There are two very different types of AI in the world right now, and most people only know about one of them.
The familiar kind, the chatbots, the writing assistants, the summarisers, were trained on text from the internet. They are genuinely useful for a wide range of tasks. But when a pharmaceutical researcher needs to predict how a molecule will bind to a disease target, or when an engineer needs to know how a new material will behave under industrial conditions, a text-trained AI is the wrong tool. The answer is not a paragraph; it is a number, a structure, a physical measurement.
This is the gap Google is now trying to fill commercially. Through its cloud marketplace, it is listing AI models from SandboxAQ, a company that builds what it calls large quantitative models: AI trained on real lab data, physical equations, and scientific measurements rather than prose. Two models come first, both arriving in Q3 2026. AQCat targets materials and catalyst research, helping teams rapidly screen candidates for things like battery chemistries, green hydrogen production, and semiconductor manufacturing. AQPotency targets drug discovery, letting researchers computationally evaluate thousands of molecular candidates before any of them go near a physical laboratory.
SandboxAQ is not a startup in the early-stage sense. It began inside Alphabet's experimental division in 2016, spun out in 2022, and has since raised over $950 million. Its backers include Google itself, NVIDIA, Ray Dalio, and Marc Benioff. The company holds a $500 million award from the US government's CHIPS programme to develop AI for semiconductor manufacturing, and its chairman is Eric Schmidt, the former CEO of Google. The institutional weight here is real.
The practical change for business operators is about access. Previously, quantitative scientific AI was available mainly to organisations with deep technical teams and specialist infrastructure. Through Google Cloud's marketplace, a pharmaceutical R&D team, a chemicals company, or a materials science group can subscribe to these models in the same way they might subscribe to any other cloud software. No specialist code, no custom infrastructure required.
Google also announced Gemini for Science alongside this, a bundle of research tools that pairs its general-purpose AI with specialist scientific capabilities. The design logic is sensible: use the conversational AI for reasoning and communication, use the quantitative model for the hard science underneath.
The drug discovery market alone was valued at roughly $112 billion in 2025 and is projected to reach around $187 billion by 2034. That is the size of the commercial opportunity Google and SandboxAQ are aiming at, alongside materials science and semiconductor manufacturing. Catalysts, separately, underpin more than 90% of all commercially produced chemical products, so AQCat's scope is wider than it might appear at first.
For industries touching pharma, chemicals, energy, or advanced materials, the relevant question is whether their R&D teams are aware that this category of tool exists and is now commercially accessible. The research that used to require a dedicated AI lab can now, at least in principle, be rented by the month.