MLOps infrastructure is crowded because teams still lack agreed tools for many parts of the machine learning workflow.
2
Investors look for startups that can grow from a specific wedge into a platform, with a defensible product and a clear path to a large user group.
3
Open source can build adoption and trust, but founders should decide early what remains free and how the company may eventually make money.
Summary
Leigh Marie Braswell and Davis Treybig discuss machine learning companies from an investor's perspective. They describe MLOps infrastructure as crowded, with teams often stitching together six to eight tools and lacking established choices for orchestration, data preparation, monitoring, and collaboration. They expect consolidation, especially where products overlap across model orchestration, data quality, feature engineering, monitoring, and model response workflows. Both investors prefer startups with a clear wedge, a defensible product, and a credible path from a point solution to a broader platform. They also discuss the long tail of companies adopting machine learning. Products aimed at SQL users and business analysts may have a large market, though they need a different product and sales approach from open source infrastructure. On open source, Davis argues that companies need a monetization strategy even if they will not build paid features for years. The conversation ends with hiring. Founders need strong networks, an authentic story, and a reason for scarce technical talent to join their company.
MLOps infrastructure has many startups because teams lack agreed choices
Leigh Marie Braswell defines MLOps as putting complex machine learning, often deep learning, into production. She says relatively few companies urgently need deep learning, while the infrastructure market around it is already extremely crowded. Engineers face complicated production work and do not have consensus tools for several tasks, including active learning and collaboration with non-technical stakeholders. Davis Treybig adds that a typical machine learning stack may require six, seven, or eight tools, which leaves data scientists overwhelmed by the work of understanding and connecting them. He also says some categories, such as feature stores, remain poorly defined, so buyers may not even know what they need.
Model orchestration, data preparation, and monitoring may merge
Davis sees a strong case for consolidation around model orchestration. Teams often combine Airflow for orchestration with MLflow or Weights & Biases for experiment and model metadata, even though these activities belong to the same training and deployment workflow. He points to Metaflow as an example of a project bundling more of that work. Leigh Marie sees overlap between labeling, data quality, feature engineering, and tools that identify skew or gaps in datasets. She also expects more connection between monitoring and the actions that follow a detected problem, such as retraining, deciding whether a model can remain in production, or triggering a fail-safe.
The machine learning stack will remain different across use cases
Leigh Marie rejects the idea that one tool will cover every part of machine learning training and deployment. Autonomous vehicles need perception, planning, sensor handling, fleet management, debugging, and large tensor logs. Robotics has different operational needs, while healthcare brings privacy concerns. Davis compares this with the modern data stack, which may look unified through branding while its surrounding tools vary by company. He says upstream work such as data and feature engineering may be more common across machine learning use cases, but serving requirements diverge sharply. Some models need specialized latency, throughput, or large-model architectures, while others can run behind a simple Flask server.
Investors want a wedge that can grow into a defensible platform
Leigh Marie says a real pain point is not enough to justify an investment. Founders Fund looks for strong teams, a defensible product, and a moat created by technical insight or an accumulating data advantage. Davis focuses on how a company can stand out in a noisy market and become large over time. He gives Scale AI as an example of a company that began with API-driven labeling and expanded from that foothold. A startup might begin with edge-model deployment, a specific user group, or one acute problem for SQL-oriented teams, then expand into a wider platform. Both investors are more skeptical of products aimed at everyone without a clear initial user.
Long-tail machine learning customers need simpler products
Leigh Marie separates companies with mission-critical machine learning from the much larger group that is still experimenting. She is still trying to understand how quickly smaller companies will grow their spending, since some may eventually become major customers. Davis says products for the long tail usually need to operate at a higher abstraction level. He points to common retail and supply-chain needs such as inventory management and revenue forecasting, and mentions Continual as a SQL-first, declarative machine learning platform that sits on a data warehouse. Leigh Marie responds that these customers need a different product and sales motion from technical users adopting open source. She also questions whether some of them need machine learning or better analytics.
Open source requires an early plan for the paid business
Davis says open source helps infrastructure companies reach technical users, build community, and reduce fears about vendor lock-in. He is comfortable with founders postponing revenue, but wants them to decide what will remain free and what could belong in an enterprise tier. If too much is released and the community depends on it, pulling features back later may cause a backlash. Davis also describes cases where it can make more sense to build an adjacent enterprise product instead of an open-core version. He compares this approach with Vercel building a complementary business around Next.js. Leigh Marie agrees that founders should explore possible paths early, even though the first monetization strategy will probably change.
Leigh Marie says hiring is one of the first questions investors ask because machine learning infrastructure talent is scarce and highly paid. Early founders can draw on personal and investor networks. She recalls that many early Scale hires came from the MIT poker club, where the founders had met people with strong technical backgrounds. Later, companies can build a distinct community through a podcast, newsletter, hackathon, event, conference, active Twitter account, or a programming language associated with the company. Davis looks for founders who are already known in the field and can recruit well. He also argues that an authentic, unusual idea about the future of machine learning can give people a reason to join an infrastructure company.
Robotics adoption and long-tail platforms surprised the investors
Davis says the volume of new startups continues to exceed his expectations. He has also been surprised by the traction of companies focused on edge and robotics deployments, despite expecting the market to first solve basic cloud deployment problems. Leigh Marie says she once thought it would be very difficult to build a successful machine learning company for the long tail, but companies such as Weights & Biases and Hugging Face changed her view. She remains surprised that robotics has appeared in the world more slowly than she expected. The cost of hardware and uncertainty about return on investment may be part of the reason, although she also considers social acceptance.
"You may eventually realize, oh god, I put a little bit too much in the open source, but I can't pull it back now because my community will revolt and then I'm dead."Davis Treybig33:44
Who should watch
You are building an MLOps or machine learning infrastructure startup and need to sharpen its initial wedge, target customer, or expansion plan.
You are deciding whether open source adoption can lead to a viable company and want practical questions to ask before releasing too much functionality.
You are hiring for a technical startup and want ideas for building a community that attracts people who could start companies themselves.
