Look across the biggest public-market and pre-IPO wealth-creation events of the past two weeks, and a clear pattern emerges: capital keeps rewarding the physical and financial infrastructure underneath the AI boom, not the application-layer products riding on top of it.
SK Hynix's $26.5 billion US listing is the clearest example. The company priced its offering explicitly to fund new high-bandwidth memory capacity for AI accelerators -- not a novel AI product, but a scarce physical input every AI lab and hyperscaler needs regardless of which model architecture ultimately wins. The market's response was unambiguous: more than seven times oversubscribed, a 14% opening pop, and a listing that immediately became the largest-ever US debut by a foreign company.
James Murdoch's SpaceX windfall tells a structurally similar story one layer removed. His roughly $120 million pre-IPO position, made in 2019 and 2020, is now estimated worth $6.6 to $7.4 billion following SpaceX's own record-setting IPO -- a return built on conviction in physical launch infrastructure years before public markets could price it, not a bet on any particular satellite-internet or space-tourism application layered on top of that infrastructure.
“SK Hynix's $26.5 billion US listing is the clearest example.”
The pattern holds even where skepticism about AI infrastructure spend is real and well-documented. VentureBeat's own survey found 86% of enterprises running their own GPUs report utilization at half capacity or less -- direct evidence that raw compute capacity is, in the near term, outrunning actual workload demand. And yet capital keeps flowing preferentially to the scarcest physical and financial infrastructure layers -- memory, launch capacity, power generation -- rather than pulling back from infrastructure investment broadly or rotating toward application-layer AI products.
The likely explanation is durability. Application-layer AI products face rapid commoditization risk as frontier labs ship new model tiers monthly and open-weight alternatives compress margins across nearly every category; physical infrastructure -- a fab, a launch pad, a power plant -- takes years to replicate regardless of how fast software iterates on top of it. Investors pricing IPOs and pre-IPO positions appear to be weighing that durability gap explicitly, rewarding scarcity and multi-year lead times over near-term growth rates in categories that could be disrupted by the next model release.
For founders and GPs allocating capital across the AI stack, the lesson from this cycle's biggest public-market winners is specific: infrastructure-layer businesses with genuine multi-year replication lead times are commanding premium reception from public investors right now, even amid real skepticism about near-term utilization, while application-layer AI companies face a much higher bar to prove durable differentiation. The bear case is that this is a temporary function of which infrastructure happens to be scarce this cycle -- memory and launch capacity today, potentially something else entirely once current buildouts mature -- rather than a permanent structural preference for infrastructure over applications. What to watch next: whether the next wave of 2026 IPO candidates outside AI-adjacent infrastructure -- in manufacturing, defense and energy, as Fortune's own H2 2026 outlook suggests -- receive anything close to SK Hynix or SpaceX-level reception, which would be the real test of whether this is an infrastructure-specific pattern or a broader durability premium across sectors.