AI helps a drug pass phase 1. At phase 2 the advantage disappears

Rene Haas, chief executive of Arm Holdings, told the BBC that AI will help cure cancer within our lifetimes. The Guardian carried the claim on Monday. Arm designs chips and licenses those designs, with no laboratory and no clinical trial of its own.

What the pipeline holds

Regulators have approved no drug that an AI discovered or designed. The US Food and Drug Administration had granted none as of July 2026.

AI drug discovery has a measurable footprint. Of 6,147 drugs in clinical development, 67 involve a molecule an AI found. That is 1.1%. Twenty-four involve a biological target an AI picked, which is 0.4%.

A peer-reviewed count presented at the American Society of Clinical Oncology this year tracked 117 AI-enabled therapies from 63 companies into human trials. Sixty of them finished phase 1. Eight finished phase 2.

The two closest to a regulator are rentosertib from Insilico Medicine, which entered phase 3 on 7 July for idiopathic pulmonary fibrosis, and Takeda's zasocitinib for plaque psoriasis. Neither of them treats cancer.

“I believe in our lifetime, AI will help cure cancer.”

— Rene Haas, chief executive of Arm Holdings, The Guardian, 8 September 2026

Rene Haas, chief executive of Arm Holdings, speaking to the BBC. Source: The Guardian, 8 September 2026

Where the help stops

Boston Consulting examined the AI-discovered molecules that reached trials and published the result in Drug Discovery Today. In phase 1 they succeed 80% to 90% of the time against a historical norm near 50%. In phase 2 they land at about 40%, the same as every other molecule.

Phase 1 asks whether a compound is safe in people. Phase 2 asks whether it works. The entire gain sits in the first question.

Haas describes the fix as scale: more data into a bigger foundation model, and the problem yields. Look at what AI was pointed at and the pattern makes sense. It made molecule design fast, and molecule design was already the cheap part. Choosing which biological target to aim at stayed where it was, and target selection is where cancer research stalls.

What AI has delivered in a hospital

More than 4 million NHS patients received a faster lung diagnosis after funding went into AI-powered X-ray tools. Reading an image sooner is worth having, and the NHS presented it as faster diagnosis rather than treatment.

We published a piece this morning from a GP who uses an AI scribe. He reads the transcript three times and trusts none of the readings.

The man making the forecast

Arm licenses its designs to about 500 companies, Apple, Samsung, Qualcomm and Nvidia among them, and employs more than 7,000 people, 3,000 of them in Britain. The company is worth $269 billion. For a different view of what a chip designer can promise, see the physicist rebuilding chip energy from the ground up.

Earlier this year Arm proposed a pay scheme that could make Haas a billionaire if he turns it into a trillion-dollar company. That is a 3.7 times move from where the shares sit now. Haas also runs SoftBank's international business, and SoftBank holds a stake in OpenAI. In April he left the board of AstraZeneca, which develops cancer drugs. Arm does not.

His forecast may still land. What the record shows today is AI clearing the safety gate and stopping at the efficacy gate, with its two leading candidates aimed at scarred lungs and inflamed skin.

This article is for informational purposes only and does not constitute investment advice.