The 2022 Alphabet spinoff, chaired by former Google CEO Eric Schmidt, will use physics trained AI to propose chip making chemistries the US currently imports, and the program's standing depends on whether it can avoid the thin track record of generative AI in drug discovery.
The US Department of Commerce is paying SandboxAQ $500 million to find out whether AI trained on the laws of physics can do something the language-model AI boom has largely failed to do: produce a deployed, approved result. SandboxAQ, the 2022 spinoff of Alphabet that The Register characterizes as being chaired by former Google CEO Eric Schmidt, will use what the company calls Large Quantitative Models to propose new semiconductor manufacturing materials domestically, according to The Register's report on the award.
This is not a factory announcement. The grant funds research and development, not chip fabrication, and SandboxAQ will not be laying down silicon. Instead, the company will run its simulation software against five specific material targets named in the announcement: novel molecules and formulations for chip manufacturing, replacements for PFAS-based chemicals (a class of long-lived compounds often called "forever chemicals") used in chip production, new fabrication catalysts, neodymium- and rare-earth-independent magnets, and fab-powering batteries using lithium from non-majority-foreign sources. The Register characterized the program as a bet on AI-driven materials discovery, and Commerce framed it as part of the broader push to onshore semiconductor manufacturing that the CHIPS and Science Act of 2022 set in motion.
That framing makes SandboxAQ's bet legible only against a well-known prior: AI-driven drug discovery. Over the past several years, language-model-adjacent generative AI has spent billions on the hypothesis that it can propose new drug candidates faster than human chemists. The practical track record of approved drugs developed largely by those tools remains thin. SandboxAQ is making a related but different claim: that an AI grounded in physical laws, chemistry, and biology rather than human text can do for industrial materials what generative AI has not done for pharmaceuticals. The company calls these systems Large Quantitative Models. Whether that distinction holds is the story.
The Register's reporting leaves the timeline vague. The grant announcement names no milestones, no delivery dates, and no specific fabs or customers that would adopt any resulting material. SandboxAQ has not, in the public record available here, deployed a material that has cleared fab qualification. The five named targets are research deliverables, not products. Large Quantitative Models are a vendor term, not a peer-reviewed category, and the company's marketing line that its models are "trained on the laws of physics, not human language" should be read as a claim to evaluate rather than a settled fact.
The $500 million itself is real and material, but it is also a small share of the CHIPS and Science Act's roughly $52 billion portfolio. Read against that backdrop, SandboxAQ's award looks less like a moonshot and more like one allocation in a series of bets on where domestic capacity is most likely to fail without help. Materials science is a reasonable bet: a new fab on American soil still depends on chemistries, catalysts, and electrolytes that the US imports.
What to watch over the next several years is straightforward. The first test is whether SandboxAQ publishes a candidate material with measured properties, not just a press release. The second is whether any of the five targets reaches a fab pilot under a real customer, with a real qualification timeline. The third is whether the Department of Commerce publishes milestones against which the grant can be evaluated, or whether the program runs as an open-ended R&D line item. The AI-drug-discovery wave did not fail because its models were wrong. It failed because translation from a promising molecule to a deployed drug is the hard part. SandboxAQ is being paid to find out whether materials science is any kinder.