Schrödinger's AI Co-Scientist Is Already Its Supercomputer's Main User

Editor's note: Bunsen is an AI product, and this piece could have run under AI. We are filing it under HPC, because everything that makes Bunsen work is traditional HPC infrastructure: job schedulers, queues, licenses, throughput contracts. The most consequential thing about the agent is what it does to demand for that stack.

Schrödinger has spent two decades training scientists to feed its internal HPC systems through purpose-built interfaces. Its AI agent needed about a month to become the machine's dominant user.

The lion's share of job submissions on the company's multi-thousand-GPU internal system now come through Bunsen, the AI co-scientist Schrödinger introduced in July, chief technology and operating officer Pat Lorton told Supercomputing News (SCN).

"It's been slowly creeping up. We're now at about 80 percent of jobs submitted coming from Bunsen," Lorton said of the roughly 200-person internal team that submits work to the machine. "We definitely expect higher supercomputing demand when you get help from an AI to create and submit jobs."

The standing infrastructure worry about agentic AI is that agents multiply inference calls and make compute spending unpredictable. Schrödinger built Bunsen on the opposite premise: the agent pauses itself while simulations run. The model never touches the scheduler, queues, or licenses, and the oldest cost controls in HPC set a ceiling the agent cannot break. The early consequence, inside Schrödinger at least, is a fuller supercomputer queue.

Days after Bristol Myers Squibb agreed to deploy Bunsen across its research organization, Lorton and Shane Brauner, Schrödinger's executive vice president and chief information officer, walked SCN through the engineering behind that discipline.

The agent that sleeps

Most agent frameworks keep the model in the loop end to end: plan, launch, poll, react. Schrödinger engineered the model out of the middle.

"We've taught the LLM agent to pause itself whenever jobs are running, instead of sitting there and using the LLM to query status," Lorton said. When a job completes, JobServer, the job-management layer that has fronted Schrödinger's compute for years, "wakes the LLM back up" and hands it the results. Otherwise, "the LLM is going to be chewing through credits just watching the jobs, and we've intentionally engineered that away, because we wanted to run so many jobs that that could get very expensive."

Argonne National Laboratory put the equivalent overhead for ChemGraph, its open-source computational-chemistry agent, at 2 to 5 percent of simulation time in SCN's reporting on the lab's continuous evaluation runs. Lorton has not run the numbers for Bunsen, but estimated that in dollar terms the orchestration share would come in "substantially lower than that."

The scheduler keeps its job

Deciding where work runs (local CPUs and GPUs, a customer's HPC scheduler, cloud capacity) is the one call Schrödinger deliberately kept away from the model. For placement, Bunsen knows JobServer's existing APIs and nothing more: it can query availability and queue length and submit jobs, but Schrödinger has taught it nothing beyond that interface, and the queues and license configuration keep working the way they always have. "Rely on that tried-and-true part of the ecosystem that works really well," Lorton said, "and don't try to have an LLM reimplement it."

Argonne reached the same rule independently; ChemGraph's agents hand work to a workflow engine that owns placement. A commercial vendor and a national lab converged on the architecture from different directions: the agent proposes, the workflow layer disposes.

The queue is the budget

Spend control falls out of the same design. In the Schrödinger ecosystem, a job does not leave the queue unless licenses are available for it, and licensing is throughput-based: customers buy the right to run a set number of CPUs and GPUs at once, usually on an annualized contract. However many campaigns the agent spins up, the theoretical maximum spend is computable in advance. There is no billable unit for Bunsen itself; the agent is a new way to consume infrastructure and licenses the customer already pays for.

"Bunsen now enables people to kick off an unbelievable amount of calculations, and obviously they don't want to go and spend $50 million in the cloud by accident," Lorton said.

After 30 years selling into pharma, where a fixed price ceiling is a precondition of the sale, Lorton finds the token-metered alternative strange. "I've been surprised how much the LLM providers have gotten away with having this essentially uncapped, limitless spend potential," he said. "Tokenmaxxing may be dead. I was surprised how long it lived."

What a run leaves behind

On observability, Lorton points at the file system before the model. Every Bunsen session runs on an underlying virtual machine the user can open: the parsing scripts the agent wrote and the logs that came back land there "as if a human were running it," alongside the LLM's reasoning traces.

The immediate payoff is triage. Simulation backends fail in convoluted ways; the meaning of a given failure log might live somewhere in hundreds of pages of documentation. Bunsen reads the log and answers on the spot: did you forget to add hydrogens? That kills the classic failure where a bolus of jobs launched before a weekend dies seven minutes in over something small and idles the cluster all night.

Brauner is more interested in what accumulates. How a scientist approaches a target today lives "in users' home directory or in somebody's head," he said; recorded sessions make it extractable, "like a computational ELN." "That kind of level of instrumentation is something we haven't really had before."

One boundary the pair drew unprompted: Schrödinger mines those usage patterns only from its internal drug-discovery team. "We don't have the right to mine that for our customers," Lorton said.

The question the whole field shares

An agent can also run a validly formed job with a poorly chosen method and return a plausible number that means nothing. SCN asked whether Schrödinger has measured the compute burned on campaigns that completed but were scientifically invalid.

"Not really," Lorton said. "We haven't done that historically either." Human scientists routinely explore calculation setups that fail, and the company treats that as a normal cost of doing science. Argonne gave SCN the same answer about ChemGraph. Nobody in the field appears to hold that number yet.

"It's an interesting thought," Lorton said, "examining how much it costs us when Claude comes up with bad ideas instead of a scientist comes up with bad ideas."

Bunsen adds review points: customers can require that no job launches without a scientist's sign-off, a gate Lorton expects to be popular, and the skills layered into the agent let it push back on a calculation that makes no sense, asking whether you are sure and recommending an alternative. "That just wasn't happening before," he said. The flip side, he conceded: higher throughput may let people submit bad ideas at higher throughput too.

Brauner's analogy is self-driving cars. Human drivers hit people too; the question is what rate "is going to need to be even better than the humans before we start really unleashing that with comfort." The "self-driving scientist," Lorton offered, at least warns you first.

The ledger Schrödinger watches is time. "If you can speed up your time to get a predictive model from six weeks to two weeks, that's an extra month the project team has where they're making accurate predictions before they send something to a wet lab," Lorton said. In his accounting, that’s worth more than the failed jobs along the way.

What BMS gets, in general terms

On Bristol Myers Squibb, both executives declined specifics (pharma guards its own disclosures) and would not characterize BMS's deployment plans or early-access history beyond the August 5 announcement. What they would describe is what any customer in BMS's position inherits.

"Bunsen just knows how to talk to it," Lorton said of Schrödinger's existing compute plumbing. "You don't have to deploy a whole new complicated tech stack. You don't have to get new security approvals. Pretty much every top-25 pharma and most biotechs are already using our platform." As for the large GPU systems pharma has been buying (BMS announced an eight-rack DGX Vera Rubin SuperPOD in July), Lorton kept it hypothetical: "If you were to buy an awesome DGX SuperPOD and you wanted to run jobs on that, Bunsen would be able to do it just like the rest of Schrödinger software would." Whether Bunsen will drive BMS's new machine is a question neither company has answered publicly.

Co-development follows a pattern the pair said they know from LiveDesign: work that touches a customer's critical IP is written and owned by the customer, while broadly useful work tends to get productized for everyone. "They would rather us productize it, and it be shared, so that they can essentially share the support burden," Lorton said.

The model layer stays deliberately uncommitted. Bunsen connects to models through the OpenAI-compatible API convention most providers and open-weight harnesses support; Schrödinger has tested it extensively with Anthropic and OpenAI models and tests notable open-weight releases in house. The reasoning follows the same data-gravity logic behind pharma's move to own its AI supercomputers. "When you see BMS and these companies purchasing these SuperPODs, and you're trying to figure out what they're doing with them, there's a non-zero chance that they end up running their own LLMs at some point, and they'll want to hook into it," Lorton said. "We want to make sure we are there if and when they do that." "That's not where our competitive advantage is," Brauner said of the model race. "We want to hook our wagon to somebody going fast."

The glue, and the demand curve

Schrödinger's own summary of Bunsen is unglamorous: connective tissue. LiveDesign holds enterprise data and models, Maestro serves expert users, and both could always reach the HPC layer, but nothing spoke natural language across all three. "It's this glue we've been waiting for forever," Lorton said: one interface that "can talk to both of those tools seamlessly, uses them exactly how they're supposed to be used, and can submit to HPC in a way that humans used to really struggle."

The commercial logic runs through the same throughput licensing that caps the spend. CEO Ramy Farid has said the licensing model is designed to capture the additional utilization Bunsen creates, and in second-quarter results ($58.9 million in revenue, a swing to profit, 27 percent ACV growth), the company told investors Bunsen's early access was successful, and new products contributed meaningfully to first-half growth.

Customers may be capped in the near term by hardware and license counts, Lorton said. "But when they start seeing the ROI of being able to access it like that, I expect demand to go up substantially." Brauner already sees the shape of it internally: "The old-school expert users now have a force multiplier attached to them, and the people that were not expert users now have that capability to really get in there and do scientific work that wasn't accessible to them before."

Bunsen remains in early access, with full commercial release expected by the end of 2026.