1. the 80% margin from 2025 was theoretical,
2. they're relying on distillation/synthetic data for training,
3. and have been very opaque about cross-subsidization of R&D with their models.
The distillation alone adds a big asterisk for comparisons.
> But the numbers are available for companies like DeepSeek
You'd rather trust self-reported figures? LMAO
1. the 80% margin from 2025 was theoretical,
2. they're relying on distillation/synthetic data for training,
3. and have been very opaque about cross-subsidization of R&D with their models.
The distillation alone adds a big asterisk for comparisons.