The DeepSeek-R1 Shock — A Low-Cost Chinese Reasoning Model Rocks the Market

On 20 January 2025, the AI company DeepSeek, based in Hangzhou, China, released two open-weights reasoning models under the MIT license: DeepSeek-R1 and DeepSeek-R1-Zero, together with six smaller dense models (1.5B to 70B) distilled from R1 onto Qwen and Llama backbones. The accompanying paper went up on arXiv two days later.

A week after the release, on 27 January, NVIDIA's share price fell about 17% in a single day to close at $118.58, erasing roughly US$589 billion in market value. It was the largest single-day loss for any company in the history of the US stock market.

A Quant Fund's Side Project

DeepSeek (深度求索) was founded in July 2023. Liang Wenfeng, who co-founded the Chinese quant fund High-Flyer in 2015, runs both—DeepSeek began as the fund's side project.

In about eighteen months it released a reasoning model comparable to OpenAI's o1. Technically and economically, that shook the industry's assumptions.

Why It Shocked

1. Performance. The R1 paper reports the model as "comparable to OpenAI-o1-1217 on reasoning tasks", and the numbers bear that out in both directions: R1 was ahead on AIME 2024 (79.8 vs 79.2 pass@1) and MATH-500 (97.3 vs 96.4), and behind on Codeforces percentile (96.3 vs 96.6), GPQA Diamond (71.5 vs 75.7), and MMLU (90.8 vs 91.8). Not a rout in either direction—which was the point. Reasoning models were not the exclusive property of closed labs.

2. Open weights. The full weights were published under MIT. Local execution, derivative models, commercial use—all permitted. OpenAI's o1, by contrast, lived behind a closed API.

3. The cost line. R1 was built on DeepSeek-V3, released a month earlier. Table 1 of the V3 technical report puts the model's full training at 2.788 million H800 GPU-hours and, "assuming the rental price of H800 is $2 per GPU hour", at US$5.576 million—of which $5.328M was pre-training, completed in under two months on a cluster of 2,048 H800s. The paper is explicit about the boundary: those costs "include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data." It is the price of the final run, not of building the model, and certainly not of building the company. Even so, set against the reputed hundreds of millions for a frontier Western model, the number reframed what the frontier was assumed to cost.

4. Hardware under US export controls. Training used the NVIDIA H800—the part NVIDIA down-specified so it could still be sold into China under US export rules. The implication that a deliberately handicapped chip had produced a frontier-class model turned into a question about the regulation itself.

5. Reinforcement learning without demonstrations. The headline methodological result belongs to R1-Zero, not R1: trained by large-scale reinforcement learning (GRPO) directly on DeepSeek-V3-Base with no supervised fine-tuning step, it developed reasoning behaviour on its own. R1-Zero's output suffered from poor readability and language mixing, so R1 added cold-start data and a multi-stage pipeline—SFT, reasoning-oriented RL, rejection sampling, then RL again—around the same idea. Conflating the two is easy and wrong: R1 is not a pure-RL model.

27 January — The Market

Through the weekend, US markets absorbed the DeepSeek news—helped by DeepSeek's app overtaking ChatGPT as the most-downloaded free app on the US App Store. On Monday 27 January, per CNBC's tally:

• NVIDIA: −17%, closing at $118.58 (−US$589 billion in market cap; more than double its own previous record, a $279 billion drop in September 2024)
• Broadcom: −17% (−US$200 billion)
• Oracle: −14%; Dell, HPE, and Super Micro all −5.8% or worse
• Nasdaq Composite: −3.1%

What the market read: if AI training is far cheaper than assumed, future NVIDIA GPU demand contracts; if frontier-class models ship open-weights, OpenAI/Anthropic API margins compress; the US chip export-control regime may have strengthened rather than slowed China. Not everyone agreed—Cantor's analysts wrote the same day that advances in AI would lead to "the AI industry wanting more compute, not less", and told clients to buy.

Aftershocks

The shock outlived a single trading day.

Versus Stargate. The juxtaposition was almost too neat. R1 shipped on 20 January; the next day, 21 January, the Trump White House announced the Stargate Project—OpenAI, Oracle, SoftBank and MGX, $100 billion committed immediately and up to $500 billion over four years. Six days after that, the market decided a $5.576 million training run was the more important number. Whatever the comparison was worth arithmetically—and it compared a capital-expenditure programme with a single training job—it planted a durable doubt about scale-first strategy.

Export-control re-debate. The strengthened end-of-Biden controls on chip exports to China came under renewed scrutiny in Washington.

US firms' strategy shift. OpenAI shipped o3-mini on 31 January, four days after the rout; the full o3 followed in April. Anthropic and Google increased investment in reasoning models.

Open-source revival. Llama (Meta), Mistral, Qwen (Alibaba), and the broader open-weights camp gained momentum from R1's gravitational pull. Alibaba released a model it claimed surpassed DeepSeek-V3 on 29 January, two days after the rout.

The Frontier Stopped Being Exclusive

DeepSeek-R1 demonstrated a form of democratisation of AI development. The frontier was no longer the exclusive province of the US tech giants. A mid-sized Chinese startup, working on export-restricted hardware and publishing a training-run cost two orders of magnitude below the going assumption, could produce a model in the same class.

If ChatGPT's release (November 2022) was a vector of "AI to every household", DeepSeek-R1 was a vector of "AI development to mid-sized companies anywhere". The industry's terrain flattened by one more degree.

The claim also survived review. On 17 September 2025 the R1 work appeared in Nature (645, 633–638) as "DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning"—received in February, accepted in July—making it one of the first frontier-scale LLMs to pass peer review rather than arrive as a company preprint. DeepSeek kept shipping: V3.1 in August 2025, V3.2 in December, and the V4 family (V4-Pro, V4-Flash) from April 2026. A model badged R2 has still not appeared.

How that flattening will combine with regulation, national strategy, and market logic is still an unfolding story.