Scaling Scientific R&D with AI Supercomputing Infrastructure

​Pharmaceutical and life sciences enterprises have confirmed AI can enhance particular person phases of discovery, improvement, and manufacturing — however the {industry}’s legacy IT infrastructure was by no means built for frontier-scale compute. This is a structural, industry-wide constraint.

The economics make the stakes clear. According to the National Institutes of Health, a discovery can take virtually 15 years to turn into an accepted drug, with a failure price exceeding 95 p.c and pushing the efficient price per profitable drug above $1 billion.

A peer-reviewed research on NIH’s PubMed Central discovered that fewer than 10 p.c of medicine getting into scientific trials are in the end accepted, with attrition reaching 40–70 p.c in Phase 2 alone. That quantity of failed information is a latent asset — however exploiting it requires computational scale legacy techniques can not present.

Federal companies affirm the hole is nationwide, not proprietary. The National Science Foundation acknowledged in August 2026 that entry to AI infrastructure remains extremely uneven whilst AI transforms scientific discovery, and it committed $100 million to new regional AI infrastructure hubs to assist shut the hole. Meanwhile, the FDA reviews its drug-evaluation middle has already reviewed greater than 500 regulatory submissions containing AI parts since 2016, elevating the efficiency and reproducibility bar infrastructure should meet.

Where a purpose-built scientific compute has been deployed, returns are measurable. A 2025 NIH-hosted evaluate discovered that digital twins in steady manufacturing have improved energetic pharmaceutical ingredient consistency to 99.95 p.c, whereas AI-linked protein construction fashions show potential to chop goal validation time from months to days.

These figures describe an {industry} the place excessive attrition, unexploited historic information, rising regulatory scrutiny, and an acknowledged nationwide compute shortfall converge on one conclusion: pharmaceutical IT should evolve from a transactional assist operate into purpose-built scientific compute functionality.

Emerj’s Matthew DeMello in dialog with Thomas Fuchs, Chief AI Officer at Eli Lilly, discusses how AI has turn into a scientific infrastructure drawback in pharma; the {industry} ought to construct objective‑constructed compute to unlock significant discovery, mannequin accuracy, and manufacturing influence.

This article examines three core insights that matter most for pharmaceutical and life sciences leaders as AI turns into a central driver of scientific work:

• Unpublished failure information for higher-accuracy drug prediction: Capture and construction failed-experiment information into coaching pipelines, so prediction fashions study from the total end result house as an alternative of solely the fraction that succeeded.
• Purpose-built molecular fashions for actual drug-design functionality: Deploy molecular generative, physics-based, and diffusion fashions alongside LLMs, so organic complexity is modeled by architectures constructed for it somewhat than compelled by means of language instruments.
• Compute as core Scientific infrastructure, not IT overhead: Fund and govern supercomputing as a scientific instrument owned by R&D, so compute capability scales with what discovery requires as an alternative of what IT budgets enable.

Listen to the total episode beneath:

Episode:  Scaling Scientific R&D with AI Supercomputing Infrastructure — with Thomas Fuchs of Eli Lilly

​

Guest: Thomas Fuchs, Chief AI Officer at Eli Lilly

Expertise: Artificial Intelligence, Machine Learning, Computational Pathology, Medical Informatics

Brief Recognition: Thomas Fuchs is Senior Vice President and Chief AI Officer at Eli Lilly and Company, the place he leads AI initiatives throughout the group. Previously, he was the inaugural Chair of the Department of Artificial Intelligence and Human Health and Dean of Artificial Intelligence and Human Health on the Icahn School of Medicine at Mount Sinai, the place he additionally directed the Hasso Plattner Institute for Digital Health. He is the founder and former Chief Scientist of Paige, an AI-focused firm in computational pathology, and serves on the board of GeneDx. Fuchs can also be an Adjunct Professor at Mount Sinai and holds a Doctor of Sciences in Machine Learning from ETH Zürich, with postdoctoral analysis in Computer Vision at Caltech.

Unpublished Failure Data for Higher‑Accuracy Drug Prediction

Thomas Fuchs highlights that pharmaceutical R&D generates way more failed experiments than profitable ones, and that these failures include the strongest sign for enhancing molecular prediction.

He argues that the majority AI techniques within the {industry} are skilled on an unrealistically slender slice of outcomes as a result of solely profitable outcomes make it into the scientific literature. For leaders, the strategic shift is to raise failed‑experiment information right into a main coaching useful resource somewhat than treating it as discardable noise.

Fuchs explains why adverse outcomes are the strongest studying sign in discovery:

“If you solely practice an AI on the optimistic outcomes — the outcomes that get printed — you’re giving it a tiny and deceptive slice of actuality. For each molecule that labored, we had hundreds of thousands that failed, and people failures are precisely the place the actual studying sign is. When you faucet into many years of adverse outcomes, you may construct fashions that perceive what doesn’t work, and that’s what permits you to design molecules you wouldn’t have considered earlier than.”

– Thomas Fuchs, Chief AI Officer at Eli Lilly

For organizations, the operational steerage is:

• Integrate failed assays, non‑binding makes an attempt, and toxicity outcomes into mannequin coaching, so techniques study from the total end result house somewhat than the sliver that succeeded.
• Use adverse‑skilled fashions to prune non‑viable candidates earlier, lowering moist‑lab burden and decreasing late‑stage attrition.
• Leverage proprietary failure information as a aggressive benefit, since solely massive enterprises possess the depth required to outperform fashions skilled solely on public literature.

By elevating unpublished failures into a primary‑class coaching useful resource, pharmaceutical leaders can construct prediction techniques that mirror the true complexity of discovery, and materially enhance the reliability of AI‑pushed molecule design.

Purpose‑Built Molecular Models for Real Drug‑Design Capability

Thomas Fuchs makes a pointy distinction that many pharmaceutical leaders nonetheless blur: massive language fashions are helpful inside discovery workflows, however they aren’t, and can’t turn into drug‑design engines.

His level isn’t about hype administration; it’s about organic complexity. A single cell operates at a stage of bodily, chemical, and temporal element that language can not encode. When organizations attempt to power drug‑design duties by means of LLMs, they constrain themselves to the representational limits of textual content somewhat than the physics of biology.

Fuchs describes how Lilly approaches the issue otherwise. Language fashions orchestrate work, deal with documentation, and assist regulatory Q&A, however the precise design and prediction engines are constructed on architectures meant for molecular conduct: generative fashions, diffusion fashions, physics‑primarily based techniques, and nucleotide/RNA fashions. These instruments function in areas the place binding affinity, toxicity, conformational change, and response dynamics will be modeled straight — not translated into sentences.

He captures the limitation of language‑primarily based approaches in a approach that’s each technically trustworthy and strategically helpful for leaders:

“The complexity of a single cell goes far past what human language may even describe. You would constrain your self in the event you constrained your self to language‑primarily based fashions. The fashions that basically drive design are completely different — molecular fashions, diffusion fashions, generative circulate fashions — they’ll come up with new molecules.”

– Thomas Fuchs, Chief AI Officer at Eli Lilly

This distinction leads on to a set of sensible selections for executives:

• Use LLMs the place language is the substrate — orchestration, summarization, regulatory assist, documentation, and workflow automation.
• Use molecular and physics‑primarily based fashions the place biology is the substrate — prediction, generative design, optimization, and exploration of chemical house.
• Architect discovery pipelines so every mannequin class operates the place it’s strongest, somewhat than forcing a single mannequin sort to do the whole lot.

According to Thomas, this division of labor is just not a technical nuance; it’s a strategic requirement. Leaders who deal with LLMs as common engines will hit arduous ceilings in accuracy and scientific validity. Leaders who pair LLMs with objective‑constructed molecular fashions can broaden the frontier of what their discovery groups can really design.

Compute as Core Scientific Infrastructure, Not IT Overhead

Thomas Fuchs frames compute as a scientific instrument, a shift that basically modifications how pharmaceutical organizations ought to govern and spend money on AI. He factors out that Lilly’s supercomputing technique is just not about sooner experiments or generic efficiency positive factors; it’s about increasing the horizon of what scientists may even try.

With the brand new system, researchers can practice frontier‑scale fashions, run physics‑primarily based simulations at significant decision, and discover chemical and genetic areas that have been beforehand inaccessible.

What stands out in his clarification is how compute modifications the scope of scientific considering. In the previous, groups have been constrained to small fashions and restricted datasets as a result of infrastructure couldn’t assist something bigger. With a thousand B300 GPUs, that ceiling disappears. Fuchs describes the supercomputer as “a telescope” — a instrument that lets scientists see additional, ask deeper questions, and design fashions that have been unattainable underneath legacy constraints. This metaphor is just not rhetorical; it displays how compute straight shapes the ambition and creativity of R&D groups.

He additionally emphasizes that compute delivers worth far past discovery. In manufacturing, for instance, Lilly used AI to optimize the drying course of for APIs — a seemingly mundane step that resulted in hundreds of thousands of further doses reaching sufferers sooner. These sorts of positive factors don’t come from summary innovation; they arrive from treating compute as a core functionality owned by R&D, with clear metrics tied to affected person influence, operational effectivity, and scientific rigor.

Fuchs’s perspective provides leaders a special lens for evaluating infrastructure investments:

• Compute determines the ceiling of scientific ambition. Without frontier‑scale capability, groups are compelled to make use of smaller fashions, narrower datasets, and extra restricted exploration of organic and chemical complexity.
• Compute creates worth far past discovery. Manufacturing, improvement, regulatory, and industrial features can generate measurable returns when superior AI techniques have entry to enough computational scale.
• Compute ought to be evaluated as a scientific functionality, not solely as infrastructure. Like a telescope or microscope, better computational energy expands the questions researchers can ask and the fashions they’ll realistically construct.

This is the strategic shift Fuchs is pushing towards: compute is now not a background expertise. It is a scientific functionality that determines how far researchers can push the boundaries of discovery, how successfully organizations can leverage their information, and the way rapidly AI improvements translate into measurable enterprise and affected person influence.