Before a chemist makes a molecule, somebody must decide which molecule deserves to be made. That decision is expensive. A computer can propose candidates far faster than a laboratory can investigate them. The interesting question is how to choose. SandboxAQ has built a company around that gap: the space between a plausible idea and something worth spending an experiment on.
- Scientific AI ranks molecules and materials before expensive testing.
- Security and navigation products tackle problems enterprises already have.
- The useful output is a better decision, followed by evidence.
The computer they did not wait for
SandboxAQ emerged from Alphabet in March 2022, with founder Jack Hidary as CEO and former Google chief Eric Schmidt as chairman. Some of its roots go back to 2016. At launch, its cybersecurity customers included Mount Sinai, Vodafone Business and SoftBank Mobile. The software ran on classical computers.
That last detail helps explain the company. Quantum physics describes the behavior of matter; quantum computing is a particular way of computing. You can use conventional hardware to calculate quantum chemistry. SandboxAQ could therefore sell useful work while quantum computers continued developing. Hidary told Reuters he wanted outside investors to help accelerate growth.
First, manufacture the evidence
Its central idea is the Large Quantitative Model, or LQM. Where a language model learns patterns in language, these models learn numerical relationships from scientific data and simulations. The aim is to predict a property that matters: how a molecule binds, how a material behaves, which catalyst merits investigation.
In September 2025, SandboxAQ released AQCat25, a public catalyst dataset containing 11 million quantum-chemistry calculations across 40,000 intermediate-catalyst systems. Generating it consumed more than 400,000 GPU-hours on NVIDIA infrastructure. Cheap predictions have a surprisingly expensive childhood.
The dataset incorporated spin polarization, a magnetic effect relevant to industrial metals. That is the kind of unglamorous detail that can determine whether a model is useful. An elegant answer about the wrong physics does not improve a factory.
“The most accurate simulation methods are too slow to search the range of materials that matter at scale.”
Stefan Leichenauer / VP of Engineering
Leichenauer described this bottleneck in the company’s June 2026 CHIPS announcement. SandboxAQ’s ReAQT platform generates scientific training data and uses models trained on it to search more broadly. The company acquired Good Chemistry in January 2024, adding computational chemists, simulation software and the Tangelo open-source toolkit.
A discovery workflow, not a promise of success.
A dollar buys a shortlist
AQPotency makes the proposition tangible. Launched in August 2026, it ranks molecule-target pairs without requiring a solved three-dimensional protein structure. The advertised starting price was as little as $1 per 1,000 comparisons. Each prediction includes uncertainty and an indication of whether the target falls within the model’s reliable range.
In October, SandboxAQ added AWS Marketplace and Amazon SageMaker availability, letting discovery teams run the model inside their own AWS environments. Researchers supply a protein identifier and molecular representations, then receive rankings to guide their next experiments.
Those boundaries matter. A high score does not establish that a compound is safe, works in a patient or becomes a medicine. A target outside the model’s useful range deserves particular scrutiny. Faster screening is valuable when it improves the experiments that follow; the score itself is an intermediate result.
comparisons at AQPotency’s advertised launch starting price.Prediction cost. Laboratory work comes separately.
The inventory problem came first
Security supplied an earlier, less cinematic application. Large organizations need to know where encryption lives before they can replace vulnerable cryptography. SandboxAQ’s security software inspects that infrastructure. Its AQtive Guard business later extended this inspection approach toward AI assets, announcing an AI security posture management offering in December 2025.
The September 2022 acquisition of Cryptosense reveals the commercial logic. Founder Graham Steel had intended to raise more capital independently. Discussions with SandboxAQ changed the calculation: complementary technology, a shared approach to cryptographic migration and greater reach made joining attractive. Growth demanded more capacity than the original plan anticipated.
This is also a useful corrective to the company’s futuristic vocabulary. An encryption inventory is ordinary operational work. It becomes valuable because a bank or hospital cannot sensibly upgrade systems it has not identified.
The planet is another instrument
AQNav applies the same appetite for measurable signals to navigation. It processes magnetic measurements to help determine position when GPS is unavailable. In September 2026, SandboxAQ announced a flight test on Northrop Grumman’s Lumberjack unmanned aircraft, pairing magnetic and visual navigation.
Adoption still involves engineering. SandboxAQ’s deployment process assesses sensors, mounting, inertial inputs and flight data before supporting certification evidence. Performance depends on the aircraft and mission. The Earth offers a signal; turning it into a dependable instrument takes integration and testing.
Pay for a better experiment
Across these businesses, SandboxAQ sells enterprise software, scientific collaboration and integration, increasingly through cloud marketplaces and familiar AI interfaces. AQCat reached AWS Marketplace in August 2026; the company also connected quantitative models to Anthropic’s Claude. Scientific specialists remain central to the product even as access becomes easier.
The market already includes established computational discovery suppliers such as Schrödinger. SandboxAQ’s particular pitch combines its scientific datasets, models and distribution with a broader portfolio in security and sensing. Buyers should compare performance on their own task.
The capital is substantial. Its April 2025 Series E closed above $450 million, bringing reported funding since spinout above $950 million. A separate $500 million CHIPS R&D agreement announced in June 2026 targets semiconductor-related materials; Commerce receives a minority, non-voting equity stake.
For a reader considering similar tools, the copyable move is specific: choose one costly decision, establish a baseline, then measure whether predictions improve the next round of real work. SandboxAQ’s most persuasive demonstration will always be the experiment somebody can afford to run because the shortlist got better.