The artificial intelligence boom is not playing out the way most predicted. Rather than enriching software companies that build large language models, the real money is flowing to infrastructure providers—the companies selling the compute capacity itself. SpaceX's latest earnings report crystallized this inversion: the aerospace company generated $2.6 billion in AI-related revenue in its most recent period, more than triple the prior year, primarily by leasing GPU and data center capacity to AI companies. The figure now exceeds SpaceX's revenue from traditional satellite and launch services, positioning the company less as a space enterprise and more as a compute provider that happens to own rockets. This shift reflects a fundamental market reality: training and deploying large language models requires staggering amounts of computing infrastructure, and companies are willing to pay premium prices for reliable access.
AMD's financial results underscore the same dynamic. The chipmaker's data center revenue reached $6.7 billion in its latest quarter, up 107 percent year-over-year, with nearly all growth driven by AI-related demand. The company has essentially pivoted from a traditional semiconductor manufacturer to an AI infrastructure specialist, with gaming and consumer products relegated to secondary importance. This concentration of profit in the infrastructure layer matters because it reveals where actual scarcity exists in the AI economy. GPUs, high-bandwidth memory, and data center capacity are finite resources controlled by a narrow set of suppliers. Meanwhile, model developers—from OpenAI to Anthropic—operate in a more commoditized layer where competitive pressures are intense and margins thin. Even as software companies raise billion-dollar funding rounds, hardware providers are capturing outsized returns, suggesting that investors' focus on model capabilities may be misaligned with where value actually accumulates.
However, this infrastructure concentration creates emerging governance tensions. The Trump administration's recent AI testing framework explicitly excludes open-source models from security assessments, focusing instead on proprietary systems accessed through compute providers. This regulatory approach inadvertently strengthens the bottleneck: by funneling oversight through infrastructure providers rather than the developers themselves, policymakers are cementing compute as the choke point for AI oversight. If AMD, SpaceX, and a handful of GPU suppliers become de facto gatekeepers for which AI systems reach scale, regulatory pressure will naturally concentrate there. For investors and technologists tracking AI's evolution, the lesson is stark: the next decade's AI giants may not be model developers at all, but the companies that own the infrastructure on which those models depend.