Did DeepSeek prove that frontier-competitive AI models can be built for a small fraction of assumed billion-dollar budgets?

Steeve Morin · Feb 24, 2025

DeepSeek's innovation came from China because constraint is the mother of innovation — being unable to buy more compute forced them into efficiency

If you can simply buy more compute you have no reason to care about efficiency; being pushed to efficiency makes you deliver efficiency

59:56 Constraint of export controls forced the efficiency breakthrough

Nabeel Hyatt · Apr 4, 2025

DeepSeek doesn't change much about how to think about the future of the model layer

He never believed $100B–$1T training runs were the only barrier to entry; if capital alone were going to win, Spark would never have invested in Anthropic against OpenAI's capital argument, and that reasoning still holds

50:31 Deepseek does not change the view that capital alone was never the only moat

Richard Socher · Apr 18, 2025

DeepSeek's breakout was driven by being the first open-source model to match or exceed closed-source models, which destroyed the narrative that you need billions of dollars to compete

Everyone said it should have been impossible; the prevailing narrative was don't even try without billions, and the timing meant it was briefly the outright best model

19:55 First open source model to match closed source disproved the billions needed narrative

Richard Socher · Apr 18, 2025

DeepSeek's stated ~$5.6m training cost is not believable; the real figure is likely around $100m, though still vastly cheaper than the billions claimed to be necessary

The quoted number is at best the final training run's electricity; GPU purchase isn't included, and reaching a final model requires hundreds or thousands of smaller ablation and hyperparameter runs that each cost real money

21:27 Stated 5 6m cost not credible real figure closer to 100m but still far cheaper than assumed

Jonathan Ross · Feb 17, 2025

DeepSeek did not show that less compute is needed; it was a simple algorithmic improvement that made generating training data easier.

They wrote the answer in a box so they knew what to look for instead of needing a human to check it — a simple algorithmic change that eased data generation for training

10:11 Deepseek breakthrough was easier training data generation not less compute

Andrew Feldman · Mar 24, 2025

DeepSeek's achievement was focused engineering rather than invention — boring from an invention standpoint but excellent engineering

They weren't confused about being model intellectuals or about breaking new ground; they were interested in simply being better, and they built a model that was plainly better at many things

30:36 Deepseek success was superior engineering execution not scientific invention

Andrew Feldman · Mar 24, 2025

Very few open source software projects have had the kind of immediate, industry-wide impact on the technical community that that model had

Most open source projects ramp gradually from 10,000 to 100,000 to a million users, whereas this landed as an immediate boom among really smart people in the industry

31:55 Deepseek had unusually immediate industry wide open source impact