Explainer: why Roche’s supercomputer is a big deal

Danish supercomputer Gefion, a public-private partnership between the Novo Nordisk Foundation and the Export and Investment Fund of Denmark (EIFO), was symbolically turned on in October 2024. The computer relies on 1,528 Nvidia GPUs. Keystone

Pharma companies are racing to acquire microchips and achieve the most powerful AI capacity in the industry. Experts say it’s encouraging for future drug development but bottlenecks remain.

Swiss drug manufacturer Roche announced last week that it would expand its collaboration with semiconductor giant Nvidia and become the pharmaceutical company with the largest graphics processing unit (GPU) footprint across the industry. GPUs are microchips made of semiconductors that can handle large calculations simultaneously, making their compute capacity key to complex AI models.

Roche’s deal comes on the back of a multi-year research collaboration struck in November 2023 between its American subsidiary Genentech and Nvidia. Both companies had vowed to accelerate new drug discoveries and delivery thanks to a next-generation AI platform.

The deal comes as pharma companies race to integrate AI into their workflows and drug development. Two other pharma giants, Novo Nordisk and Eli Lilly, made similar announcements in the past 18 months. These cross-industry mega-deals matter as they indicate which drug manufacturer will have the most powerful AI capacity in the coming years.

“Computer chips, and GPUs in particular, are the hottest commodity right now. Everybody wants to get their hands on Nvidia chips,” said Christian Hein, an independent consultant in AI in biopharma and healthcare, with previous work experience at Novartis and Amgen.

What’s in the deal?

The agreement will boost Roche’s semiconductor arsenal, which is set to reach over 3,500 cloud-based and physical GPUs – the largest capacity ever announced by a single pharmaceutical company.

Roche’s GPUs will be set up across Europe and the United States and serve what Nvidia calls an AI factoryExternal link, a data centre “specifically optimised for artificial intelligence workloads”. The AI factory will support the drug developer’s R&D, manufacturing, diagnostics, digital pathology and digital health units.

Roche acquired GPUs from Nvidia’s latest product range Blackwell, which the semiconductor company claims is the largest GPU ever built. Blackwell microchips have 2.5 times more transistors and are up to 30 times more efficient than those from the previous range, called Hopper, released in 2022. Roche expects its new microchips to be running in the second half of 2026, and believes its AI factory will be in full gear by early 2027. 

As is customary with commercial deals, financial details weren’t disclosed by either of the companies. Although Nvidia likely sells the circuits in bulk, when the model was first announced in March 2024 CEO Jensen Huang said Blackwell chips would costExternal link between $30,000 (CHF23,700) and $40,000 per unit. Roche didn’t disclose its AI budget, but a spokesperson said the manufacturer viewed “the AI Factory as a critical, long-term strategic investment in Roche’s future”.  

Is Roche’s deal better than its competitors’? 

While silicon, the most common material for the basis of chips, might not be rare, four minerals central to its conductivity are dependent on China and Russia. On top of geopolitical risks, the race for AI supremacy is also putting intense pressure on demand and turning GPUs into rare resources. The semiconductor market was valued at $775 billion in 2024 and could reach $1.6 trillion by 2030, according to recent McKinsey estimatesExternal link. 

In May 2024, clinical biotech firm Recursion said it completed the largest supercomputer in pharmaceutical history, with 504 Hopper GPUs. But it was beaten to the title less than six months later by a Danish supercomputer from Novo Nordisk Foundation and the Export and Investment Fund of Denmark, which used 1,528 of the same chips. A year later, in October 2025, Eli Lilly announced “the world’s largest, most powerful AI factory”, with 1,016 Blackwell Ultras. By the time Roche’s new AI factory is operational, Nvidia will have already launched a new range of processorsExternal link.

>>Is pharma living up to its AI promise? Listen to our podcast episode:

“There is not a single CEO who is not under pressure from his board to say, ‘I’ve got my AI strategy, I’ve got it figured out’,” Hein said. What isn’t clear from these announcements, including Roche’s, is how the chip footprint will translate into use cases. “These big partnerships sound very sexy and I’m not saying companies can’t pull it off, but making it work in real life is a much bigger challenge,” the expert added. 

“The new GPUs will power solutions across the entire value chain in both pharmaceuticals and diagnostics,” a Roche spokesperson said about the company’s two main divisions. Roche has said its AI factory has already powered digital transformation across the organisation, including helping scientists test hypotheses at scale and ramping up production of new drugs twice as fast, thanks to virtual replicas of production lines known as digital twins. 

Future challenges

But AI in pharma still hasn’t reached its full potential, analysts say, and challenges remain, including fully integrating the technology in all processes and being able to attract talent. “Drug development processes are very complicated and fragmented,” Hein said.

Not only is biological data harder to digitalise and navigate because it is “messy, unstructured and fragmented” compared to text, but manufacturing medicine requires experiments in specialised laboratories, including clinical trials, to ensure safety for patients.

Swiss bank UBS reported in March that AI in the pharma industry is currently most impactful in operations and accelerates regulatory preparation and financial processes. “AI can help identify targets in early drug discovery, but […] it is ultimately still limited in its ability to reliably design drugs that achieve a predicted clinical outcome,” the bank’s analysis said.

>> Find out how one tech company has been venturing into drug development:

According to UBS’ findings, Roche dropped a drug for idiopathic pulmonary fibrosis, an incurable lung disease that impacts breathing, because the “AI-identified target” performed worse than placebo in phase two clinical trials.

“We’re going to see major benefits of AI across pretty much every single step of the workflow of pharma companies, but you’re only going to reap real benefits when you actually review your processes and job titles and review the way you work,” said Hein, predicting that changing an industry that hasn’t seen major changes in the last 20 years could take three to five years.

In the meantime, drug developers are also facing a global, cross-industry war for AI talent. Some tech companiesExternal link are offering up to $100 million in bonuses. Pfizer CEO Alfred Bourla reportedly earnedExternal link $27.6 million in 2025. Pharma’s only advantage in this talent war would be its mission-driven work, said Hein: “Pharma after all, is about saving lives and curing major diseases”.

Edited by Virginie Mangin/gw