AI Infrastructure Capital Fans Out From Chips to Custom Silicon, Power and Autonomy

Fractile's step-up to a $6.5bn valuation on the back of an Anthropic chip agreement is the clearest datapoint yet that frontier model developers are willing to underwrite bespoke inference silicon to escape margin capture by incumbent accelerator suppliers. That same logic explains SpaceX's reported talks to acquire AI coding startup Cognition — vertical integration of the software layer where model capability converts into engineering throughput. The constraint binding all of it is electricity, which is why Bloom Energy's 40% reduction in fuel-cell installation time matters disproportionately: time-to-power, not cost-per-watt, is now the gating factor for datacentre buildouts, and Bloom is compressing exactly that variable. On the mobility edge, Uber's launch of self-driving rides in Zagreb with Pony.ai and Verne shows autonomy commercialising in second-tier European markets where regulatory friction is lower. Institutional positioning confirms the theme: Coatue's Philippe Laffont aggressively added semiconductor exposure in the second quarter alongside private AI and space bets. The counterweight is valuation and structure — analysts are explicitly warning that leveraged semiconductor products such as SOXL no longer offer expected returns commensurate with their volatility drag, a signal that the easy beta phase of this trade is over and dispersion within the complex will widen from here.