DeepSeek and Chinese open source models keep undercutting OpenAI on cost

[Karen88]

The ongoing China versus US AI story keeps getting more interesting and the Independent piece digs into how DeepSeek and the broader Chinese AI ecosystem keep rattling OpenAI and the rest of the Western AI industry

What started with DeepSeek's R1 release caused something like a trillion dollars in market losses in the US because it beat major benchmarks while apparently being built for a tiny fraction of what OpenAI reportedly spent on GPT4, and that gap in cost efficiency is still the thing nobody in the West has a great answer for

Since then the pattern has basically repeated with newer Chinese models like Kimi K3, cheap or free open source releases that are good enough to make people question whether the massive spending by US labs is actually necessary to stay competitive

Theres also a real question of whether OpenAI has any legal recourse if it turns out Chinese labs trained on their outputs, and legal experts seem to think that path is murky at best since copyright law generally doesnt protect AI generated material the same way it protects human made content

What I find most interesting is how this keeps forcing a rethink of the entire premise that says you need hundreds of billions in compute spending to build a frontier model, China keeps showing that constraints can breed efficiency rather than just falling behind

Feels like this tension between the open efficient Chinese approach and the expensive proprietary Western approach is going to define a lot of how this decade in AI actually plays out

China is preparing its citizens for AI upheaval in a way the West can only dream ofwww.independent.co.uk

[Rustic Stuart]

The efficiency gap is still the wildest part of this whole story, if DeepSeek really did what they claim for a fraction of the cost then a lot of the Western spending logic just doesnt hold up

[Rhys74]

I remain a little skeptical of the exact cost figures Chinese labs report, doesnt mean the models arent good but the marketing incentive to claim you did it cheaper is pretty obvious

[Current]

OpenAI going after DeepSeek legally for distillation always felt like a weak position to me given how much OpenAI itself trained on scraped web data without permission

[BanterQueen]

This is basically the same story on a loop every few months, cheap Chinese model shocks markets, US labs say they cant compete on cost, then a new expensive US model comes out and the cycle resets

[Cheeky Shaun]

Open source Chinese models undercutting proprietary US ones on price is going to keep happening and honestly I think its good for consumers even if its bad for OpenAI's margins

[GlassKnight]

The national security framing around this stuff always feels a bit overblown to me, at the end of the day people are going to use whatever model is cheapest and good enough

[SingularityNodeOwl]

Constraints breeding efficiency is a great way to put it, American labs have basically had unlimited capital so they never had to get clever the way Chinese labs did under export restrictions

[Blake_73]

Copyright law not covering AI generated content the same way is going to be a mess for years, feels like the courts are way behind the actual technology here

[Highland Dylan]

If Chinese open weight models keep closing the gap for free the whole subscription business model for chatbots might be in real trouble long term