Benchling and Baseten announced Benchling Inference, giving biotech customers scalable, cost-effective GPU capacity to train and run scientific models, without managing infrastructure. It comes preloaded with today’s top scientific models and the integrations to make in silico discovery work out-of-the-box for biopharma companies.
Between 2020 and 2025, the number of new scientific AI models released annually grew from 28 to more than 380. These models are now standard in R&D workflows, but the compute layer hasn’t kept up. Drug discovery is bursty by nature: teams wait on the physical lab, data comes in waves, then need to run 100,000 predictions in a few hours before going quiet for days. For most computational teams, that plays out as HPC queues with multi-week backlogs, GPU reservations sitting idle between data collection cycles, and predictions rationed during active campaigns.
Benchling Inference is built on the Baseten Inference Stack, a tightly integrated combination of a high-performance runtime (custom kernels, speculative decoding, KV cache optimizations) and inference-optimized infrastructure spanning 15+ cloud providers, with cold starts in 5–10 seconds. Benchling adds a biotech layer on top with pre-configured defaults for scientific models and deployment options for organizations with strict data residency requirements. By aggregating demand across the industry, Benchling also brings better economics to biotech startups.
Also Read: Owkin to Build AI Agents as Part of a Multi-Year K Pro License Agreement With AstraZeneca
With Benchling Inference, scientists can deploy third-party models or serve internal models built on their own experimental data from a unified compute environment. For teams with data sovereignty requirements, the Baseten Inference Stack runs identically in Baseten Cloud, inside a customer’s virtual private cloud (VPC), or a hybrid of both so predictions never have to leave their environment. Computational scientists working in Jupyter notebooks or via SDK can call inference directly through Benchling.
“Biotech has entered a new era where AI models trained on proprietary experimental data could unlock breakthroughs that weren’t possible before. The bottleneck has been infrastructure and biotech research labs should not have to become GPU experts to run frontier models on their data. By partnering with Benchling, we bring six years of inference expertise directly into the environments where the science happens” said Amir Highighat, CTO & Co-Founder of Baseten.
“Access to compute is becoming a strategic advantage. But we hear from computational scientists that getting inference to work in drug discovery is harder than it should be; workloads are bursty, the data is sensitive, compute costs are too high,” said Ashu Singhal, co-founder and President of Benchling. “We’ve been running Baseten internally for Benchling’s Model Hub and learned a lot about tailoring inference for drug discovery. Now we want customers to have the same access.”
Source: PRNewswire