Erik Bernhardsson

I expect AI to produce major productivity gains and some very concrete scientific benefits. Computational biology is an especially compelling example: using large amounts of GPU compute to help discover medicines is a much more useful frame than treating AI as merely a chatbot or a cheaper way to write code. But capability is not the whole economic story. What teams can build depends on access to compute, infrastructure, vendors, and usable developer tools. GPU access is still awkward: suppliers often want long, fixed commitments while actual demand is uncertain and changes quickly. Reducing that infrastructure burden matters because most organizations do not want to become experts in scheduling accelerators and operating distributed systems just to deploy useful software. I also do not think cheaper code means every company will build everything internally. AI improves software vendors too, and shared products still benefit from accumulated knowledge, distribution, reliability, and product judgment. Knowledge with feedback loops measured in months or years remains valuable because it cannot be instantly recreated by generating more code. So I see a large upside, but its distribution will depend on the practical layers around the models. Flexible compute access, strong software ecosystems, open infrastructure, and genuine competition—including independent infrastructure companies—will shape whether AI becomes broadly useful or remains concentrated behind a few difficult platforms.

The biggest update would be evidence that AI cannot reliably improve work with long feedback loops—science, infrastructure, and complex product development—even when paired with strong tools and abundant compute. That would weaken my expectation of broad productivity and scientific gains. In the other direction, a major AI-enabled scientific result, such as a medicine successfully discovered through GPU-intensive computational biology, would make the upside much more concrete. I would also update if compute stopped being a meaningful bottleneck, or if cheap code genuinely caused companies to abandon shared software vendors at scale. Those outcomes would change not just the magnitude of AI’s impact, but the economic structure through which it arrives.