NVIDIA's shares fell about 17% on 27 January 2025 because DeepSeek, a Chinese AI lab, showed that a frontier-class reasoning model could be built with far less GPU compute than investors had assumed. If top models need fewer chips, the reasoning went, demand for NVIDIA's most expensive accelerators might grow more slowly. This post was written on 28 January 2025, the day after that sell-off, and is kept as a dated commentary, with a short update at the end.
This article was originally published by me on LinkedIn.
Key takeaways
- DeepSeek-R1, released in January 2025, matched OpenAI's o1 on many reasoning benchmarks and was published with open weights.
- The widely quoted "under $6 million" figure is DeepSeek's own estimate for the final training run of its base model, DeepSeek-V3, not the full cost of building R1.
- The efficiency came from architecture and software choices: a mixture-of-experts design, low-precision FP8 training and careful use of export-restricted H800 GPUs.
- NVIDIA lost about 17% in one day, the largest single-day fall in market value recorded for a US company at the time.
- Cheaper training does not automatically mean fewer chips; lower cost per model can increase total demand.
The problem DeepSeek set out to solve
DeepSeek's paper (problem statement from their paper, published on arXiv as "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", January 2025) starts from a simple observation. Post-training, the stage after a model's large pre-training run, has become an important part of the pipeline. It improves accuracy on reasoning tasks, aligns models with social values and adapts them to user preferences, while needing relatively little compute compared with pre-training.
OpenAI's o1 models were the first to use inference-time scaling: letting the model reason through a longer chain of thought before answering, which brought large gains in mathematics, coding and science problems. How to scale reasoning at test time effectively was still an open question. Earlier work had tried process-based reward models, reinforcement learning and search methods such as Monte Carlo tree search and beam search, but, as the paper put it, none had reached general reasoning performance comparable to o1. DeepSeek's answer was to lean mainly on large-scale reinforcement learning.
Here's the scoop
Affordable AI development
DeepSeek showed that top-tier AI models do not have to cost a fortune. The headline claim, as reported at the time, was that its model was built for less than $6 million using around 10,000 NVIDIA chips. Compare that with OpenAI, which was reported to be spending over $5 billion a year. AI progress can be much more budget-friendly than many assumed.
Those numbers need care. The figure DeepSeek actually published was for DeepSeek-V3, the base model R1 was built on: about 2.788 million H800 GPU-hours for the final training run, which at an assumed rental price of $2 per GPU-hour gives about $5.6 million. DeepSeek said this excludes earlier research, experiments, data and staff. The run used a cluster of 2,048 H800 GPUs. The "10,000 chips" figure comes from press reports about GPUs held by DeepSeek's parent, the hedge fund High-Flyer, and was not confirmed by the paper. The OpenAI figure was a press estimate of spending and losses for 2024, not an audited number.
Worked example: does the $5.6 million add up?
| Item | Value | Note |
|---|---|---|
| GPU-hours (V3 final run) | 2.788 million | As published by DeepSeek |
| Assumed price | $2 per H800 GPU-hour | DeepSeek's own assumption |
| Compute cost | 2.788 M x $2, about $5.6 M | Final run only |
| Cluster size | 2,048 GPUs | As published |
| Wall-clock time | 2.788 M / 2,048, about 1,360 hours | Roughly 57 days of continuous running |
The sum is internally consistent. What it leaves out, the hardware purchase, failed runs, the R1 reinforcement learning stage and the team, is where most of the real cost of a frontier lab sits.
Open-source magic
DeepSeek-R1 was released with open weights under the MIT licence, which supports transparency and collaboration in AI research. Unlike the closed flagship models from OpenAI and Google, anyone could download and run it. Strictly, it is open-weight rather than fully open source: the training data and the complete training code were not published. Meta's Llama and Google's smaller Gemma models were also available as open weights, but R1 was the first open model to sit this close to the best closed reasoning models.
Efficiency over power
DeepSeek focused on efficiency, getting high performance without huge amounts of compute. That challenges the usual hardware-heavy strategy and could make companies rethink how they allocate resources in AI research. The main techniques, as described in the V3 technical report, were:
- Mixture of experts (MoE): the model has 671 billion parameters in total, but only about 37 billion are active for each token. As a rough illustration, compute per token scales with active parameters, so this needs around one eighteenth of the arithmetic of a dense 671-billion-parameter model.
- FP8 mixed-precision training: using 8-bit floating point for much of the training arithmetic cuts memory and bandwidth needs compared with 16-bit formats.
- Multi-head latent attention: compresses the attention key-value cache, which reduces memory use during inference.
- Communication and pipeline scheduling: custom scheduling overlapped computation with GPU-to-GPU communication, which mattered because the export-compliant H800 has lower chip-to-chip interconnect bandwidth than the H100.
Market shake-up
DeepSeek's success stirred up the AI market. NVIDIA's share price dropped about 17% on 27 January 2025 after the R1 release drew wide attention (the earlier version of this post said 16%; the closing fall was just under 17%), wiping out close to $600 billion of market value in a day. Analysts argued that DeepSeek's efficient methods could change the economics of AI, making it easier for new players to enter and pushing the largest technology companies to move faster.
Global AI scene
DeepSeek's rise showed how quickly China is advancing in AI, challenging the lead of American technology companies, and it did so on GPUs whose capability had been limited by US export rules. This could lead to a more competitive and diverse global AI industry.
Why efficiency may not mean fewer chips
The fear behind the sell-off was simple: if a model costs a tenth as much to train, buyers need a tenth as many GPUs. The counter-argument, made by many in the industry within days, is the Jevons effect: when a resource becomes cheaper to use, total use often rises. Cheaper training and inference let more companies build and deploy models, and reasoning models in particular spend much more compute at inference time, since they generate long chains of thought for every answer. Efficiency gains can shift spending from training to inference rather than remove it.
What reviewers and analysts probed
- Whether the cost figure compared like with like: final-run compute versus total programme cost.
- Which GPUs were really used, and whether export rules had been circumvented; these claims were reported and disputed, not settled in public.
- How much R1 benefited from learning from other models' outputs (distillation), which some US companies alleged.
- Whether open weights from a Chinese lab raised security or censorship concerns for enterprise users.
In a nutshell
DeepSeek's efficiency-driven approach shows that great things in AI can be achieved without breaking the bank. It could inspire more innovation and wider adoption of AI technology, and it put efficiency per unit of compute alongside raw scale as a measure of progress.
Update
NVIDIA's share price recovered in the months that followed and the company went on to new record valuations during 2025, as spending on AI data centres kept growing. The January 2025 fall is now usually read as a sharp repricing of how much compute frontier models need, rather than a turn in demand.
FAQ
Did DeepSeek really train R1 for $6 million?
No single published figure says that. About $5.6 million is DeepSeek's estimate for the final training run of its V3 base model at an assumed GPU rental price. The total cost of research, hardware and the R1 stage was not disclosed.
Which GPUs did DeepSeek use?
The V3 report describes a cluster of 2,048 NVIDIA H800 GPUs, a version of the H100 with reduced interconnect bandwidth made for the Chinese market under US export rules at the time.
Why did NVIDIA fall so far in one day?
Its valuation rested on expectations of very large, sustained GPU demand. A credible sign that top models might need far fewer GPUs hit those expectations directly.
Is DeepSeek-R1 open source?
Its weights were released under the MIT licence, so it can be downloaded, run and modified. The training data and full training pipeline were not released, so it is more precisely an open-weight model.
Your comments will be moderated before it appears here.