Google Cloud will offer specialized AI models designed for scientific purposes.

      Google is incorporating SandboxAQ’s ‘large quantitative models’ into its cloud marketplace, integrating Gemini with AI trained on scientific equations and lab data.

      The large language models that drive most of the AI sector excel at language generation but are often unreliable with numerical data. Google’s recent action acknowledges the need for different models in scientific contexts.

      The company announced that it will begin offering specialized AI models from SandboxAQ via Google Cloud, introducing what SandboxAQ refers to as large quantitative models to its cloud marketplace. This initiative aims to enhance access to AI tailored for drug discovery, materials science, and semiconductor manufacturing for enterprises and researchers alike.

      The emphasis on this distinction is crucial. Large language models are trained on text and excel at generating it, while large quantitative models, as described by SandboxAQ, are trained on numerical data and scientific equations, making them better suited for chemistry, biology, and physics, where accurate answers often take the form of numbers or structures rather than detailed prose.

      Researchers will have the opportunity to combine these models with Gemini on Google Cloud, using the language model for reasoning and interfacing, while leveraging the quantitative model for the underlying scientific analysis.

      Google has introduced this marketplace feature alongside Gemini for Science, which is a suite of tools and experiments designed to enhance the research workflow itself. It builds on projects that the company has been developing for some time, including its AI co-scientist, the AlphaEvolve coding agent, an empirical research assistant, and NotebookLM, and is positioned as a means to expedite the routine and labor-intensive aspects of the scientific method rather than replacing scientific professionals.

      This framing aligns with Google’s commitment to scientific advancement. DeepMind’s work on protein structures has already transformed certain areas of drug development, and a separate initiative has produced an AI that identified more new materials in a year than the scientific community had cataloged throughout its entire history. The common factor is that the most valuable AI in scientific fields tends to be specialized and trained on actual measurements rather than general knowledge sourced from the internet.

      The commercial rationale is clear. Google is competing with other major cloud providers to attract enterprises running AI operations, and scientific and industrial R&D represent a valuable segment that general chatbots do not effectively cater to.

      By offering specialized models through the marketplace, which already features a diverse catalog of third-party systems, Google can meet that demand without the necessity of developing every domain-specific model internally.

      This also aligns with a wider trend of converting AI capabilities into tangible laboratory results. DeepMind's drug-discovery spinoff, Isomorphic Labs, is progressing toward clinical trials, while competitors in the industry are racing to translate algorithmic potential into effective treatments and materials beyond just theoretical applications. By providing quantitative models to enterprise researchers, Google aims to establish itself as the foundational infrastructure in this competitive environment.

      Google has stated that these capabilities are already being utilized by partners in a private preview for real-world R&D, albeit with limited details disclosed regarding the organizations involved and the outcomes achieved.

      The marketplace listing signifies a significant shift: a category of AI that was previously limited to specialized laboratories is now accessible for research teams to utilize on a rental basis. The outcome of this initiative—whether it will yield groundbreaking discoveries or simply more efficient data processing—is what the private previews aim to determine.