Arm CEO Rene Haas believes artificial intelligence could help humanity cure cancer within his lifetime. But his prediction comes with a major qualification: the technology’s medical promise will depend on whether the semiconductor industry can supply the chips, memory, data centers and electricity needed to run powerful AI systems.
“I’ve always thought that the killer app for AI is health,” Haas said in an interview with the BBC released Tuesday. “AI is going to not only shorten the amount of time that those drugs can be invented, it’s going to shorten the amount of time that you test them.”
“I believe in our lifetime, AI will help cure cancer,” he added.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Haas’s comments capture the widening gap between AI’s ambitions and the physical infrastructure required to support them. The industry is increasingly presenting AI as a tool for solving some of humanity’s most difficult problems, from drug discovery and climate modeling to robotics and advanced manufacturing. Yet the systems needed to deliver those breakthroughs are placing unprecedented pressure on semiconductor supply chains.
Haas has led Arm, the British chip-design company whose technology is used in products ranging from smartphones to data-center processors, since February 2022. He joined the company in 2013 after spending seven years at Nvidia, where he was vice president of its computing products business.
His background gives him a direct view of the infrastructure demands created by the AI boom. Arm does not manufacture chips itself; instead, it licenses processor designs to companies that build chips for phones, servers, vehicles and other devices. As AI workloads spread across those markets, demand is rising for both high-performance data-center processors and more efficient chips capable of running AI applications at the edge.
“Right now, it’s quite constrained. We need more fabs before we can put a data center in space, I’ll tell you that much,” Haas said. “We’ll probably only put data centers in space when the biggest impediment to data centers is the cost of the data center.”
His reference to space-based data centers was partly humorous, but it underscored a serious point: the industry is still struggling to expand capacity on Earth. Semiconductor manufacturers are operating in what Haas described as an “absolutely supply-constrained environment,” and he expects shortages to persist.
AI’s Hardware Bottleneck Is Broader Than Processors
The shortage is not limited to the graphics processors and custom accelerators used to train and run AI models. One of the most important constraints is high-bandwidth memory, or HBM, which allows AI accelerators to move large volumes of data quickly.
Modern AI systems require enormous amounts of memory because their models contain billions or even trillions of parameters. As models become larger and more capable, the amount of memory needed to train and operate them increases. HBM has therefore become a critical component in the AI supply chain, and its production is concentrated among a small number of manufacturers.
But that concentration creates a vulnerability. Even if chipmakers can produce more AI accelerators, they may not be able to ship complete systems without sufficient supplies of advanced memory and packaging capacity. In many cases, the limiting factor is not the processor itself but the ability to combine the processor, memory, and networking components into a functioning AI system.
Micron Chief Operating Officer Manish Bhatia described the memory shortage as “really unprecedented” in a January interview. He said demand for HBM was consuming manufacturing capacity and contributing to shortages of memory used in more traditional products, including smartphones and personal computers.
Qualcomm CEO Cristiano Amon also warned during the company’s February earnings call that an “industry-wide memory shortage and price increases” could affect the size of the handset market during the year. The comments show that AI demand is beginning to influence markets far beyond data centers.
The pressure is also spreading to advanced packaging, the process used to connect processors and memory in high-performance systems. Packaging has become strategically important because AI accelerators cannot deliver their full performance without fast connections to memory. Expanding packaging capacity can be as difficult and time-consuming as increasing chip production itself.
Building Fabs Takes Years, While AI Demand Is Rising Now
The semiconductor industry is responding with a wave of investment. Chipmakers, memory manufacturers and governments are committing tens of billions of dollars to new fabrication plants, packaging facilities and research programs.
But supply cannot expand quickly. A new semiconductor fabrication plant can cost tens of billions of dollars and typically takes two or three years to build, followed by a lengthy process to install equipment, qualify production lines and reach high yields.
The development has created a mismatch between the speed of AI investment and the pace of industrial expansion. Technology companies can deploy new models and order additional computing capacity within months, while the factories needed to produce the required hardware may take years to complete.
The result could be periodic shortages, higher prices, and greater competition among customers. Large cloud providers and AI companies are likely to receive priority because of their purchasing power and long-term contracts, potentially leaving smaller businesses, universities and startups with less access to advanced computing.
The shortage could also encourage companies to develop smaller, more efficient models that require less computing power. Improvements in algorithms, model compression, specialized chips and software optimization may reduce the amount of hardware needed for some applications. But efficiency gains may not fully offset demand if AI use continues expanding into new industries.
Data Centers Face An Energy And Public-Acceptance Problem
Even when chips are available, AI companies need somewhere to operate them. Data centers require large amounts of electricity, cooling, land, and network infrastructure.
In the United States, more than 1,400 data centers had been built or approved for construction by the end of last year, as technology companies and infrastructure developers raced to accommodate generative AI and other compute-intensive applications.
The expansion is increasingly generating opposition from local communities concerned about electricity consumption, water use, noise, land development and environmental effects. A Gallup survey of 1,000 U.S. adults published in May found that seven in 10 respondents opposed building AI data centers in their local area, with nearly half saying they were strongly opposed.
That resistance could become a significant constraint on AI growth. Data-center developers may have access to capital and hardware but still face years of delays while securing permits, connecting to power grids, and negotiating with local governments.
Electricity availability is becoming a huge issue. Large AI facilities can require as much power as a small city, and clusters of data centers can place pressure on regional grids. Utilities may need to build new generation, transmission lines, and substations before facilities can begin operating at full capacity.
Water use is another concern, especially in regions where data centers rely on evaporative cooling. Developers are exploring alternative cooling systems, including closed-loop and liquid-cooling technologies, but those approaches can increase construction costs and complexity.
AI Could Accelerate Cancer Research, But It Cannot Eliminate Clinical Risk
For healthcare, the potential benefits of more computing power are substantial. AI systems are already being used in drug discovery, protein analysis, medical imaging, patient monitoring and the identification of potential therapeutic targets.
AI can search large biological databases, predict how proteins interact, identify patterns in medical images and help researchers prioritize compounds for laboratory testing. In principle, these tools could reduce the time and cost required to move from a biological hypothesis to a candidate treatment.
But Haas’s prediction should not be interpreted as evidence that a universal cancer cure is imminent. Cancer is not a single disease but a collection of more than 100 diseases, each involving different genetic mutations, biological mechanisms, and responses to treatment.
AI may improve the odds of finding effective therapies, but it cannot remove the fundamental uncertainty of human biology. A compound that performs well in a computer simulation or laboratory experiment may fail in animals or humans. A treatment that works for one genetic subtype of cancer may be ineffective for another.
Clinical trials remain essential. Researchers must establish that a treatment is safe, determine the appropriate dosage, measure its effectiveness, and understand its long-term risks. Regulatory review and manufacturing also take time, even when the underlying discovery process is accelerated.
AI could therefore have its greatest near-term impact not by producing a single cure for cancer, but by helping develop more precise treatments for specific cancer types. It may also improve early detection, identify patients most likely to respond to a therapy, and help doctors select combinations of existing treatments.