Assertion Supported AI assessment confidence: 95% certainty 4/5 debate potential 2/5

DeepSeek-R1 cost roughly $300,000 to train, 10x cheaper than DeepSeek-V3

Sebastian Raschka · State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka · Jan 29, 2026 · at 31:06

AI researcher Sebastian Raschka compares the training cost of DeepSeek's base model (V3) against its reasoning model (R1) to illustrate the cost-effectiveness of RL post-training.

0:00 / 0:24exact quote · 24.2s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“Deep seek version three, they had like a five million dollar price tag on that given the, I think two dollars per GPU, they assume whether that's a correct assumption or not it's a different question, but if you compare it relative to the cost of R one, I think R one was about 300,000 dollars when they trained it, they had a number in the nature version of the paper. So it's basically more than 10 times cheaper than pre-training.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

DeepSeek-R1 cost roughly $300,000 to train, making it more than 10 times cheaper than DeepSeek-V3's estimated $5 million pre-training cost.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

According to official technical reports published by DeepSeek, DeepSeek-V3's pre-training required 2.664 million GPU hours costing approximately $5.576 million (at an assumed $2.00/GPU hour rental rate). The DeepSeek-R1 paper reports that its post-training stage required 147,000 GPU hours, totaling $294,000 (roughly $300,000). Comparing $294,000 to $5.576 million shows R1's post-training cost was approximately 19 times cheaper than V3's pre-training cost, satisfying the claim that it was 'roughly $300,000' and 'more than 10 times cheaper than DeepSeek-V3's estimated $5 million pre-training cost'.

DeepSeek-V3 Technical Report (arXiv:2412.19437) reports pre-training costs of $5.576 million based on 2.664 million Nvidia H800 GPU hours at $2.00/hour.
DeepSeek-R1 Technical Paper (arXiv:2501.12948) reports post-training costs of 147,000 GPU hours evaluated at $294,000 (roughly $300,000).

Checked against

[1] arxiv.org [2] arxiv.org [3] huggingface.co [4] stanford.edu [5] muxup.com [6] deeplearning.ai [7] ycombinator.com [8] mlq.ai

DeepSeek R1 paper training cost 300000 DeepSeek V3 technical report pretraining cost 5.58 million DeepSeek R1 technical report post training cost "DeepSeek-R1" arXiv "294" OR "300" "GPU hours" "2501.12948" "DeepSeek-R1" "GPU hours" OR "cost" OR "294,000" OR "300,000"

assessed as of 2026-07-30

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

DeepSeek-R1 DeepSeek-V3

⚑ Spot an error? Report it

More from Sebastian Raschka

Opinion

Text diffusion models will not replace autoregressive Transformers at state-of-the-art

“So it is a interesting direction to go into these diffusion, diffusion models as alternative to the auto regressive transformers, but it is not I would say the replacement at the state of the art.”

Sebastian Raschka Jan 29, 2026 ▶ 12:57 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

Prediction Not checkable as stated

Future LLMs will prioritize architectural efficiency over larger model sizes

“I wouldn't expect bigger architectures. I would expect a more efficient architectures tweaks getting, The same modeling performance for less compute”

Sebastian Raschka Jan 29, 2026 ▶ 17:09 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

Insight

Pre-training is no longer where the low-hanging AI gains lie

“Pre-training is not dead, but pre-training is boring. So it's not where the low hanging fruit is anymore.”

Sebastian Raschka Jan 29, 2026 ▶ 17:32 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

Insight

RLVR unlocks pre-training knowledge rather than teaching LLMs new math

“The knowledge is already there in the pre-training, and this just unlocks it. It's just like a step that maybe shows the model how to use its own knowledge, basically.”

Sebastian Raschka Jan 29, 2026 ▶ 24:33 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

Prediction Not checkable as stated

Process Reward Models will eventually become standard in LLM post-training

“I think it is promising and we will see it working at some point. I think it's just like right now it's still Tricky to make it work, but I am quite sure we'll see it as part of the standard repertoire at some point.”

Sebastian Raschka Jan 29, 2026 ▶ 26:39 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

Insight

Bigger LLM gains will come from multi-model process refinement, not scaling

“That's where you make the bigger gains rather than scaling the model size. I think that's one of those things where you will see more progress coming from.”

Sebastian Raschka Jan 29, 2026 ▶ 28:04 State of LLMs 2026: RLVR, GRPO, Inference Scaling — Sebastian Raschka

DeepSeek-R1 cost roughly $300,000 to train, 10x cheaper than DeepSeek-V3

More from Sebastian Raschka

Turn any episode into a week of clips.