“Scientists have felt left out of the AI craze. That’s the chasm we’re closing”, Bill Fitzgerald, SandboxAQ

The AI dividend has been paid out unevenly. Lawyers draw on it, marketers have absorbed it into muscle memory, financial analysts lean on it before their first coffee. The people who spend their lives at the bench, running the actual science, have mostly watched the revolution happen to everyone else. That imbalance is what drives Bill Fitzgerald’s work, and he has decided it is his to fix.

As VP of Growth and Ecosystems at SandboxAQ, Fitzgerald leads growth for the company’s AI Simulation business, overseeing the commercialisation of software and consulting services across novel catalyst development and molecule discovery. His organisation is chartered with expanding the use of Large Quantitative Models, or LQMs, to enrich research datasets grounded in the physical world. He knows the machine from the inside. Before SandboxAQ he was Head of Biotechnology for Google Cloud, building partnerships with venture capitalists, biotechs, academia and pharma, and working on the commercialisation of AlphaFold. More than 20 years in the biotech and biopharma ecosystem sit behind him, along with advisory seats at MassBio and Scientific American.

The occasion was SandboxAQ’s announcement on 2 July 2026 that it would make its LQMs accessible through Google Cloud’s Marketplace, letting users tap physics-grounded scientific models directly from the conversational AI tools they already use, with no specialised code or infrastructure required. AQCat, aimed at materials and catalyst discovery, comes first, with AQPotency for drug discovery to follow, both expected in the third quarter.

LQMs are AI models built on real-world lab data and scientific equations, engineered for the more than $50 trillion global quantitative economy that spans biopharma, energy, advanced materials and financial services. We spoke about distribution as strategy, the limits of the models everyone else is racing to build, and the community he thinks the industry forgot. What follows is edited for length and clarity.

In a conversation with The SourceCode, Fitzgerald spoke about the comapny’s plans and the importance of LQM

Start with the choice itself. SandboxAQ can, and does, distribute on its own. Putting the LQMs inside a marketplace is a deliberate decision. What is the thinking?

The main reason is to meet people where they are already doing the work. Machine learning workflows are running inside the biggest catalysis and life sciences companies today, and the field has moved quickly. The most valuable researchers are the ones who can move fluidly between the wet lab and the computational lab. A distribution mechanism like Google meets those customers inside the pipelines they already run, so as they assemble their own data alongside external and newly generated sources, they can inject our models directly into that flow for prediction accuracy and simulation.

It is one distribution mechanism among several, not the only one. We will also release models on our own platform. The strategy is to democratise access to these highly advanced scientific models and make them as easy to reach as possible, and Google is one of the key routes to that.

So Google is a beginning, not an endpoint. Is there a pipeline of further marketplace partnerships behind it?

There are, you will see a steady cadence of model releases distributed through the various avenues our hyperscaler partnerships open up. Those partnerships work at different tiers. At the simplest level, a model sits in a partner’s marketplace. Beyond that, we integrate more deeply into the workflows they already run, such as a Gemini scientific workflow, a Cloud for Science workflow, or a Microsoft discovery workflow. We intend to reach distribution across each of them.

The first model out is AQCat, and it goes after adsorption energy calculation, the opening step in catalyst and materials discovery. Why has that one calculation been so expensive, and what shifts once it runs at scale?

Consider what a heterogeneous catalysis company is actually for. Its job is to build and expand catalysis, not to build simulation models. A model like AQCat costs tens of millions of dollars in compute, and tens of millions more in scientific capital- the people who build it and capture the data behind it. That puts simulation out of reach for most of the companies that would benefit from it. Now those companies can simulate what is happening in their scale-out factories and across their development process.

Before a model like AQCat, innovation in catalysis came at a steep price. You had to pull people off a deployment or manufacturing line, gamble on reappropriating machinery and reagents, and physically test an idea. We have lowered the bar for simulation. The alternative was a costly, multi-million dollar joint venture to develop something like a sustainable aviation fuel. Ask a large oil and gas company to turn fats and biologics into fuel and you are asking it to spend millions. Simulating that problem on a system already built on first principles compresses the time to innovation and cuts the cost of getting there.

SandboxAQ has wired its models to both Gemini and Claude, rivals in the same market. Most companies pick a side. Why sit across two frontier LLMs at once, and what does that say about where you are taking the business?

I spent several years running the biotech organisation at Google, and I watched a lot of people struggle to find their footing with AI. I worked on the commercialisation of AlphaFold, and people would rush to the model, use it, and then ask what came next. The LLM market is bifurcated. Anthropic holds the largest share at 35%, Gemini is distributing at enormous scale, and plenty of people are still on ChatGPT. Our aim is simply to put the models where the customers already are and make deep science as easy to reach as possible.

Our LQMs are recognised as fundamentally different from what the LLM providers offer. Those providers are built on a transformer architecture descended from the BERT model Google introduced years ago. That kind of model reads context and predicts the next answer, but it does not perform the real computation an LQM does. Ask it to run a virtual screen of 10,000 ligands and it will return something that looks academic and reads well, with charts and graphs that seem convincing. A newcomer would be impressed. A computational biologist would see immediately that nothing had actually been computed.

What we do with the LLM providers is enhance that experience for the scientific community. The model still delivers the polished executive summary researchers are comfortable seeing every day. Underneath it, AQCat, AQPotency and our other models add the computational layer, so instead of a generalised graph you get real science, output that can be validated and defended when a scientist takes it to a development partner.

AQPotency carries you into drug discovery, a market forecast to grow from roughly $112 billion in 2025 to about $187 billion by 2034. That is crowded, well-funded ground. What is the specific problem you mean to solve at the early stage, and who are you really up against?

Our biggest competitor is traditional science itself, because a seasoned drug hunter still needs a substantial validation step before trusting a computer and an in silico simulation. What we do is make the selection process inherently more accurate as researchers move through the therapeutic pipeline. Early on, instead of sending thousands or tens of thousands of candidates to a CRO and grinding through a loop of down-selection, payment and iteration that eats weeks and months, they can simulate those candidates and bring their strongest 50 screens to the lab. The goal is to eliminate what is already known, and a great deal of what is known lives in physical science, in a compound that is toxic, or an absorption calculation that tells us a molecule will not bind under a given heat scenario.

I have spent a long time studying how pharma companies are organised, and they are organised very deliberately. They are structured by therapeutic area so that a failing asset can be cut cleanly and the rest of the business insulated from the risk. That focus comes with blinders. There is no incentive to share that a compound carried a toxic attribute, a dosing failure or a clinical problem, say across a GLP-1 drug and a blood pressure drug. The underlying knowledge is often pure physics and chemistry, well understood, but the data is never saved, because saving it and modelling it is expensive.

If I had a single dollar and I were running AstraZeneca, I would spend it on the best drug hunter, not on compute from NVIDIA. A lot of the market is chasing the AI hardware boom. Eli Lilly bought a billion dollars in chips, Roche followed with a billion of its own, and now everyone is hunting for the workloads to justify them, because they were never organised to generate those simulations in the first place. What we bring is a toolkit of workflows, potency and affinity binding predictions, toxicology, and knowledge graphs that map where all the data sits, so a company can reach an AI-driven outcome. We get customers 50, 60, 70% of the way there and strip out the costly steps, leaving them to fine-tune on their own data.

Here is the counterintuitive part. Materials discovery ships first, drug discovery second, though most people would rank medicine as the bigger prize. What is the logic underneath that order?

We have an exceptional team on the chemical simulation side, and the AQCat model was published in Nature. The value it can add to this market is substantial, and materials is a less noisy market to simulate into. The materials science landscape is structured in a way that lets us target customers precisely: battery companies, photovoltaic companies, and go after them directly. The biological landscape is far more fractured and heavily regulated, so leading with hyperscalers there makes sense because it eases our route to market.

Pull back to the horizon. As one year closes and the next opens, where is AI actually heading? Still the age of agents, or something that turns back towards the science?

The real shift in AI will be towards explainability. I think we have reached the ceiling on what LLMs alone can do, and most users cannot get near the top of that range anyway. The next move is genuine integration of AI into workflows. We are building our models around that idea. The GPCR workflow we released at BIO a few weeks ago walks a researcher through a sequence of tasks and explains each step so that it can stand on its own. That matters, because the past year has been a race to benchmark who is fastest, who can push furthest. There is real value in that, but the emphasis is turning towards outcomes and the ability to explain them. With the FDA loosening its stance on AI and the White House hardening its thinking on how the technology is used, the bar for explainability will be high as we move into 2027.

Last word is yours. What have we not talked about that matters most to you?

On a personal level, I believe an entire community has been left out of the AI boom, and that community is scientists. Getting this science into their hands is why I am here. Press releases got easier to write, imagery got easier to produce, work got easier for lawyers and finance teams, and scientists sat back asking what about them. That is the chasm we are closing in the quantitative sciences, alongside our work inside the LLM ecosystem. We do a great deal without LLMs, too. People can work with LQMs directly. We run frontier partnerships, including a $200 million partnership with MapLight and the Bain collaboration, both already announced.

There are a couple of hundred thousand people on Claude for Science right now who cannot see the computation they need, which is why we launched with Anthropic. Jonah Cool, their head of life sciences, recognises that we fill that gap, and Google recognises that we can help fill it too. The point I want to leave you with is simple: we want to broaden the reach of AI to a community that has yet to be served by it.

The researchers the boom passed by

Fitzgerald keeps circling back to the same conviction, that the AI revolution has so far rewarded the professions that write, sell and litigate, while the people running the underlying science have been left to watch. His answer is not another frontier model but a layer beneath the ones already in wide use, computation that turns a plausible-looking output into something a scientist can defend. The marketplace strategy, the deliberate refusal to pick a single LLM, the decision to lead with materials rather than medicine, all of it serves that same end. Whether SandboxAQ can close the gap it describes will depend on adoption inside laboratories that have long treated simulation as a luxury. For now, the ambition is unambiguous, and it is aimed squarely at the researchers the last few years passed by.