For most of the past decade, conversations about artificial intelligence centered almost exclusively on demand: which industries would adopt the technology, which jobs might be displaced, and which consumer products would be transformed. A new analysis from CaixaBank Research shifts that lens, training attention on the producers, infrastructure builders, and capital allocators who actually make AI systems possible in the first place.
This supply-side framing has deep historical precedent. During the electrification wave of the late nineteenth century, public discourse fixated on what factories and households would do with electricity, while the less glamorous story of generator manufacturers, copper wire producers, and utility companies quietly determined the pace and shape of the entire transition. Something similar played out during the semiconductor boom of the 1980s, when chip fabricators and equipment makers held more structural power over the computing revolution than many of the software firms that captured the headlines.
AI is now at a comparable inflection point. The companies mining and curating training data, manufacturing the specialized GPUs that power large models, and building out the data center capacity to run inference at scale form a kind of invisible backbone. Their investment cycles, supply constraints, and pricing decisions will ultimately govern how quickly AI capabilities reach end users — and at what cost.
CaixaBank's researchers appear to be signaling that financial analysts and policymakers have been underweighting this layer of the stack. It is a reminder that transformative technologies are not conjured purely by visionary software teams; they are assembled, piece by physical piece, through supply chains that carry their own bottlenecks, geopolitical vulnerabilities, and economic logic.
Understanding the supply side of AI may be among the more important analytical tasks of this decade — not merely for investors seeking an edge, but for governments attempting to map where strategic dependencies actually lie.