Innovation often slows not because better ideas are unavailable, but because existing systems become too expensive and complex to replace. Once infrastructure, capital, regulation, and industry habits align around a dominant technology, switching to alternatives becomes increasingly difficult—even when better options begin to emerge. This pattern is already visible in artificial intelligence today and has repeated across multiple technological revolutions.
AI Infrastructure and Rising Lock-In
In 2025 and 2026, leading technology companies are expected to invest hundreds of billions of dollars into AI infrastructure. The focus is heavily concentrated on scaling existing approaches—larger models, more GPUs, bigger data centres, and expanded computing clusters. These investments have delivered significant improvements, but they have also created a deeply entrenched ecosystem.
As more capital flows into a single direction, it becomes harder for alternative AI approaches to compete. Research funding, procurement decisions, and talent pipelines increasingly align with the dominant transformer-based architecture. Even promising alternatives such as neuromorphic chips, state-space models, liquid neural networks, photonic computing, and analogue systems face structural disadvantages because they do not fit into the existing industrial and financial framework.
Semiconductor Dependence and Industrial Constraints
The same lock-in effect is visible in the semiconductor industry, which underpins AI development. Advanced chip manufacturing relies on a highly concentrated global supply chain, including specialised tools and equipment. For example, extreme ultraviolet lithography systems are supplied by a single dominant manufacturer, and these machines cost hundreds of millions of dollars.
This creates a system where technological progress is tightly bound to industrial capacity. Even well-funded companies face significant barriers to catching up once they fall behind. Some firms have already exited the most advanced segments of chip manufacturing because re-entry is economically and technically prohibitive. Over time, rational investment decisions reinforce the very structure that limits future flexibility.
Historical Precedents of Technological Lock-In
This pattern is not unique to AI or semiconductors. History shows similar dynamics in earlier technological transitions.
In the early electricity era, Thomas Edison’s direct current (DC) systems were eventually overtaken by alternating current (AC), which proved more efficient for long-distance transmission. Despite the technical superiority of AC, transition was slow and resisted because large investments had already been made in DC infrastructure. Economic incentives shaped adoption as much as engineering merit.
A similar dynamic appeared in nuclear energy. Experimental molten salt reactor research in the United States during the 1960s demonstrated promising safety and efficiency characteristics. However, the technology was not pursued at scale, partly due to institutional priorities and the dominance of existing reactor systems. Meanwhile, other countries have since revisited related research, highlighting how early abandonment can resurface later in different geopolitical contexts.
Economist W. Brian Arthur described this phenomenon as path dependence: once a technology accumulates enough supporting infrastructure, suppliers, users, and investment, it becomes self-reinforcing. Success itself creates resistance to change.
The Risk Facing Artificial Intelligence
AI may now be entering a similar phase. Early signs suggest that scaling existing models continues to deliver improvements, but the rate of progress may not remain linear. As returns from scale begin to stabilise, future breakthroughs may depend more on efficiency, architecture redesign, memory systems, and alternative computing paradigms.
However, financial systems tend to favour known trajectories. Capital is more likely to fund expansion of existing infrastructure than uncertain alternatives that require rebuilding ecosystems from the ground up. This creates a structural bias toward continuation rather than exploration.
Conclusion
The central risk is not that AI progress will stop, but that it may become concentrated in a narrow technological pathway simply because too much has already been invested in it. Once a system becomes economically and industrially embedded, it begins to protect itself through inertia, incentives, and dependency.
Historical transitions suggest that change only occurs when either the cost of maintaining the old system becomes too high or when a new approach becomes too effective to ignore. The critical question for AI is whether innovation will diversify before current infrastructure hardens into long-term lock-in—or whether progress will remain constrained by the success of the systems already in place.
