Underscore VC

The Scientific AI movement that we are catalyzing at TetraScience is bigger and more important than any one company. Either we all succeed together in replatforming science for the era of AI or we collectively fail. While we — led by me — can often be highly non-consensus, we are nonetheless authentically, extraordinarily, and passionately collaborative. I am eternally grateful for our partners at NVIDIA and Thermo Fisher Scientific and I love that the three of us are closely collaborating together as we move forward. #ScientificAI #BetterTogether

One of the clearest themes from NVIDIA’s Kimberly Powell at #JPM2026: The life sciences are moving from “AI experiments” to AI factories — repeatable systems that turn data + models + compute into production workflows. NVIDIA’s JPM updates reinforce that direction: BioNeMo is expanding the platform and ecosystem for AI-driven biology and drug discovery, and NVIDIA is building out the stack from models to lab infrastructure and enterprise adoption. Kimberly also called out TetraScience on-screen and from the stage as a trusted enterprise platform for scientific data and discovery. That’s an important point: in complex fields like the life sciences, the data layer isn’t plumbing — it’s the substrate that determines whether “AI” is a demo or an operating capability. The practical takeaway for R&D leaders: If you want agents and foundation models to be useful at scale, you need standardized, contextualized, governed scientific data and a discovery layer that makes it usable across teams, instruments, and time. Compute is accelerating quickly — the durable work is building the data foundation and the factory-like workflow delivery model. NVIDIA’s roundup of what they announced at JPM is worth reading for the full context. Link in comment. #ScientificAI #LifeSciences #BioNeMo #ScientificData #JPM2026

Humanoids are finally getting real, and last night’s 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗜𝗥𝗟 event drove that home for me. Three simple takeaways: 𝟭. 𝗧𝗵𝗲 𝗺𝗮𝗿𝗸𝗲𝘁 𝗶𝘀 𝗵𝘂𝗴𝗲, 𝗯𝘂𝘁 𝘁𝗵𝗲 𝘁𝗶𝗺𝗶𝗻𝗴 𝗶𝘀 𝗼𝗳𝗳. There’s a believable path to a ~$1T+ (you read that right, with a T!) humanoid market by 2050—but expecting most real-world use cases to mature in the next 3-5 years is extremely wishful thinking. We should be planning for a significant reset in excitement when people realize “they don’t work" and we go through a period of disillusionment with generalized robotics. In 2026-2027, expect this backlash as deployments hit messy reality and a lot of the well financed startups can't make the math work to keep their operations going. A lot of venture dollars are going to get flushed in this category but the people working at those companies -- those are who to keep track of, for the best companies will be founded in 5 years after all that expensive learning by those exact people as the technology improves...     𝟮. 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹-𝗳𝗶𝗿𝘀𝘁 𝗶𝘀 𝘁𝗵𝗲 𝗼𝗻𝗹𝘆 𝗰𝗿𝗲𝗱𝗶𝗯𝗹𝗲 𝘄𝗲𝗱𝗴𝗲. Companies could chase home or retail use cases, but many are deliberately starting in industrial environments where teams already deploy advanced automation and can tolerate imperfect-but-learning systems. The key is picking work where ~95% reliability still creates ROI instead of waiting for 99.9% perfection. We are betting on industrial application vs. in the home, primarily to the the complexity of varied environments and safety issues (i.e. do you want a robot helping clean up your kids toys with them in the room?)     𝟯. 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗮𝗻𝗱 𝗵𝗼𝗻𝗲𝘀𝘁𝘆 𝘄𝗶𝗹𝗹 𝗯𝗲 𝘁𝗵𝗲 𝗺𝗼𝗮𝘁. In the long run, the winners won’t be the flashiest demo videos—they’ll be the companies that do what they say they can do on-site, learn from every failure, and steadily improve 𝘳𝘦𝘭𝘪𝘢𝘣𝘪𝘭𝘪𝘵𝘺 and 𝘴𝘢𝘧𝘦𝘵𝘺. In a world where standards for humanoids don’t really exist yet, trust and execution are the real differentiators.     Fittingly for Inference IRL, the conversation wasn’t about any single datapoint or demo, but about how you connect market size, timelines, safety, reliability, and real customer workflows into a coherent view of where humanoids actually go from here. So appreciative of Aya Durbin, who shared her well‑earned perspective as the Humanoid Application Product Lead, Boston Dynamics, and brought exactly this discussion to life for the community at Underscore VC. Thank you Aya!

TetraScience this morning announced we'll be collaborating with Thermo Fisher Scientific to accelerate Scientific Data and AI enablement across the laboratory. Together, we'll be addressing THE fundamental industry constraint: Biopharmaceutical organizations face rising complexity and costs as they bring new therapies to market. Scientific AI can help—but only when it runs on harmonized, AI‑native scientific data. Through this collaboration: - TetraScience will integrate Thermo Fisher’s instrumentation and informatics with the Tetra Scientific Data Foundry and Scientific Use Case Factory, transforming experimental data and results into standardized, AI‑enabled scientific workflows. - This collaboration demonstrates how Thermo Fisher’s solutions integrate into an open, interoperable scientific data and AI operating system, the Tetra OS, reinforcing our commitment to remain fully vendor agnostic.   - Tetra AI, combined with Thermo Fisher’s digital and AI capabilities, will help power high‑value use cases that improve reproducibility, throughput, and scalability across R&D, manufacturing, and QC. The decision by our two organizations to collaborate reflects an acknowledgement that Scientific AI is an enterprise capability that requires an independent, AI-native scientific data and AI foundation. Initial efforts will focus on high‑value workflows at leading global biopharmaceutical organizations already using combined solutions from both companies. Link to the full announcement in comments below. #JPM2026