A 2026 review of 4,667 medical artificial intelligence studies found that 88.2% remained at the preclinical stage. Only 2.4% were randomised controlled trials, the kind of research needed to show whether an intervention improves outcomes in real patients rather than performing well in a carefully prepared dataset.
Nature Reviews Clinical Oncology identified more immediate and better-supported uses in oncology, including:
- finding potentially eligible patients for clinical trials
- checking trial eligibility criteria
- extracting information from medical records
- monitoring studies under human supervision
Artificial intelligence is also being applied to drug design, medical imaging, treatment matching and the search for new uses for existing medicines. Dana Pe’er, an American Association for Cancer Research forecaster, has said the technology is already advancing work in those areas, while stressing the need to protect patient privacy and address data ethics.
The practical promise is substantial. It is simply not yet equivalent to showing that a computer can eliminate cancer.
Better medical data may matter more than bigger computers
Professor Chris Bakal of the Institute of Cancer Research in London offered a more grounded version of the future Haas described. Bakal, who is also chief executive of Sentinal4D, said the central question is no longer whether researchers use artificial intelligence, but what information they use to train it.
His laboratory works with data generated directly from patient samples rather than material gathered from the public internet. The system also does not require a giant data centre.
“The future of medical AI will not belong to whoever builds the biggest computer,” Bakal said. “It will belong to whoever has the right measurements.”
Models trained on precise biological observations could predict how cancers respond to drugs and potentially reduce the time required to develop treatments. Yet researchers continue to face poorly defined clinical questions, inconsistent datasets and biological information distorted by multiple overlapping factors.
Stanford researcher Ethan Goh has similarly warned that strong scores on technical benchmarks do not automatically produce clinical benefits. A model can look excellent in testing and still struggle with unfamiliar hospitals, incomplete records or patients who differ from its training data. Medicine has a persistent habit of being more complicated than a demonstration video.
Arm’s new processor business raises the stakes
Haas’s optimism also sits inside a major commercial shift at Arm. The Cambridge-based company became central to global computing by licensing power-efficient processor designs used in hundreds of billions of phones, vehicles, smartwatches and other devices. It traditionally supplied the architecture while customers built the finished chips.
That boundary is now moving. On 24 March 2026, Arm launched its first production data-centre processor, the 136-core Arm AGI CPU. Meta served as lead partner and co-developer, giving the Facebook owner a processor designed for its expanding artificial intelligence infrastructure.
According to Arm’s regulatory filings, customer demand across its 2027 and 2028 financial years has exceeded $2 billion. Haas told the BBC that interest had been “off the charts,” although substantial revenue depends on production increasing as planned.
Selling complete processors puts Arm into markets occupied by some of the companies that license its technology. The strategy could produce billions in annual revenue, as Reuters reported, but it also creates a delicate relationship with customers that may now see their supplier as a competitor. Nothing sharpens a partnership quite like discovering both companies want the same sale.
Haas said Arm’s energy-efficient designs are already used in about half of artificial intelligence data centres worldwide. Earlier in the summer, the company’s rising Nasdaq share price briefly made it the most valuable UK-based company on record by market value.
Will humanoid robots really be widespread within five years?
Haas also predicted rapid growth in humanoid machines, saying artificial intelligence would enable robots to see, learn and take on new assignments. Within a decade, he expects them to play a large role in manufacturing, cleaning, security, construction and repair work.
A hotel robot initially programmed to make a bed, he suggested, could learn to arrange towels or empty bins. He expects widespread humanoid use within five years and argues that forecasts of mass job replacement are overstated because new opportunities will accompany changes to existing work.
Independent projections are less enthusiastic. Gartner expects fewer than 20 companies to reach production-scale humanoid deployments in manufacturing or supply chains by 2028. More than 13,000 humanoid robots shipped during 2025, according to reporting by the Associated Press, but researchers cited in that coverage said high prices continued to block broad adoption.
The machines are improving, but reliable operation in uncontrolled workplaces remains difficult. A factory floor, hotel room or construction site presents changing objects, people and safety risks that cannot always be resolved by adding another training dataset.
Chip shortages remain the immediate obstacle
Before artificial intelligence can transform medicine or provide an army of adaptable cleaners, companies need enough processors and data-centre capacity to run it. Haas said the industry remains “absolutely in a supply-constrained environment,” with shortages slowing the construction of computing infrastructure.
Developers are planning multi-gigawatt data centres in France and the United States, while some companies have proposed placing computing facilities in space. Haas was blunt about the order of operations: “We need more fabs before we can put a data centre in space.”
Advanced chip manufacturing remains concentrated around Taiwan Semiconductor Manufacturing Company, better known as TSMC. Despite British government discussions about strengthening the domestic semiconductor supply chain, Haas does not believe the United Kingdom needs to build leading-edge fabrication plants.
Such factories require enormous investment, specialised workers, extensive natural resources and a broad network of suppliers. The UK government’s 2026 AI Hardware Plan takes a more targeted approach, prioritising specialist manufacturing, compound semiconductors and photonics instead of attempting to reproduce Taiwan’s advanced foundry industry from scratch.
That position may disappoint officials who still view Arm as the foundation of a larger British chip-manufacturing sector. Haas’s answer is effectively that Britain should focus on areas where it has a credible advantage, rather than buy an extremely expensive factory for the national display cabinet.
Arm remains British, although its ownership is not
SoftBank acquired Arm in 2016, a deal later criticised by British politicians concerned about the loss of a strategic technology company. Arm returned to public markets through a partial flotation on New York’s Nasdaq in 2023 rather than listing in London, adding another entry to the UK’s long-running debate about keeping major technology businesses at home.
Haas said half of Arm’s employees remain in the United Kingdom and described the company as “by far and away the largest employer in Cambridge.” Its headquarters, research base and engineering workforce therefore remain deeply tied to Britain, even as its ownership and stock-market identity stretch elsewhere.
The company now stands to benefit from almost every ambitious artificial intelligence prediction, from data-centre expansion to autonomous robots and computational drug discovery. Whether those predictions arrive on Haas’s timetable is another matter. Arm can supply more computing power. Cancer researchers still need the right biological measurements, robust trials and proof that promising models help actual patients.
