AI compute may split into two markets, and Morgan Stanley says policy is the only lever that matters

A global market for AI compute is being priced like one economy, but it could be administered as two. Morgan Stanley's note forecasts that demand will outstrip supply for years — and sees bilateral US and Chinese policy intervention as the lever that sorts the world's model ecosystems into separate jurisdictions, with separate capital stacks, separate customer bases, and separate compute footprints. Consensus gets the demand side right; it gets the lever wrong. The lever is policy, and it is bilateral.

The demand side has not yet arrived at scale. Median enterprise token spend sits below USD 11 per user per month, a humbling figure for a market that the consensus treats as supply-constrained. If the customer base is still rounding error, supply-side policy is the only variable that can bend the curve, and both Washington and Beijing know it.

The mechanism is jurisdictional sorting. US and Chinese AI policy intervention now acts as the actual product roadmap for who gets to train what, and where. The recent AI-infrastructure stock weakness is technical, not fundamental — which is why it is the wrong signal to fade the call. Bifurcation is the watchable variable, and the next two quarters of named policy decisions are the only data that matters.