DeepSeek says its hit AI model cost just $294k to train

https://www.reuters.com/world/china/chinas-deepseek-says-its-hit-ai-model-cost-just-294000-train-2025-09-18/

19•jonbaer•4mo ago

Comments

onion2k•4mo ago

Maybe, if you don't include the >$10m investment in H800 hardware. Still a lot cheaper than competitors though.

48terry•4mo ago

Yes, if we include a cost they didn't include, the cost would be different.

beaner_count•4mo ago

More like, if you exclude costs, things cost whatever you want to tell people they cost.

jml7c5•4mo ago

No, their calculation is based on a rental price of $2/hour.

yorwba•4mo ago

Right, but they didn't use rented GPUs, so it's a purely notional figure. It's an appropriate value for comparison to other single training runs (e.g. it tells you that turning DeepSeek-V3 into DeepSeek-R1 cost much less than training DeepSeek-V3 from scratch) but not for the entire budget of a company training LLMs.

DeepSeek spent a large amount upfront to build a cluster that they can run lots of small experiments on over the course of several years. If you only focus on the successful ones, it looks like their costs are much lower than they were end-to-end.

yunohn•4mo ago

No, they’re saying training a model, specifically DeepSeek, costs X using N hrs of Y GPU rental.

yorwba•4mo ago

If by "they" you mean DeepSeek, they're not saying this, since you might not actually be able to rent a cluster of 512 H800s wired together with high-bandwidth interconnects at that GPU-hour price point. If you rent smaller groups of GPUs piecemeal in different locations and try to transfer weight updates between them over the internet, it'll kill your throughput.

PaulRobinson•4mo ago

They made this claim in a peer reviewed paper submitted to Nature, but it’s not clear how peers could evaluate the truth of this claim.

If it’s true, and the consensus is that we are hitting limits of how to improve these models, the hypothesis that the entire market is in a bubble over-indexed on GPU costs [0] starts to look more credible.

At the very least, OpenAI and Anthropic look ridiculously inefficient. Mind you, given the numbers on the Oracle deal don’t add up, this is all starting to sound insane already.

[0] https://www.wheresyoured.at/the-haters-gui/

fspeech•4mo ago

These numbers were easily supported by those who attempted to replicate the RL portion of their work. The foundational model training is harder to verify but is also not central to the paper.

Poland to probe possible links between Epstein and Russia

Effectiveness of AI detection tools in identifying AI-generated articles

Warsaw Circle

Reverse Engineering Raiders of the Lost Ark for the Atari 2600

The AI4Agile Practitioners Report 2026

Digital Independence Day

What a bot hacking attempt looks like: SQL injections galore

Show HN: FlashMesh – An encrypted file mesh across Google Drive and Dropbox

Show HN: AgentLens – Open-source observability and audit trail for AI agents

Show HN: ShipClaw – Deploy OpenClaw to the Cloud in One Click

Unlock the Power of Real-Time Google Trends Visit: Www.daily-Trending.org

Explanation of British Class System

Show HN: Jwtpeek – minimal, user-friendly JWT inspector in Go

Willow – Protocols for an uncertain future [video]

Feedback on a client-side, privacy-first PDF editor I built

Clay Christensen's Milkshake Marketing (2011)

Show HN: WeaveMind – AI Workflows with human-in-the-loop

Show HN: Seedream 5.0: free AI image generator that claims strong text rendering

A contributor trust management system based on explicit vouches

Show HN: Analyzing 9 years of HN side projects that reached $500/month

The Floating Dock for Developers

Arcan Explained – A browser for different webs

We are not scared of AI, we are scared of irrelevance

Quartz Crystals

Show HN: I built a free dictionary API to avoid API keys

Show HN: Kybera – Agentic Smart Wallet with AI Osint and Reputation Tracking

Show HN: brew changelog – find upstream changelogs for Homebrew packages

Any chess position with 8 pieces on board and one pair of pawns has been solved

LLMs as Language Compilers: Lessons from Fortran for the Future of Coding

Projecting high-dimensional tensor/matrix/vect GPT–>ML