Finding a single viable molecule usually takes a decade. Cost: billions. Result? Most candidates die in the lab anyway. The industry calls it failure. We call it Tuesday. A wave of AI startups promised to fix the bottleneck, but most just handed slightly sharper knives to people who already know how to hold them. They didn’t solve the real problem. They just made the pain marginally less annoying.
SandboxAQ has a different take.
They think the model isn’t the issue. It’s the door.
Physics over Text
SandboxAQ joined forces with Anthropic. Now, their scientific models live inside Claude. You talk. The machine calculates. No specialized infrastructure needed. No PhD in computing required.
Five years ago, this spun out of Alphabet. Eric Schmidt sits on the board as chairman. They’ve raised over $950 million, though they’ve also dabbled in cybersecurity. The heavy lifting happens with what they call LQMs: large quantitative models.
Not large language models. These don’t guess the next word based on Twitter posts. They respect the rules of the physical world. Physics-grounded. They run quantum chemistry. They simulate molecular dynamics and microkinetics—basically watching how chemical reactions unfold, particle by particle, before anyone touches a beaker.
“LQMs are AI models engineered for the ‘quantitative economy’ — a $50 trillion sector spanning biopharma to energy.”
They aren’t building a chatbot for emails. They are chasing the sectors AI was actually supposed to change.
Who Can Use It
Other well-funded players like Chai Discovery and Isomorphic labs bet everything on the science itself. SandboxAQ bets on the user.
“For the first time… someone can access [this] in natural language.” — Nadia Harhen, Sandbox AQ
Previously, if you wanted SandboxAQ’s models, you had to bring your own servers. Now, the conversation does the heavy lifting. This changes the customer profile entirely. It wasn’t always open to the general public anyway, but it is now open to those who don’t code for a living.
Their typical client works in a pharma or industrial lab. They are experimentalists. Research scientists. They are looking for new materials, yes, but more importantly, they are looking for things that actually work outside the simulation.
Harhen admits their clients come from the scrapheap. People who tried every other software stack first. Who saw promising data turn to dust when moved from a screen to the real world. The translation failed. So now, they try this.
It might actually work this time.
Or maybe we are still just shouting at machines and hoping for a miracle. Who knows? The data will tell us, eventually. If it ever runs long enough.
