The large language models that power most of the AI industry excel at generating text but struggle with numerical accuracy. Google's latest marketplace move acknowledges that science demands a different kind of artificial intelligence. By adding SandboxAQ's large quantitative models (LQMs) to Google Cloud, the tech giant is betting that domain-specific AI, trained on equations and laboratory data rather than prose, will unlock faster breakthroughs in drug discovery, materials science, and semiconductor design.
SandboxAQ, a company spun out of Alphabet's moonshot lab, develops AI systems that learn from structured numerical datasets and scientific equations. Unlike traditional large language models (LLMs), which predict the next word in a sentence, LQMs predict molecular structures, material properties, and physical interactions—outputs that are numbers or coordinates, not paragraphs. This distinction is critical in fields like chemistry and biology, where precision matters and hallucinations can lead to wasted lab work or flawed hypotheses.
Through Google Cloud Marketplace, enterprise research teams will be able to access these specialist models alongside Google's own Gemini AI. The integration allows researchers to use Gemini for natural language reasoning and interface, while the quantitative model handles the underlying scientific computation. For example, a scientist could ask Gemini to 'find a stable catalyst for ammonia synthesis at room temperature,' and the system would query the LQM for materials candidates, returning structures ranked by predicted stability and catalytic activity.
Why science needs different AI
The announcement comes as the AI industry grapples with the limitations of LLMs in handling quantitative tasks. Benchmarks show that even advanced models like GPT-4 can make basic arithmetic errors or fail to understand probabilistic reasoning. In scientific research, such errors are unacceptable. The National Institutes of Health and other funding bodies have increasingly called for AI that can be trusted with experimental data.
Large quantitative models address these limitations by training on curated scientific datasets—crystal structures, reaction databases, genomic sequences—rather than internet text. This narrow focus allows them to achieve high accuracy on specific tasks, though it also means they cannot generalize to unrelated domains. Google's strategy is to complement this specialization with the broad linguistic capabilities of Gemini, creating a hybrid workflow.
Alongside the marketplace listing, Google introduced Gemini for Science, a suite of tools designed to streamline the research process. This bundle includes the AI co-scientist unveiled earlier this year, AlphaEvolve—a coding agent that automates experiment design—an empirical research assistant, and NotebookLM for summarizing papers. The aim is not to replace human scientists but to take over the routine, time-consuming steps of the scientific method: literature review, hypothesis generation, experimental planning, and data analysis.
DeepMind's scientific lineage
Google's push into scientific AI builds on work from its DeepMind subsidiary. DeepMind's AlphaFold revolutionized protein structure prediction, enabling researchers to model thousands of proteins in hours instead of years. More recently, an AI from Google Research discovered hundreds of thousands of new materials in a single year—more than science had catalogued in its entire history. These successes have demonstrated that narrow AI, trained on real measurements, can deliver outsized impact in scientific domains.
The commercial logic is clear. Google Cloud competes with Amazon Web Services and Microsoft Azure to be the default platform for enterprise AI workloads. Scientific and industrial research and development is a high-value, high-growth segment that general-purpose chatbots serve poorly. By offering specialist models through its marketplace, Google can capture demand without building every domain model itself. The marketplace already hosts thousands of third-party AI models, and adding SandboxAQ's LQMs fits a pattern of curating best-in-class tools.
Real-world applications and competitive landscape
Sectors from pharmaceuticals to semiconductors stand to benefit. Drug discovery pipelines typically take over a decade and cost billions; AI models that can predict pharmacokinetics, toxicity, or binding affinity reduce the number of failed candidates. In materials science, LQMs can propose new alloys or polymers with desired properties—strength, conductivity, heat resistance—saving years of experimental trial and error. For semiconductor fabrication, AI can optimize lithography processes or suggest novel dielectrics.
SandboxAQ's models are already used in private preview by partners in pharmaceutical and materials companies, though the company has not disclosed specifics. The marketplace listing makes these capabilities available to a wider audience on a pay-as-you-go basis. Researchers at universities, startups, and large corporations can now rent scientific AI without building the infrastructure or expertise from scratch.
The competitive landscape is heating up. Microsoft Azure offers similar scientific AI tools through its Azure Quantum Elements platform, which includes AI for chemistry and materials simulation. Amazon Web Services provides access to models from startups like Schrödinger and D. E. Shaw Research. Google's differentiator is the tight integration with Gemini and its existing ecosystem of cloud services, as well as DeepMind's research pedigree.
Yet challenges remain. Training LQMs requires high-quality, curated datasets that are often proprietary or siloed. Regulatory hurdles in industries like pharmaceuticals mean that AI-discovered candidates still need extensive real-world validation. And convincing risk-averse researchers to trust AI predictions is a cultural shift that takes time. Google's private previews are designed to gather evidence that these models produce reliable results—faster spreadsheets or genuine discoveries will determine the technology's adoption.
Isomorphic Labs, DeepMind's drug-discovery spinout, is moving toward clinical trials for several candidates, leveraging AlphaFold and other DeepMind systems. The marketplace listing does not directly include Isomorphic's work, but it positions Google Cloud as the infrastructure layer for the broader AI-in-science race.
From a strategic perspective, Google's move also protects against the potential commoditization of LLMs. As open-source models catch up and costs drop, the profit margins on general-purpose AI shrink. Specialized models for verticals like science, healthcare, and engineering offer higher margins and deeper customer lock-in. The LQM marketplace is a bet that the future of enterprise AI is not one giant model but many narrow ones, each optimized for a specific domain and connected through a common cloud platform.
For scientists, the promise is a significant acceleration of the research cycle. Instead of spending months manually searching for candidate molecules or materials, they can generate hypotheses in hours. Instead of sifting through thousands of papers, they can ask an AI to summarize the state of the art. The bundle of tools under Gemini for Science—co-scientist, coding agent, research assistant—is designed to cover the entire workflow from idea to lab experiment.
The true impact will take years to measure. Scientific discovery is famously resistant to shortcuts; many AI-generated leads prove false when subjected to real-world testing. But by putting quantitative models directly into the hands of researchers, Google is lowering the barrier to experimentation. Whether the result is a flood of new drugs and materials or simply a more efficient way to verify existing hypotheses, the infrastructure being built today will shape the future of scientific AI.
Source: TNW | Google News