The artificial-intelligence revolution has reached an uncomfortable limit.

The problem may not be intelligence.

It may not be data.

It may not even be algorithms.

It may simply be that there are not enough chips to run everything AI promises to do.

Arm CEO Rene Haas has warned that the world's chip supply and data-center infrastructure are struggling to keep pace with AI's ambitions, even as he predicts the technology could ultimately transform medicine and potentially help humanity cure diseases such as cancer.

His message is remarkably optimistic about what AI could accomplish—and remarkably pessimistic about how difficult it may be to provide the hardware needed to get there.

“I've always thought that the killer app for AI is health,” Haas said in an interview with the BBC. He argued that AI could dramatically shorten the time required to discover new drugs and accelerate testing.

That vision is enormous.

AI systems are already being used to analyze biological data, predict molecular structures, search chemical combinations and assist researchers with drug discovery.

But increasingly capable models require increasingly powerful computers.

And that is where the bottleneck appears.

Haas says the industry first needs to figure out how to make enough chips to power the next generation of AI.

The warning comes from one of the most important figures in the semiconductor industry.

Haas has led Arm since 2022 and previously spent seven years at Nvidia, where he served as vice president of its computing-products business. That background gives him a unique perspective on both processor architecture and the AI infrastructure explosion.

His concerns are not limited to the number of GPUs available.

Memory supply is also expected to remain tight for some time.

That matters because AI systems are extraordinarily memory-intensive.

Training and running advanced models requires moving huge amounts of information between processors and high-speed memory.

A shortage of memory can therefore limit the performance of the entire system even when enough compute capacity exists.

This is becoming one of the defining bottlenecks in the AI economy.

Demand for high-bandwidth memory has surged as hyperscale data centers expand.

Companies such as Samsung and other memory manufacturers are investing heavily to increase capacity.

But new semiconductor facilities take years to construct and qualify.

That creates a structural mismatch.

AI software can be developed relatively quickly.

Hardware production cannot.

The result is a world where demand can grow faster than manufacturing capacity.

That problem becomes even more complicated when governments and local communities resist new data centers.

Haas warned that U.S. data-center growth faces opposition because of environmental concerns and shortages of resources.

The issue is bigger than chips.

A giant AI data center needs electricity.

Lots of it.

It needs cooling.

It needs water in many locations.

It requires land and high-capacity network connections.

And it has to be integrated into an electricity grid capable of handling enormous new loads.

That means the physical infrastructure supporting AI is becoming just as important as the AI models themselves.

It is one of the biggest misconceptions surrounding the technology boom.

People often describe AI as software.

But the most powerful AI systems are industrial-scale machines.

They require factories, power plants, data centers and global supply chains.

The companies building the models therefore depend on an extraordinary number of infrastructure providers.

Nvidia supplies accelerators.

TSMC manufactures advanced processors.

Samsung and other memory makers supply crucial memory.

ASML provides the lithography equipment needed to manufacture leading-edge chips.

Arm provides processor architecture used across enormous parts of the computing ecosystem.

Each layer matters.

A shortage at any point can slow the entire chain.

That creates a strategic problem for the United States and other countries racing to develop AI leadership.

Governments want domestic AI capacity.

Companies want more data centers.

Investors want faster growth.

But the physical world has limits.

Electricity grids cannot be expanded instantly.

Chip factories cannot be built instantly.

Memory production cannot double overnight.

Water infrastructure cannot simply appear where demand is greatest.

The AI industry's enormous growth forecasts therefore depend on a huge infrastructure buildout occurring simultaneously.

And that buildout is becoming increasingly expensive.

The contradiction is fascinating.

AI could eventually make many systems dramatically more efficient.

It could accelerate scientific research.

It could improve energy usage.

It could make drug discovery faster.

It could help engineers develop better technologies.

But before society reaches those benefits, enormous amounts of resources must be devoted to building the computers that make AI possible.

Haas's healthcare prediction illustrates why the investment may still be worth it.

Drug discovery is notoriously slow.

Researchers can spend years testing compounds and eliminating candidates that fail.

If AI can help narrow the search space and predict successful compounds earlier, the economic value could be enormous.

The benefits would extend beyond pharmaceutical companies.

Faster drug development could reduce healthcare costs, improve treatment options and potentially accelerate breakthroughs for diseases that have resisted conventional research.

But those models still have to run somewhere.

A system capable of searching billions of molecular combinations requires substantial compute.

Running such workloads at scale requires data centers.

Data centers require chips.

Chips require semiconductor manufacturing capacity.

And semiconductor manufacturing capacity requires enormous capital expenditure.

This explains why investors are increasingly looking beyond the headline AI companies.

The biggest opportunities may exist not only among model developers but throughout the infrastructure chain.

That includes chipmakers, foundries, memory manufacturers, equipment providers, networking companies, data-center operators and energy suppliers.

Arm is part of that ecosystem because its architecture underpins a huge portion of modern computing.

The company has historically been strongest in smartphones and embedded systems, but it is increasingly positioning itself for growth in data centers and AI.

Its importance could increase if AI workloads continue spreading across different types of computers rather than relying entirely on specialized accelerators.

At the same time, Arm faces intense competition and investors have assigned the stock an enormous valuation premium.

Yahoo Finance recently highlighted Arm's efforts to expand into AI accelerators with Samsung, although analysts have questioned how much of that opportunity will immediately translate into data-center revenue.

That highlights a key point.

Being strategically important to AI does not automatically make a company a great stock.

The economics still matter.

Revenue must justify investment.

Margins must support valuations.

And demand must remain strong enough to generate returns on enormous capital spending.

Haas's comments nonetheless reinforce a much broader theme.

The AI race may eventually be constrained less by algorithms than by physical resources.

The companies that solve those constraints could become some of the most important businesses of the next decade.

That includes chip manufacturers.

Memory producers.

Semiconductor-equipment companies.

Data-center builders.

Power suppliers.

And perhaps companies developing technologies that make existing computing infrastructure dramatically more efficient.

For now, the AI boom continues at extraordinary speed.

Investors are celebrating larger models, bigger data centers and rising chip demand.

But behind the excitement is a hard physical reality.

There must be enough processors.

There must be enough memory.

There must be enough electricity.

There must be enough data-center capacity.

And there must be enough infrastructure to connect all of it.

Haas believes AI could ultimately help humanity tackle some of its most difficult medical problems.

That future may be closer than many people think.

But before AI can help cure cancer, someone has to build the machines powerful enough to run the models that might discover the cure.

The next AI bottleneck may therefore be surprisingly old-fashioned.

Not intelligence.

Not software.

Hardware.