Cut Drug Development Costs by 30% with New Approaches

We were told AI would cure disease in years. That’s a fallacy.

Today, over 90% of clinical drug candidates still fail. It costs more than $2 billion and over a decade to bring a single drug to market. And now the industry faces a brutal paradox, which is the promise of exponential progress stuck inside a system governed by diminishing returns.

This is pharma’s version of Eroom’s Law, i.e. the inverse of Moore’s Law, where the return on drug development declines as science advances. This disconnect has created the “valley of death”: the chasm between discovery and clinical validation, where risk is high, capital is scarce, and most ideas go to die.

Pharma’s response? Avoid early risk. Wait for efficacy signals. Let small biotechs take the first leap. The result: a broken innovation model, a biotech echo chamber, and too few new answers for patients.

The Myths Holding Back Progress

Myth #1: Pharma knows how to develop drugs better.

Not true. Drug research and development remains probabilistic, messy, and context-specific. Big Pharma excels at scaling, not always at discovering. There’s no secret playbook locked inside corporate vaults.

Myth #2: AI can develop drugs better.

Also false, or, at best, wildly overstated. Today’s AI has helped with discovery, but not with development. It can suggest new molecules or predict binding, but it can’t yet navigate the complex, high-dimensional path to safe, effective therapies.

The current wave of hype assumes that AI can inform drug development just by reading scientific papers. But as AI pioneer Yann LeCun has argued: true intelligence doesn’t come from reading alone – it comes from interacting with the world and learning from experience.

Why Drug Development Needs Physical AI

Drug development is deeply physical. It spans chemistry, immunology, toxicology, pharmacokinetics, clinical design, and real-world outcomes. Optimizing this requires more than clever language models, it demands systems that can sense biology in motion.

Pharma has tried to bolt AI onto existing pipelines. But without coherent, standardized, longitudinal, multi-modal biological data, AI cannot reason. Instead of a world model for biology, we get piecemeal tools trained on static PDFs or siloed snapshots.

If a car breaks down, would you read the logbook or look under the hood?

Why Pharma Can’t Do This Alone

Even if large pharma wanted to build this, they would hit three walls:

1. Siloed data: Fragmented across departments, trials, and vendors.

2. No longitudinal integration: Incompatible formats across the preclinical-to-clinical arc.

3. Lack of engineering culture: Building physical AI requires sequencing platforms, Machine Learning pipelines, cloud infrastructure, and rapid iteration: things pharma doesn’t do natively.

Too often, valuable samples (e.g., blood, tissue) are collected but never turned into usable data. Or they’re analyzed and then locked inside PDFs, inaccessible to any AI.