Bernstein rates SpaceX Outperform with a $239 PT but says the valuation hinges
less on rockets than on whether AI compute can move to orbit. Its model assumes
about 3,600 launches a year by 2031, roughly 170,000 satellites and about 20 GW
of orbital compute. The thesis: terrestrial data centres face rising constraints
on power, land and cooling; if Starship achieves rapid two‑stage reusability and
sharply cuts launch costs, orbit could become a compute venue. Bernstein flags
new bottlenecks — chip supply, satellite thermal management, radiation
hardening, communications and regulation — and notes the scale it models would
demand about 2.8 mln wafers a year, or the equivalent of roughly five dedicated
fabs. Conclusion: ground AI data centres remain the primary revenue stream this
decade; orbital compute is a long‑dated option.
