The Economist has laid out a case that AI-driven computing could create a self-reinforcing capital accumulation loop that operates differently from any previous technology cycle, and the implications for who captures the surplus from that loop are the most important strategic question in startup formation right now.
The mechanism The Economist describes is worth stating plainly before evaluating it. Compute investment produces AI systems capable of accelerating research and development. That accelerated R&D produces productivity gains and new technologies faster than human-paced innovation could. A portion of those gains is reinvested into more compute, which produces more capable AI systems, which accelerates the cycle further. The loop is not guaranteed to be smooth or continuous, and it depends on a set of conditions that are not all currently satisfied, but the directional logic is grounded in dynamics that are already observable at smaller scale in frontier AI labs. OpenAI, Anthropic, and Google DeepMind are all using AI systems to assist in designing better AI systems. The question The Economist is raising is what happens when that process becomes the dominant driver of technological progress rather than a supplementary tool for human researchers.
For founders, the important reframe is that this is not primarily a labor displacement story, though it has labor consequences. It is a capital accumulation story. The entities that control large compute infrastructure, access to energy at scale, and the capital to reinvest productivity gains into the next generation of systems are structurally positioned to capture compounding returns in a way that looks more like ownership of a productive natural resource than participation in a competitive software market. Microsoft, Google, Amazon, and Meta have spent the past three years making capital commitments to AI infrastructure that dwarf anything in the history of the technology industry. Those commitments were made in anticipation of exactly the dynamic The Economist is describing: a world where compute is not just an input to software products but a self-amplifying productive asset.
The honest answer is that most startups cannot participate in the capital accumulation loop at its core, because the core requires access to chips, energy, and financing at scales that are not available to companies without hyperscaler balance sheets or sovereign-level capital backing. Training frontier models, running autonomous R&D pipelines at meaningful scale, and reinvesting productivity gains into the next generation of compute all require infrastructure that costs hundreds of millions to billions of dollars to build and operate. That is not a startup-accessible market in any conventional sense. The companies that are positioned to capture the compounding returns from the loop are the ones already inside it, and the barriers to entry are defined by physical infrastructure rather than intellectual property or software architecture.
What startups can realistically do is occupy the application and workflow layer that sits on top of the infrastructure, which is a genuinely valuable position if the underlying productivity gains materialize as expected. Every previous general-purpose technology cycle, electricity, computing, the internet, produced an infrastructure layer controlled by large capital allocators and an application layer where startups created enormous value by figuring out what the new capability was actually useful for. The AI cycle is following a similar pattern, with the important caveat that the application layer is itself being automated faster than in previous cycles. A startup that builds a useful AI-assisted workflow today may find that the same workflow is reproducible by a foundation model with better tool use in eighteen months, which compresses the defensibility window in ways that software businesses built on more stable abstractions did not face.
The founders who will navigate this environment best are those who think carefully about where durable value accrues in a world where AI can increasingly replicate software functionality. The answers cluster around a few categories: proprietary data that AI systems need but cannot generate themselves, distribution relationships that provide direct access to high-value customer contexts that API providers cannot reach without a partner, regulatory and compliance positioning in industries where deployment requires credentials and accountability structures that pure software companies cannot provide, and network effects that make a product more valuable as it accumulates users in ways that are not easily bootstrapped by a new entrant with better underlying models.
How the macroeconomic framing should change company-building strategy
The Economist's argument implies that the productivity gains from AI will not be evenly distributed, which is historically consistent with how general-purpose technology cycles have played out. Electrification created enormous aggregate wealth and also concentrated industrial advantage in ways that took decades of regulatory intervention to partially redistribute. The internet created enormous aggregate value and also concentrated search, social, and commerce markets in ways that remain the defining feature of the technology industry twenty-five years later. If AI creates a more powerful compounding loop than either of those predecessors, the concentration dynamics could be more pronounced and faster-moving than the patterns that founders and investors currently use as mental models.
For practical company-building, this argues for being explicit about the moat question earlier in a startup's life than has historically been necessary. The standard advice to find product-market fit before worrying about defensibility assumes a competitive timeline that may not hold when AI can compress the time from zero to credible competitor. A startup that finds traction in a valuable workflow needs to be asking what makes this difficult to replicate at the same time it is asking whether users want it, rather than sequencing those questions with a gap of several years between them.
The Economist's loop argument is worth reading not as a prediction of a specific outcome but as a framework for thinking about where the leverage in the next decade of company-building actually sits. Founders who understand the mechanism will make better decisions about what to build, where to build it, and how to position their companies relative to the infrastructure layer that is compounding beneath them. Those who treat AI as another software cycle will be solving for a competitive environment that may already be changing in more fundamental ways than the framing allows for.
Also read: A Waymo drove away with a passenger's luggage at San Jose airport and the incident exposes a product gap that safe navigation scores cannot fix • Apple Silicon's AI inference reputation is giving Apple pricing power it did not have to earn through software • University of Toronto students hit 50,000 tokens per second on FPGA hardware and the inference economics story is more important than the headline
