Frontier-model release-interval saturation inflection?

By end of 2028, the gap between major frontier AI model releases shrinks to something like 2-3 months on average—not the 6-12 month intervals we saw in 2023-2024. OpenAI, Anthropic, and Google DeepMind are all shipping incremental capability updates faster than they can write blog posts about the last one. This happens because HBM memory production finally catches up with demand in mid-2027, labs stop fighting over chips, and the scaling laws that demanded massive training runs every 18 months give way to smaller, continuous refinement cycles. The race becomes about who can iterate fastest, not who can build the biggest model.

Through 2028, frontier labs keep releasing major models every 6-9 months, maybe even stretching longer—the same drumbeat we've heard since 2023. Chip supply stays tight even as TSMC ramps production; demand grows faster than supply. The complexity of training the next truly novel capability (reasoning depth, novel modality, architectural breakthrough) keeps requiring long, focused engineering cycles. Labs don't fragment into rapid-release cadences because the bottleneck isn't iteration speed—it's the physics and math of getting the next genuine leap right. Releases stay sparse, high-stakes moments.