A new multi-agent AI system from Google DeepMind, Co-Scientist, recently identified drug candidates for liver fibrosis; one blocked 91% of a scarring-linked response in lab tests, according to Google DeepMind. This Gemini-powered system iteratively generates, debates, and evolves novel hypotheses, accelerating discovery for complex medical challenges. Scientific discovery is accelerating through powerful AI and supercomputing, but the methods and economic structures supporting this acceleration are becoming increasingly complex and varied. The National Laboratory of the Rockies' Kestrel supercomputer, for example, supported over 500 modeling and simulation projects in Fiscal Year 2025, according to National Laboratory of the Rockies. Such reliance on high-performance computing is now pervasive. This rapid acceleration creates a fragmented, opaque economic landscape for research, where access to cutting-edge tools and their commercialization models are as complex as the discoveries themselves. The future of scientific breakthroughs will increasingly depend on the strategic adoption and ethical navigation of advanced computational tools, potentially creating a divide between those who can harness these technologies and those who cannot.
AI and HPC: New Frontiers in Discovery
1. Google DeepMind's Co-Scientist
Co-Scientist, a multi-agent AI system built with Gemini, excels at accelerating hypothesis generation and experimental design in complex biological and chemical research. It iteratively generates, debates, and evolves novel hypotheses, as demonstrated by its identification of drug candidates for liver fibrosis; one blocked 91% of a scarring-linked response in lab tests. This system, also applied to antimicrobial resistance and plant immunity, orchestrates specialized agents for generation, reflection, and meta-review, managed by a supervisor. Its broad applicability across scientific domains suggests a paradigm shift in how initial research questions are framed and pursued, though its sophisticated computational infrastructure demands significant investment.
2. NLR's Kestrel Supercomputer
The Kestrel supercomputer provides high-performance modeling, simulation, and data-intensive computations across diverse fields. In Fiscal Year 2025, it supported over 800 users and more than 500 projects, yielding over 700 technical outputs, including 293 peer-reviewed publications, according to National Laboratory of the Rockies. NLR's ElectroCat team, for instance, leverages Kestrel for machine learning to find alternatives for scarce metals in battery technologies. This level of computational power and output solidifies supercomputers as critical national research infrastructure, yet their restricted access to specific institutions creates a potential bottleneck for broader scientific engagement.
3. Generative AI Technologies
Generative AI, including Large Language Models (LLMs) and diffusion models, addresses data scarcity, drafts scientific text, and assists knowledge discovery. Initially developed to overcome high-quality scientific data limitations, according to Nature, these tools are now widely used for drafting, translation, and knowledge exploration since ChatGPT's 2022 launch, according to pmc.ncbi.nlm.nih.gov. While improving efficiency in scientific communication, their potential for generating inaccurate or biased information necessitates rigorous human oversight, challenging the traditional peer-review process.
4. Knowledge-Guided Deep Learning Approaches
Knowledge-guided deep learning, exemplified by physics-informed neural networks, enhances AI model reliability and interpretability by embedding prior scientific knowledge directly into their architecture, according to nature.com. This approach improves robustness and trustworthiness, reducing reliance on purely data-driven insights. Its effectiveness, however, hinges on significant domain expertise for knowledge integration, suggesting a growing demand for interdisciplinary researchers who can bridge AI and specific scientific fields.
5. Machine Learning for Materials Science
Machine learning accelerates the discovery and optimization of novel materials. NLR's ElectroCat team, using the Kestrel supercomputer, applies ML to find alternatives for scarce metals in battery technologies, according to National Laboratory of the Rockies. This bypasses lengthy experimental cycles by identifying patterns and predicting material properties. While speeding up material discovery and reducing costs, its reliance on large, high-quality datasets means that advancements in data generation and curation are as crucial as the algorithms themselves.
6. Multi-Agent AI Systems
Multi-agent AI systems, like Co-Scientist, excel at collaborative problem-solving and complex task distribution in scientific research. Composed of specialized agents (e.g. generation, reflection, meta-review) orchestrated by a supervisor, according to Google DeepMind, this architecture distributes tasks, accelerating complex scientific inquiry. While offering robustness through distributed intelligence, the intricate design and orchestration of these agents pose significant challenges for debugging and interpretability, potentially limiting their widespread adoption without standardized development frameworks.
7. Virtual Modeling and Simulation
Virtual modeling and simulation test hypotheses, optimize designs, and explore complex systems in virtual environments. NLR's Kestrel supercomputer supported over 500 such projects in FY2025, according to National Laboratory of the Rockies. This approach, exemplified by tools like BioReactorDesign, accelerates discovery by reducing the need for physical prototypes and experiments. While reducing costs and enabling exploration of impractical scenarios, the accuracy of these models remains contingent on robust input data and underlying physics, emphasizing the continued need for empirical validation.
The Evolving Economics of AI-Driven Research
| Pricing Model | Description | Prevalence (Primary Model) | Implications for Researchers |
|---|---|---|---|
| Hybrid Models | Combines elements of different pricing structures, such as a base fee plus usage-based charges. | 28.3% of respondents, according to complexdiscovery | Offers flexibility but can lead to unpredictable costs; requires careful budget planning. |
| Per-Document Models | Charges based on the number of documents processed or reviewed by the AI system. | 28.3% of respondents, according to complexdiscovery | Provides clear cost per unit of work; suitable for projects with predictable document volumes. |
| Per-GB Models | Pricing determined by the gigabytes of data processed or stored by the AI service. | 11.3% of respondents, according to complexdiscovery | Scales with data volume; effective for large datasets but costs can grow quickly with data expansion. |
| Per-Token Pricing | Costs are calculated based on the number of computational tokens consumed by the AI, often used for generative AI outputs. | 5.7% of respondents, according to complexdiscovery | Granular cost control tied to AI processing; can be complex to estimate for open-ended tasks. |
| Flat Monthly Subscriptions | A fixed fee for unlimited or tiered usage within a specific timeframe. | 5.7% of respondents, according to complexdiscovery | Predictable budgeting; may not be cost-effective for low-usage scenarios or highly variable workloads. |
The diverse and often opaque pricing structures for AI services reflect a nascent market. Institutions must carefully evaluate costs against promised research acceleration. This fragmentation suggests that economic models for AI integration are becoming as complex as the scientific discoveries they facilitate, potentially creating significant barriers to entry for underfunded research initiatives.
Navigating the Accelerated Future of Science
The integration of AI introduces new economic models and demands a strategic approach, balancing advanced computation with fundamental biological exploration. While hybrid and per-document models dominate generative AI-assisted review, outcome-based pricing, though uncommon at 3.8% according to complexdiscovery, shifts how AI's value is perceived and monetized. This proliferation of fragmented pricing, from per-document to per-token, risks creating a two-tiered scientific community where access to cutting-edge acceleration is dictated by financial agility, not scientific merit. For example, Google DeepMind's Co-Scientist offers unprecedented speed in drug discovery, contrasting with Florida State University chemists synthesizing new molecules from sea sponge bacteria, according to Florida State University News. This methodological divergence requires the scientific community to actively decide how to integrate advanced computational tools while sustaining foundational exploratory science. By 2026, institutions must develop clear strategies to navigate this divide, or risk leaving valuable insights uncaptured.s undiscovered.
