USG blocking frontier models - and the impact on compute revenue being surprisingly positive? My thesis:

The US government has blocked the release of Anthropic Mythos, and is slow walking the release of openAI GPT5.6 to individual customers (enterprises probably) approved by the government one by one

This means Chinese open source open weight models are going to catch up to parity with legacy GPT5.5 and Opus 4.8 and even slightly surpass them in 3-6 months to global retail and enterprises outside of hand selected American companies close to the administration

This means we are going to increasingly see enterprise workflows happy to plug in open source models, simply because they are as good or even slightly better than the frontier lab models available to you as a business in Europe or Singapore (which isn't Mythos or GPT5.6 to be clear)

This means the continued commodification of tokens. It's difficult for OpenAI and Anthropoc to charge 75% margins for serving these non frontier tokens when the open sourced models are 5% the cost and at parity functionally.

So enterprises can incorporate a lot more AI workflows, as the cost of tokens ("the cost of intelligence") gets commodified and drops heavily

Where is the bottleneck to serving more and more of these cheap tokens now? Inference compute. Inference compute becomes the bottleneck, and with that, the usual infra buildout (memory, foundry, power, photonics etc)

\>"We heard this already bro. Infra is the bottleneck, picks and shovels yadda yadda"

Yes, but the new point that I am trying to make is that previously the frontier labs could capture a significant portion of the economic excess generated by incorporating AI into Enterprise workflows. That portion is now significantly cut. Instead, much more of the value capture shifts downstream, to the unavoidable layer - to the compute layer itself.

Tl;Dr - my hypothesis is that the US government blocking frontier models for global release -> open source models get a boost in adoption for global enterprises -> value capture by model labs dwindles -> more value is available to be captured by the unavoidable bottleneck, the inference compute layer.

So I would infer this development is strongly bullish $MU $SNDK $DRAM $INTC $TSMC etc in the medium term