When money freezes progress: AI and the high cost of technological lock-in

Innovation slows when the infrastructure built to accelerate it becomes too expensive to abandon. That pattern is visible in artificial intelligence today, but it has appeared many times before. Big technologies do not win only because they are better. They win because capital, regulation, manufacturing, and habit gather around them. Once that happens, a system can keep funding the familiar even after a better path starts to emerge.

That is the risk now around AI. In 2025 and 2026, the largest technology firms are expected to spend hundreds of billions of dollars on AI infrastructure. The spending is concentrated in one direction: larger models, larger clusters, more GPUs, more data centres, more power. This has already produced real gains. It has also created a heavily installed base. When a company has committed enormous sums to one architecture, it becomes harder to support alternatives that do not use the same equipment, the same supply chains, or the same assumptions. The question stops being only technical. It becomes financial.

More from Opinion

That matters because alternative approaches are not absent. Neuromorphic chips, state-space models, liquid neural networks, photonic systems, analogue computing, and other designs all point to a different future for machine intelligence. Some may fail. Some may prove useful only in narrow cases. But the market does not treat them as equal contenders. They compete against a system already anchored to the transformer era. Research money follows the dominant path. Procurement follows the dominant path. Talent follows the dominant path. The result is not always the best technology. It is often the technology that has already absorbed the most capital.

The same logic appears in semiconductors. AI depends on leading-edge chips, and leading-edge chips depend on a narrow set of suppliers and tools. ASML remains the only commercial supplier of extreme ultraviolet lithography machines, and its newest machines cost hundreds of millions of dollars each. This is not a normal market where switching is easy. It is a dense industrial chain built over decades. A company that falls behind cannot simply buy its way back in next quarter. Even chipmakers with strong balance sheets face practical limits. GlobalFoundries, for example, stepped away from the race to develop the next generation of chip manufacturing. Once a company exits that race, re-entry becomes extremely difficult. That is how lock-in works. Each rational decision makes later alternatives less rational.

This pattern is older than AI and older than semiconductors. In the history of electricity, Thomas Edison had a large investment in Direct Current (DC) infrastructure. Alternating Current (AC) eventually proved better for long-distance transmission, but Edison did not welcome the transition. He resisted it because the shift threatened the value of the system he had already built. The engineering case for AC was strong. The economic case for preserving DC was strong too, at least for the people already invested in it. The final outcome was not decided by engineering alone. It was decided by the cost of change.

Nuclear power offers another example. In the 1960s, the Molten Salt Reactor Experiment at Oak Ridge showed that thorium-based molten salt reactors could run at low pressure and produce less long-lived waste than conventional uranium reactors. The technical promise was real. The experiment ran for years and generated important evidence. Yet the programme was not taken forward as the leading civilian nuclear path. The reasons were not only scientific. They also involved institutions, budgets, and the gravity of the existing reactor industry. Alvin Weinberg, one of the key figures behind the work, was later pushed out as the US system moved away from his preferred direction. That is the central lesson. Superior ideas do not always fail on merit. They fail because they arrive late or because the surrounding system has already decided where its money belongs. Adding a final irony, China recently launched a $350 million molten salt reactor programme, building on research first developed at Oak Ridge and left behind by the United States.

Economist W. Brian Arthur gave this tendency a clear name: path dependence. Once a technology gains enough users, suppliers, compatible tools, and political support, switching becomes expensive even when a better alternative exists. The first winners do not remain dominant only because they are best. They remain dominant because the ecosystem around them is now built to keep them in place. This is not conspiracy. It is inertia with incentives behind it.

That helps explain why AI can continue to receive enormous funding even if the returns from raw scale begin to flatten. Early signals of diminishing returns are already visible in discussions among researchers and practitioners. More compute still helps. More data still helps. But the gains may not keep rising at the same pace. If that is true, then the next breakthroughs may come from efficiency, modularity, reasoning, memory, retrieval, or entirely different architectures. Yet capital often prefers the known curve. It funds the path it already understands. It does not easily fund the path that would make yesterday’s infrastructure look overbuilt.

That is the real danger. Once a technology becomes a financial regime, it starts defending itself. Investors want returns on sunk costs. Regulators prefer familiar systems. Builders keep extending what already exists. Even sceptics inside the system learn to speak its language. At that point, the best argument for a new direction is no longer that it is better in theory. It is that the old direction is beginning to cost too much to preserve.

History suggests that technological transitions do not happen because institutions become wiser on their own. They happen when the cost of staying becomes impossible to ignore or when a new approach becomes too effective to dismiss. The question for AI is whether that turning point will arrive before the infrastructure around today’s dominant models hardens into a generational lock-in. If it does not, the most powerful obstacle to progress may not be technical failure. It may be the success of the present system itself.

(Sreejith Sreedharan is a technology analyst and author of Future of Work – AI-Augmented Autonomous Decentralised. He works on organisational AI readiness and created the AI Instinct Index®, a psychometric diagnostic designed to assess behavioural readiness for AI adoption and adaptive capacity in constraint-heavy environments. Views expressed in the above piece are personal and solely those of the author. They do not necessarily reflect Firstpost’s views.)