector on January 27 was roughly $1 trillion. The headline of Bloomberg’s market report captured the moment: “Nvidia’s $589 Billion DeepSeek Rout Is Largest in Market History.”

The market reaction was probably overblown. NVIDIA’s stock had risen 171% in 2024 and 239% in 2023, so even after the January 27 drop, the stock was still far above its level of two years earlier. And the $5.6 million figure, as we will see, was not the full cost of building DeepSeek’s capability. But the reaction revealed how fragile the assumptions underlying the AI investment boom had become. The market had been pricing in a future of ever-increasing compute demand. DeepSeek suggested that the future might be more complicated.

On the same day, DeepSeek reported a “large-scale malicious attack” on its services and temporarily limited new user registrations. The attack was likely a distributed denial-of-service (DDoS) attack, which is common for popular online services. Some commentators speculated that the timing was convenient cover for capacity constraints, but no evidence has emerged to support this speculation.

---

The Training Cost Question

The single most-debated aspect of DeepSeek-R1 was its training cost. The figure that drove the market panic was $5.6 million. Understanding where this figure came from, and what it does and does not mean, is essential to understanding what DeepSeek actually achieved.

The $5.6 million figure comes from the technical report for DeepSeek-V3, the base model on which R1 was built. The report states that V3 required 2.788 million H800 GPU-hours for its full training. The H800 — an NVIDIA chip exportable to China until October 2023, when US export controls were tightened — was the chip DeepSeek used to train V3 and R1. DeepSeek, like many Chinese AI companies, had stockpiled H800s while they were still legal.

If you take 2.788 million GPU-hours and multiply by a rough rental rate of $2 per GPU-hour, you get about $5.576 million. This is the figure that was widely cited in the press after the R1 release. It became the basis for the claim that DeepSeek had trained a frontier model for a fraction of what American companies were spending.

But the $5.6 million figure is misleading in important ways.

First, it refers to V3’s final training run — the actual computation that produced the model. It does not include the cost of the salaries of the researchers and engineers who designed the model. It does not include the cost of the data used to train the model. It does not include the cost of the failed experiments that did not produce a publishable model. And it does not include the cost of the hardware itself — the GPUs that DeepSeek used.

Second, the $5.6 million figure is for V3, not for R1. R1 is a reasoning model built on top of V3. In September 2025, Reuters reported that DeepSeek had said R1’s specific training cost was only $294,000, using 512 H800 chips. But this $294,000 figure refers only to the reinforcement-learning fine-tuning that turned V3 into R1 — not to the cost of training V3 itself, which R1 depends on. Epoch AI, an independent research organisation, estimated that the total cost of training R1 starting from V3 was around $1 million.

Third, the total cost of building DeepSeek’s capability — including the hardware, the salaries, the failed experiments, and the ongoing compute costs — was much higher than $5.6 million. CNBC reported in January 2025 that DeepSeek’s total hardware spend was about $500 million. The Register estimated in September 2025 that the purchase cost of the 256 GPU servers used to train the models was over $51 million. Other estimates have suggested the real total cost was over $1 billion.

So the honest way to describe the situation is this: DeepSeek’s final training run for V3 cost about $5.6 million in compute. The total cost of building DeepSeek’s capability was much higher — probably hundreds of millions of dollars. But even the higher figure is striking, because American frontier labs were believed to be spending billions of dollars on training their largest models. DeepSeek had achieved comparable results for significantly less.

The reason DeepSeek could do this is a subject of intense interest in the AI research community. The company made several technical innovations that reduced the cost of training. It used a more efficient attention mechanism, called Multi-head Latent Attention, that reduced the memory required during training.

It used a fine-grained mixture-of-experts (MoE) architecture — where different “experts” handle different inputs — that allowed V3 to scale to roughly 671 billion total parameters while activating only about 37 billion per token, keeping both training and inference compute far below what a dense model of the same size would demand. And it used a number of engineering optimisations that squeezed more performance out of each GPU-hour.

The lesson that the AI industry took from DeepSeek was that the cost of training frontier models might be falling faster than the American frontier labs had assumed. This was an uncomfortable lesson for companies like OpenAI, whose valuation depended on the assumption that frontier AI required enormous, expensive, hard-to-replicate compute. It was an even more uncomfortable lesson for NVIDIA, whose valuation depended on the assumption that the demand for its chips would continue to grow rapidly.

---

The Political and Industry Reaction

The political reaction to DeepSeek-R1 was swift. On January 27, 2025, at a House Republican retreat in Miami, President Trump was asked about DeepSeek. He said: “The release of DeepSeek, AI from a Chinese company should be a wakeup call for our industries that we need to be laser-focused on competing to win.” In the same comments, he described DeepSeek’s lower-cost approach as a “positive” development. The two statements — that DeepSeek was a “wake-up call” and that its approach was “positive” — captured the dual nature of the American response: alarm that China had caught up, and respect for the technical achievement.

Sam Altman, the CEO of OpenAI, responded on X on January 28. He wrote: “deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price. we will obviously deliver much better models and also… it’s legit invigorating to have a new competitor!” This was a gracious concession from the CEO of the company whose model R1 had matched. It was also a signal that OpenAI did not intend to change its strategy in response to DeepSeek.

Demis Hassabis, the CEO of Google DeepMind, offered a more measured assessment. He called DeepSeek’s model “probably the best work from China” — a significant concession from the head of one of the world’s leading AI labs. In a separate appearance, he reportedly said that the model was “not a significant technological” advance. These two statements may seem contradictory, but they likely reflect different aspects of the assessment: DeepSeek’s engineering was impressive, but the underlying techniques were not fundamentally new.

Behind the scenes, the reaction was more pointed. David Sacks, the White House AI advisor, raised the possibility that DeepSeek had used OpenAI’s models to train R1 — a practice known as distillation, which OpenAI’s terms of service prohibit. The Financial Times reported that OpenAI had told US lawmakers it had evidence that DeepSeek had distilled from its models. The claim was not proven. Independent researchers noted anecdotal evidence — R1 sometimes identified itself as an OpenAI model in its responses, which could suggest that it had been trained on OpenAI’s outputs — but no systematic proof of distillation was publicly released. DeepSeek did not publicly respond to the allegations in detail.

The distillation question is important because it goes to the heart of the open-weight debate. If DeepSeek trained R1 by distilling from OpenAI’s models, then DeepSeek’s achievement was, in a sense, parasitic on OpenAI’s work. It would mean that DeepSeek had not independently developed the capability that R1 demonstrated, but had rather copied it from OpenAI. This would weaken the claim that DeepSeek had fundamentally changed the AI race. On the other hand, if DeepSeek did not distill from OpenAI — or if distillation is a legitimate technique that all AI companies use — then DeepSeek’s achievement stands on its own merits.

The truth is not yet clear. The distillation allegations remain allegations, not proven facts. They are taken seriously by the American AI industry and by American policymakers, but they have not been independently verified.

---

The Geopolitical Backlash

The release of R1 also triggered a geopolitical backlash that continued through 2025. The concerns were about data privacy and national security. The DeepSeek app, which millions of people downloaded in late January 2025, collected user data and sent it to servers in China. This raised the same concerns that had been raised about TikTok and other Chinese apps: that the Chinese government could access the data, and that the app could be used for surveillance or influence.

The regulatory response was swift. On January 28, 2025, Italy became the first country to block DeepSeek nationwide. The Italian Data Protection Authority, the Garante, ordered the DeepSeek app blocked for Italian users, citing concerns about data protection and the lack of transparency about how user data was handled. Apple and Google pulled the app from Italian app stores within hours — the fastest nationwide restriction any Western government had placed on a Chinese AI product.

On February 5, 2025, South Korea’s industry ministry temporarily blocked employee access to DeepSeek on security grounds. South Korea’s Personal Information Protection Commission opened an investigation. On February 15, DeepSeek voluntarily removed its app from South Korean app stores pending a privacy review. Australia and Taiwan also banned the use of DeepSeek on government devices. Belgium considered a ban. The US Pentagon blocked employees from accessing DeepSeek.

In June 2025, a US official told Reuters that DeepSeek aided China’s military and had used Southeast Asian shell companies to evade US chip export controls. The US House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party issued a report accusing DeepSeek of espionage. The accusations were serious, but they were also contested. DeepSeek did not publicly respond to the accusations in detail, and the Chinese government denied that it used Chinese companies for espionage.

The geopolitical backlash created a complicated situation. DeepSeek’s models were open weights, which meant that they could be downloaded and run on anyone’s hardware, without sending data to China. Researchers and developers who were concerned about data privacy could download the model weights and run them locally. In fact, by April 2025, the DeepSeek models were available on Microsoft’s Azure AI Foundry and on GitHub, where they could be used without sending data to China. But the DeepSeek app, which was the easiest way for ordinary users to access R1, did send data to China, and was therefore subject to the same concerns as other Chinese apps.

This created a split in how DeepSeek was received. In the research community, R1 was widely adopted and built upon. Hugging Face, the leading platform for open-source AI models, hosted the R1 weights (downloaded millions of times), launched the “Open-R1” project to reproduce and extend DeepSeek’s work, and Western AI companies — including Meta and Microsoft — used R1 for distillation experiments. Within weeks of the release, the open-source community had produced a number of derivative models. Perplexity released “R1-1776,” a modified version of R1 designed to remove the political-content filtering that DeepSeek had built in. In the consumer market, however, DeepSeek faced growing restrictions, and the app was banned in an increasing number of countries.

---

The Export Controls Question

The DeepSeek-R1 release also raised serious questions about the effectiveness of US export controls on advanced chips. Since 2022, the United States had been restricting the sale of the most advanced AI chips to China, in an effort to slow China’s AI development. The controls had been tightened in October 2023, closing a loophole that had allowed the sale of the H800 chip.

DeepSeek had trained V3 and R1 using H800 chips that it had acquired before the October 2023 tightening. This was legal — the controls were prospective, not retrospective. But it raised the question of whether the controls had actually slowed China’s AI development. If DeepSeek could build a frontier model using chips that were two years old, then the controls might not be as effective as intended.

The think-tank community was divided. RAND published an analysis titled “DeepSeek’s Lesson: America Needs Smarter Export Controls,” arguing that the controls needed to be tightened and expanded. The Center for Strategic and International Studies (CSIS) published a “deep dive” on DeepSeek that examined the geopolitical implications. Epoch AI estimated that the controls had given the United States a lead of about four years, but that DeepSeek had closed much of that gap. Other analysts argued that the controls had actually accelerated China’s development of more efficient training techniques, by forcing Chinese companies to do more with less.

The truth is probably a combination. The controls did slow China’s access to the most advanced chips. But they also incentivised Chinese companies to develop more efficient techniques, and they did not prevent China from building frontier AI capability. The net effect is still being debated, and the DeepSeek-R1 release is the most prominent data point in that debate.

---

What DeepSeek-R1 Actually Changed

DeepSeek-R1 changed the AI race in several ways.

First, it demonstrated that frontier AI capability could be built outside the small group of American companies that had dominated the field. This was not a new demonstration — Chinese companies like Alibaba, Baidu, and ByteDance had been building capable models for years — but R1 was the first Chinese model that was widely acknowledged to match or exceed the best American models on reasoning tasks. The fact that it was released as open weights, and that the technical paper described the training methods in detail, made the achievement undeniable.

Second, it demonstrated that frontier AI capability could be built for less money than the American frontier labs were spending. The exact figure is contested, as we have seen, but the broad conclusion is clear: DeepSeek achieved comparable results with significantly less compute. This called into question the “scaling laws” thesis that had dominated the AI industry — the idea that the path to more capable AI was simply to train larger models on more data with more compute. DeepSeek suggested that there were other paths, and that the scaling laws might not be as rigid as had been assumed.

Third, it demonstrated the power of the open-weight approach. R1’s release under the MIT license allowed the global research community to study, modify, and build on the model. Within weeks, the open-source community had produced improved versions, distilled versions, and modified versions. This kind of rapid, distributed innovation would not have been possible if R1 had been a closed-source model available only through an API. The release also put pressure on the American frontier labs to justify their closed-source approach. If DeepSeek could release a frontier model as open weights, why couldn’t OpenAI or Google?

Fourth, it intensified the geopolitical competition over AI. DeepSeek’s success was a demonstration that the US export controls had not prevented China from building frontier AI capability. This has led to calls in the United States for tighter controls, and to debates about whether the controls are working as intended. It has also led to a backlash against DeepSeek in several countries, with restrictions on the use of the DeepSeek app and accusations of espionage.

Fifth, it raised the question of what “open source” means in the context of AI. DeepSeek’s R1, despite being released under the MIT license, contains built-in filters that censor content related to topics that are sensitive in Chinese politics. For some open-source advocates, this political filtering disqualifies R1 from being considered truly open source. For others, the filtering is a separate issue from the openness of the weights. This debate is not just about DeepSeek — it is about the broader question of what “open source” means for AI, and DeepSeek’s R1 is the most prominent test case so far.

---

The Aftermath

DeepSeek did not stop with R1. On May 28, 2025, the company released DeepSeek-R1-0528, a significant upgrade to the original R1, reportedly achieving “o3-level” performance on some benchmarks. On August 21, 2025, DeepSeek released V3.1, an upgrade to the V3 base model. On September 29, 2025, DeepSeek released V3.2-Exp, an experimental version. On November 27, 2025, DeepSeek released DeepSeekMath V2, a specialised mathematics model. On December 1, 2025, the V3.2 upgrade was applied to both the deepseek-chat and deepseek-reasoner API endpoints. As of late 2025, no “R2” had been released; the V3.x line and R1-0528 represent the practical follow-ups to the original R1.

The original R1 weights remain on Hugging Face as the canonical reference. They have been downloaded millions of times. They continue to be used by researchers, developers, and companies around the world. The release of R1 was, by any measure, one of the most consequential events in the history of the AI industry.

The market, meanwhile, recovered from the January 27 panic. NVIDIA’s stock, which had fallen 17% on “DeepSeek Monday,” rose over the following months as investors concluded that the demand for AI compute was not going to decline.

By the end of 2025, NVIDIA’s market capitalisation was higher than it had been before the DeepSeek panic. The market had decided that DeepSeek was, on balance, good for the AI industry — it increased interest in AI, it demonstrated the value of AI capability, and it did not fundamentally undermine the case for investing in AI infrastructure.

But the assumptions underlying the AI investment boom had been shaken. The scaling laws thesis was no longer as dominant as it had been. The closed-source approach of the American frontier labs was no longer as obviously correct as it had been. And the assumption that the United States had a decisive lead in AI was no longer as secure as it had been. DeepSeek-R1 did not end the AI race. But it reset it.

---

---

What would change if more people understood the story behind DeepSeek-R1? Who benefits from the current state of affairs, and who is left out? The conversation is worth having — with colleagues, with students, with anyone who uses technology without thinking about where it comes from.