Memories are overvalued

The market is not pricing AI on scarcity lasting forever. It is pricing the idea that AI demand will continue to grow faster than compute and memory supply. We think that assumption is wrong.

Supply is coming faster than expected, especially from China, custom silicon, more memory capacity and cheaper lower-tier GPUs. When supply catches up, token costs fall and current forward estimates can compress sharply.

A correction may take 1-2 years, let's see how things turn out.

China is no longer just trying to buy Nvidia chips. It is building its own AI stack. Huawei’s Ascend chips are not as good as Nvidia for the highest-end workloads, but they are becoming good enough for many domestic use cases. Chinese companies are already shifting some training and inference work onto local chips. For them, “good enough, cheaper and available” may matter more than having the best GPU.

Token costs should also keep falling. Better open-source models, smaller models, quantization, MoE architecture, lower-precision training, better inference software, and cheaper local hardware all push in the same direction. A lot of enterprise AI does not need the best Nvidia clusters. It needs cheap inference at scale.

HBM is extremely tight today, which is why SK Hynix, Samsung and Micron have been rewarded. But memory is still a cyclical industry. When everyone adds capacity, pricing eventually weakens. If AI demand growth slows, or if customers optimize memory usage, the market can move from shortage to oversupply quickly.

These are great companies just we don't see the trend supporting them.

Our thesis is simple: AI usage will keep growing, but the cost per token will fall dramatically. That is good for AI adoption, but bad for companies priced as if compute scarcity is permanent.

We would not be surprised if some AI hardware and memory valuations halve from today's levels. The demand is real, but the scarcity premium may not be.