Google Cloud just made a calculated bet on a different sort of artificial intelligence. One that doesn’t chase fluent prose or clever banter. Instead, it tackles equations, measurements and the stubborn realities of the physical world.
The company announced today it will sell specialist models from SandboxAQ through its marketplace. These aren’t large language models. They’re large quantitative models. Trained on numerical data, scientific laws and lab results rather than internet text. The distinction matters. LLMs stumble on numbers. These models aim to get them right.
From Alphabet Spinout to Cloud Marketplace Staple
SandboxAQ began inside Alphabet’s X lab. It spun out in 2022. Backed by investors including Eric Schmidt, Marc Benioff and funds from T. Rowe Price. The firm raised hundreds of millions. Its valuation climbed past $5 billion. Now Google Cloud offers its technology to a wider audience. Researchers and enterprises can rent the models. Combine them with Gemini for reasoning while the quantitative systems handle the science.
Jack D. Hidary, CEO of SandboxAQ, put it plainly. “Partnering with Google Cloud enables us to deliver enhanced value to customers by leveraging their industry-leading cloud computing and AI infrastructure with our groundbreaking Large Quantitative Models (LQMs). This collaboration accelerates our efforts to develop Quantitative AI technologies and applications that address critical challenges—such as breakthroughs in Alzheimer’s research and other key areas of biopharma work.” (SandboxAQ)
Thomas Kurian, CEO of Google Cloud, echoed the sentiment. “SandboxAQ’s partnership with Google Cloud will help support the development of new models and solutions and bring them to enterprise customers more quickly.” (SandboxAQ)
The first models available include AQCat and AQPotency. AQCat identifies catalysts and materials. Useful for battery development, semiconductor manufacturing. AQPotency screens molecules. It finds candidates likely to bind to disease targets. Early private previews already run at partner labs. Results stay confidential for now. But the shift from specialist tool to rentable service stands out.
LLMs dominate headlines. They generate text brilliantly. Yet science demands precision. A wrong number in chemistry can invalidate an experiment. A hallucinated structure wastes months in the lab. Quantitative models sidestep that. They operate on physics, chemistry equations and proprietary datasets. No fluent paragraphs. Real outputs. The approach echoes DeepMind’s earlier wins. Its protein folding system reshaped drug development. Another effort discovered more new materials in a year than scientists had catalogued historically. (The Next Web)
Google pairs the marketplace addition with Gemini for Science. A set of tools built around existing projects. The AI co-scientist. AlphaEvolve coding agent. Empirical research assistant. NotebookLM. The bundle targets tedious parts of research. Literature review. Hypothesis generation. Data analysis. It won’t replace scientists. It aims to free them from drudgery.
But. The real test lies in results outside benchmarks. SandboxAQ already works with UCSF’s Stanley Prusiner lab. With Sanofi. The Michael J. Fox Foundation. In one neurodegenerative project it expanded chemical space from 250,000 to 5.6 million molecules. Hit rate jumped 30-fold compared to traditional screening. (Google Cloud Blog)
Those gains come from techniques like absolute free energy perturbation and generative methods trained on real experimental data. Not abstract theory. Cloud infrastructure makes the heavy computation feasible. SandboxAQ built its platform cloud-native. Scaling simulations that once took years down to weeks.
Competitors chase similar ground. DeepMind’s Isomorphic Labs pushes drug candidates toward clinical trials. Other hyperscalers court industrial R&D. Google positions its marketplace as neutral infrastructure. Third-party models sit alongside its own. Customers pick what fits. For quantitative science, SandboxAQ fills a gap general models can’t touch.
The timing aligns with bigger bets. SandboxAQ received a $500 million CHIPS Act award. Focused on AI for new semiconductor materials. Government interest in domestic tech supply chains runs high. Quantitative models that predict material properties faster could accelerate that work. Reduce reliance on trial-and-error fabrication.
Financial services and navigation also use these systems. Yet life sciences remain the headline application. Drug discovery costs billions and takes decades. Failure rates stay punishing. Tools that reliably predict molecular behavior cut both time and expense. SandboxAQ released a dataset of 5.2 million AI-generated compounds. Tied to experimental validation. Designed to train even better predictors. (Reuters)
Recent moves expand reach. SandboxAQ integrated its models with Anthropic’s Claude in May 2026. Another channel for quantitative AI to meet language capabilities. The pattern repeats. Connect the specialist engine to a conversational front end. Let domain experts query in natural language. Get back trustworthy numerical insight.
Critics wonder if this produces genuine discoveries or merely faster spreadsheets. Private previews will decide. Early signals from partners suggest more than incremental gains. Superior hit rates. Expanded search spaces. Predictions aligned with lab reality.
Google Cloud competes fiercely for AI workloads. General chat models draw consumers. High-margin enterprise R&D demands accuracy on physics and biology. By hosting SandboxAQ’s technology, Google captures usage without building every vertical model in-house. Smart infrastructure play. One that acknowledges limits of text-only training.
SandboxAQ itself keeps expanding. Cybersecurity applications. Advanced sensing. Quantum-inspired techniques run on classical hardware for now. The firm maintains ties to its Alphabet origins even as it operates independently. Chairman Eric Schmidt guides strategy. The combination of quantum thinking and modern AI defines its edge.
Researchers logging into Google Cloud next quarter will see new options. AQCat for materials screening. AQPotency for potency prediction. Integrated with Gemini. Backed by vast compute. The barrier to entry drops. What once required proprietary labs or massive grants becomes accessible to mid-sized research teams.
Whether that sparks a wave of new therapies, better batteries or novel chips remains to be seen. The infrastructure is now in place. The experiments can begin at scale. Science has always moved forward through better tools. This latest one arrives via the cloud. Ready to rent. Built for numbers instead of narrative.
And the race intensifies. More integrations will follow. More datasets will train sharper models. The question isn’t if quantitative AI matters for industry. It’s how quickly the predictions turn into products that reach patients, factories and consumers.
WebProNews is an iEntry Publication