$TSMC - TSMC has exceptional customer service and partnership approach that sets them apart from competitors, making them the clear winner in semiconductor manufacturing. Their willingness to run experiments at their own expense and adapt processes based on customer feedback demonstrates a unique competitive advantage.
$INFRA - AI inference infrastructure will become one of the largest markets in the world, with inference eventually becoming a majority of global GDP. The demand for AI tokens far exceeds supply, creating a multi-decade supply shortage and massive opportunity for infrastructure providers.
$DATACENTER - Trillion-dollar data centers are inevitable due to economies of scale in token production, similar to how semiconductor fabs have scaled. The largest companies will be those that produce the most tokens and own the supply chain.
$POWER - Power availability and energy efficiency are critical constraints for AI infrastructure. Countries will need to dedicate majority of energy to data centers, making power infrastructure and efficiency improvements essential for scaling AI workloads.
$AIMODELS - AI models will evolve to use massive amounts of compute with mixture of experts architectures and dynamic computation. Future models will be able to attend to billions of tokens of context and run distributed across multiple racks, enabling superhuman capabilities.
Bearish:
$NVDA - NVIDIA's GPU architecture has fundamental thermal and latency limitations that prevent it from efficiently scaling for modern AI inference workloads. Their chips thermal throttle when trying to increase FLOPS utilization and have 4000ns chip-to-chip latency versus theoretical 2-3ns limits.
$TRAINING - The market incorrectly focused on training infrastructure while inference will become the dominant workload. Hardware designed before ChatGPT is fundamentally retrofit for modern models and not optimized for inference at scale.