Daily Alpha · Substack
· Post-Market Alpha · by Buzzberg Research
Substack supplied one actionable positioning change and one architecture-level inference thesis; the remaining public previews did not disclose enough detail for investment use.
Themes on this desk
AI infrastructure re-entry
Fidenza Macro re-established long exposure to AI infrastructure and semiconductors after exiting in June.
MoE serving economics
SemiAnalysis argued that Mixture of Experts changes active tensors, co-location needs and data movement, not merely parameter count.
Fidenza Macro re-enters long AI infrastructure and semis after June exit
Fidenza Macro states it went long AI infrastructure and semiconductor names again last Friday, having exited in June and sidestepped the intervening drawdown; the decision was posted in the subscriber chat.
A macro-focused letter re-establishing long AI infrastructure/semis after a profitable flat period is a discrete positioning shift that could signal dip-buying demand returning to the AI complex.
Watch Whether the letter discloses specific names, sizing, or stop levels in the full post, and whether AI infrastructure/semiconductor indices outperform from the stated entry date.
Source →SemiAnalysis: MoE has changed inference serving economics, not just parameter counts
SemiAnalysis argues Mixture of Experts, now widely used in frontier models, changed both the structure of serving and the economics of useful inference, doing more than increasing parameter count.
If MoE structurally lowers the cost of useful inference, it shifts demand toward memory bandwidth and interconnect rather than raw FLOPs, with implications for accelerator and networking suppliers.
Watch Whether the full article quantifies cost-per-token or serving-efficiency gains and names which hardware architectures benefit or lose.
Source →