Will AI Keep Pretaining Scaling? XAI Grok 3.5 Will Tell Us

Epoch AI’s estimates AI compute has risen by 4x per year over the last 15 years, while AI algorithmic improvements have divided the compute needed by about 3 each year. This means that over time, the effective compute is growing by about 12x per year, with about 40% of this gain coming from algorithmic improvements.

XAI increased their compute 36 times over 9 months from Grok 2 to Grok 3.5. XAI is increasing compute by another 40 times over the next 9 months. This would be a 1500 times increase over 18 months and whatever algorithmic improvement XAI made. If XAI lagged 3X on AI algorithmic efficiency then they would not be pull away with Grok 3 and Grok 3.5 but would become the leader. The next step up with others falling behind on AI data center construction speed would separate Grok 5 in Q1/Q2 2026 with the compute difference between GPT2 and GPT3. It would be 80% of the logarithmic compute difference between full models.

XAI Grok 3.5 used ten times the compute of the next largest training (OpenAI GPT 4.5). IF Pretraining is dead then XAI Grok 3.5 will not be the best model when it is released. If it is the best and Grok 4 with even more training is better in a few months and if Grok 5 with ten times more training than Grok 3.5 in early 2026 we will see if scaling is still hugely important. XAI will also increase the post training.