AI growth is no longer constrained by software demand, but by simultaneous physical bottlenecks across the infrastructure stack.

Thesis: “AI growth is no longer constrained by software demand, but by simultaneous physical bottlenecks across the infrastructure stack.”

For years, the technology industry behaved as if software could scale independently from the physical world.

AI is changing that.

Today, AI growth is increasingly constrained not by demand, but by infrastructure scarcity. Not one scarcity. Multiple simultaneous bottlenecks appearing at the same time:

• electricity and power contracts,
• HBM memory,
• advanced chip fabrication,

That changes the operating model entirely.

Normally, when an industry faces a bottleneck, the response is relatively simple. You identify the current constraint and focus almost entirely on it. Optimising before the bottleneck only creates inventory waiting in line. Optimising after the bottleneck changes very little because input cannot increase anyway.

But AI currently has several bottlenecks at once.

That is what makes this moment unusual.

The industry does not only need more chips. It also needs more electricity, more memory, more fabrication capacity, more cooling, more datacentres, and more time. Unfortunately, time is the one thing infrastructure does not compress easily. New fabs take years. New power infrastructure takes years. New supply chains take years.

This is starting to create strange strategic behaviours.

Long-term electricity contracts are becoming strategic assets on their own. Even expensive power agreements may become valuable simply because they guarantee supply. In a constrained environment, certainty starts mattering more than optimisation.

The same logic is now appearing in semiconductor strategy.

TSMC remains the dominant advanced fabrication provider, but demand has become so concentrated that companies are increasingly forced to think differently about supply resilience. Apple reportedly exploring Intel manufacturing again is interesting not because Intel suddenly became technologically dominant again, but because predictable supply can matter more than absolute performance when scarcity appears.

That creates an uncomfortable possibility for the industry: companies may temporarily return to older or less efficient technology simply because it is available.

The same pattern is emerging with memory.

HBM has quietly become one of the most strategic resources in AI infrastructure. Without enough high-bandwidth memory, scaling advanced AI systems becomes extremely difficult. Yet memory production cannot suddenly expand overnight. Manufacturing capacity remains finite.

This creates another asymmetry.

Large hyperscalers can absorb scarcity more easily than smaller players. They can delay generation refreshes, negotiate earlier access to supply, and smooth shortages internally. Customers may simply continue consuming whatever instance generations remain available, often without visibility into the infrastructure constraints happening underneath.

The long tail does not have that luxury.

Smaller providers and smaller companies typically buy closer to market conditions. They feel shortages faster, pay higher marginal prices, and have less negotiating leverage when new supply appears.

Ironically, many of today’s hyperscalers originally built their success on serving the long tail. AI infrastructure scarcity may now reinforce the opposite dynamic: concentration of power toward the largest operators with the strongest purchasing capability and the deepest infrastructure control.

AI is still creating enormous opportunity.

But it is also quietly reintroducing something the software industry spent decades abstracting away:

physical limits.