[Adobe Stock]
The throughline is a shift from “look what it can do” to “prove it holds up under real conditions.” Experts predict the commercial emergence of optical and physics-native computing, the maturation of Agentic AI in R&D, a widening gap between the AI “haves and have-nots” in pharma, and the urgent need for Explainability to satisfy global regulators.
Here’s what they’re calling.
Compute shifts from AI hype to simulation bottlenecks
Chene Tradonsky
The computer simulations that drive scientific discovery have a hardware problem: they are too slow and consume too much power. Digital processors struggle with the complex calculus known as partial differential equations (PDEs), the math that underpins everything from aerospace dynamics to climate modeling. These workloads are a natural match for light-based computation, yet they have largely remained theoretical. Chene Tradonsky, CTO and Co-Founder at LightSolver believes that experimental phase is over and free-space optical systems are ready to step in:
“In 2026, optical and photonic processors will begin to show their most practical impact in an unexpected area… they will finally move out of the lab to help solve partial differential equations (PDEs), the core mathematical work behind many of the scientific and engineering simulations that fill today’s HPC centers. This will not look like a sudden revolution in computing. It will be a gradual, steady shift as leading HPC centers plug photonic technologies into their existing simulation workflows to attack the slowest, most power-hungry PDE kernels.”
Tradonsky predicts physics-native computing—hardware that solves equations by mimicking the physics being modeled—will emerge as a recognized category alongside CPUs, GPUs, and quantum processors.
From chat to agents
Mirit Eldor, Managing Director, Life Sciences, Elsevier
While 2025 was noisy with talk of “agents”—AI systems that act autonomously rather than just chat—the technology has yet to fundamentally reshape how R&D labs actually function. Adoption estimates have been pessimistic, suggesting a slow roll-out for autonomous science. However, Mirit Eldor, Managing Director, Life Sciences Solutions at Elsevier argues that 2026 is the year this potential converts into kinetic energy, driven by successful pilots that are finally moving the needle:
“2026 will be the true ‘Year of the Agent’. Agentic AI was talked up in 2025, but didn’t have a huge impact on R&D processes like drug development. But now that is changing, as examples from Lilly (Kernel Lilly), the EPFL (ChemCrow), and others show that agents are making the difference in life sciences.”
Mike Connell
The constraint in R&D has rarely been a lack of ideas; it is the inability to filter them efficiently. In the race to move from concept to viable product, Generative AI is a powerful idea engine, but it lacks the discernment to serve as a “truth engine” for high-stakes decisions. Mike Connell, COO at Enthought suggests the industry’s focus is correcting toward prediction rather than just generation, prioritizing tools that can validate the right ideas fast:
“True ROI for R&D will not come from the tool that generates 10,000 new ideas. It will be from the tool that confidently tells you which 9,999 to ignore.”
Researcher trust is still the bottleneck
Stuart Whayman
Despite the hype, adoption remains uneven among scientists who require absolute accuracy. Many corporate researchers still refuse to use AI for core tasks like designing experiments or generating hypotheses, citing a fundamental lack of trust in the outputs. Stuart Whayman, President, Corporate Markets at Elsevier highlights the skepticism that still pervades the lab, noting that trustworthy, purpose-built tools are the only way to break the stalemate:
“Currently, 44% of corporate researchers would not use AI to write or draft papers, 47% would not use it to generate hypotheses, and 49% would not use it to design experiments. Meanwhile, only 27% say that AI tools are trustworthy.”
Whayman thinks 2026 is when that starts to shift—not because the models get better, but because research-specific platforms with traceable sources make the outputs defensible.
Unified AI replaces point tools
Cameron Ross
Researchers have spent the last few years toggling between fragmented apps for search, summarization, and coding. This “siloed” approach leaves significant value on the table, as roughly a third of researchers have yet to integrate AI into their workflows at all. Cameron Ross, SVP Generative AI, Corporate Markets at Elsevier predicts a move toward consolidation, where unified platforms bring models and data together to answer complex queries in one place:
“In 2026, AI silos in research will start to come down and R&D organizations will embrace unified AI. There is currently a rapid move from AI point solutions in R&D to unified AI tools that bring together models and data to answer any question a researcher has.”
Data volume forces new interfaces
Neil Ward
The sheer scale of biological data is outpacing the human ability to write code to analyze it. With population-scale studies on the horizon, the industry faces a data crunch that specialized bioinformatics code simply cannot handle alone. Neil Ward, VP and General Manager, EMEA at PacBio sees a future where major AI players partner with science firms—like the recent collaboration between 10x Genomics and Anthropic—to build natural-language interfaces that tame this complexity:
“Analyzing one person’s genome involves tens of thousands of lines of code, and population-scale studies could generate up to 15× more data than YouTube over the next decade.”
Ward sees major AI players partnering with science firms to build natural-language interfaces for genomics data, “like ChatGPT but purpose-built for science.”
Huntington’s and repeat expansion disorders
Not everything is a forecast. Neil Ward points to a 2025 milestone that sets up 2026:
“2025’s landmark success for slowing Huntington’s disease marks one of the first major breakthroughs for a group of genetic conditions known as repeat expansion disorders. In 2026, we can expect the Huntington’s success to spark a surge of research and investment into tackling other repeat expansion disorders…”
Ward argues long-read sequencing is the only technology that can accurately measure these repetitive DNA regions—positioning it as foundational for a wave of research into ALS, FTD, and similar conditions.
Drug development climbs the prediction stack
Rob Bradley
Structure prediction was the proof point, but the next frontier is tackling the different layers of biological organization—from molecules to cells, to tissues, and eventually multi-organ systems. Rob Bradley, Co-Founder at Synthesize Bio says the payoff is higher up the stack, where decisions get expensive. He believes that while making accurate predictions gets harder as you move up that stack, the utility for clinical success skyrockets:
“A rough estimate of a protein’s structure usually isn’t very useful, but a rough estimate of clinical trial success can be incredibly valuable. I’m especially excited about predicting toxicity and drug responses across complex tissues and organs.”
Bradley doesn’t see AI replacing scientists—he sees it giving them “superpowers” to analyze data faster and infer complex biological responses from sparse experimental data.
The AI haves and have-nots
Mike Hampton
The cost of computing power is beginning to warp the traditional competitive landscape. Historically, nimble biotechs led early innovation while big pharma acquired the winners. But as R&D becomes increasingly compute-intensive, requiring massive investment in GPUs and simulation infrastructure, the advantage is shifting. Mike Hampton, Chief Commercial Officer at Sapio Sciences warns of a structural divide where the “haves” can out-simulate the “have-nots”:
“In 2026, we expect to see an acceleration of the current trend where large pharmaceutical companies continue to invest in GPUs and AI technologies that support in silico design, simulation, and decision-making at scale… rapidly creating a growing divide between the AI haves and have-nots.”
The pre-AI model—where nimble biotechs led early innovation and big pharma acquired the winners—may be inverting. Strategic GPU investments are letting large players move faster on the compute-intensive work that used to be neutral ground.
Governance moves from policy to practice
Michelle Gyzen
For regulated industries, the question is not if AI will be adopted, but how it will be controlled. As global guidance accelerates, static dashboards are giving way to dynamic mapping capabilities, but the core requirement remains accountability. Michelle Gyzen, Senior Director, Strategic Regulatory Solutions at IQVIA anticipates a crackdown on “black box” algorithms, pushing regulatory teams to adopt governance models that prioritize explainability above all else:
“Regulators across the world will increase expectations when it comes to AI adoption in submission and review workflows. This means regulatory teams must strengthen their AI governance models to meet new accountability standards. These models should account for explainability as it rises to the top of every AI requirement in regulated environments.”
Future pharmacovigilance inspections will begin examining how companies validate and monitor AI systems that influence safety reporting.
Digital twins inch toward something auditable
Gen Li
Digital twins have been stuck in the pilot phase, often lacking the rigor required for regulatory approval. Sponsors have explored the technology to optimize protocol designs and reduce costly amendments, but uncertainty around regulation for digital trial arms has slowed broader adoption. Gen Li, Founder and President at Phesi argues that 2026 is the year they become credible, moving from experimental pilots to valid clinical tools:
“After years of experimentation, 2026 will mark the year digital twins move from pilot to practice in clinical development. To unlock the full value of digital twins, sponsors must earn regulatory trust through rigorous validation, ethical data governance, and clear documentation.”
The shift is being enabled by regulators like the FDA expanding their AI frameworks and finalizing risk-based guidance.
Manufacturing and commercialization absorb the AI wave
Maria Whitman
Discovery got the early AI spotlight because it’s open-ended and photogenic. But the margin pressure is downstream. Neil Smith, CPG President at Schneider Electric, argues AI will start playing a much bigger role in pharma manufacturing and data transfers—mainly to increase practical patent life and maximize margins.
On the commercial side, the boom in obesity and cardiometabolic drugs is forcing a total rewrite of the sales playbook. The traditional models cannot handle the unique blend of primary-care scale and specialty complexity required by these new blockbusters. Maria Whitman, Managing Partner, Healthcare at ZS predicts that companies will turn to Direct-to-Patient (DTP) strategies and new partnership models to manage affordability and capacity:
“Obesity and cardiometabolic growth will rewrite pharma’s commercial playbook, blending primary-care scale with specialty complexity using new activation models, partnerships and focus on affordability, capacity and patient support.”
Quantum shifts toward repeatable utility
Alex Shih
Quantum computing forecasts have often sounded like countdown clocks waiting for a singular explosion. But the gap between leaders and laggards is widening, and organizations that treat quantum as a strategic capability are already integrating it with HPC centers. Alex Shih, VP of Product at Q-CTRL frames 2026 not as a moment of sudden breakthrough, but as the year of repeatability, where value is engineered through software and systems integration:
“2026 will be the year quantum advantage becomes repeatable, increasingly expected in specific domains, and more importantly, understandable in its value. Rather than a single “breakthrough moment,” 2026 will mark the normalization of quantum utility, where advantage is engineered through software, systems integration, and disciplined execution—not promised by qubit counts alone.”
