Aditya Naganath thinks many MLOps companies struggled because enterprises built narrow, customized tooling in-house.
2
AI infrastructure is a separate layer from LLM Ops because every AI workload needs compute, while LLM Ops tools are optional abstractions around models.
3
The large opportunity is in AI coworkers that can reliably reason through useful units of knowledge work, since those products could be priced against labor spend.
Summary
Aditya Naganath discusses how he evaluates AI companies and why many earlier MLOps startups failed to become large businesses. He looks for a novel approach, thoughtful product execution, and a credible go-to-market advantage. He sees data platforms as larger markets because enterprises broadly need data storage and compute, while many MLOps tools fit narrow workflows that companies can customize internally. He separates AI infrastructure from LLM Ops, arguing that compute providers benefit from unavoidable demand for training and inference, while evaluation and orchestration tools still face build-versus-buy pressure. At the application layer, he is most positive about coding assistants and enterprise search. He expects the next wave to involve AI coworkers that can perform whole units of work, though this requires more reliable reasoning. He also discusses data center inefficiency, GPU reliability, networking, and why inference may become much larger than training.
Investors need a specific reason a startup can rise above similar teams
Aditya Naganath says an interesting idea often attracts 10 or 12 solid teams, so team quality alone does not settle the investment decision. He wants to see a novel approach and a clear explanation of why the company can win despite the competition. That reasoning should reach into the product's foundations and its go-to-market plan. Distribution can help, especially when founders already have a community or strong following, but Aditya warns that fast early revenue can hide weak foundations. Customer churn and fading growth expose companies that have not executed thoughtfully.
Many MLOps startups owned only a narrow, customizable part of the workflow
Aditya says MLOps companies need to prove that buyers receive enough value to purchase their product instead of building it internally. He points to model A/B testing, model evaluation, and feature engineering as areas where enterprise teams often created tailored solutions themselves. Even when a tool was adopted, it usually captured a small portion of the budget because it fit one narrow slice of the workflow. Scale AI found a stronger position by handling data labeling, work that customers did not want to do in-house. Aditya contrasts this with data platforms, which benefit from the broad growth of enterprise data.
AI infrastructure has unavoidable demand, while LLM Ops still faces build-versus-buy pressure
Aditya divides the new infrastructure market into LLM Ops and core infrastructure for compute, storage, and networking. Training and inference create demand for GPUs, so companies that serve compute at lower cost and latency can monetize an unavoidable part of the stack. He names Together AI as an example of software that improves utilization of large GPU clusters, and Vast Data as a company working close to the data center and training-data layer. He remains cautious about evaluation and data-curation tools because many products look similar and enterprises may build them for their own needs.
Coding assistants and enterprise search are early application markets with clear buyers
Aditya calls software development an obvious first application for language models. GitHub Copilot helped educate the market, and coding assistants can be useful even when suggestions are wrong because a developer can ignore an incorrect completion. That gives the product a tolerance for error alongside a productivity benefit. He also names enterprise search as a strong market. Glean used existing connectors to applications and data sources to build a privacy-aware assistant over company data. The difficulty is greater than searching a few PDFs because access permissions and changing enterprise data must be handled across many sources.
Language models let non-technical workers build bespoke workflows, though production is harder
Aditya describes LLMs as a higher level of abstraction that puts more capability into model weights. Sales, HR, and marketing workers can therefore build custom applications and workflows without writing code or asking a developer for help. Demetrios Brinkmann points out the gap between a low-code prototype used by a small team and a production feature inside a larger application. Aditya's answer does not remove that gap. His point is that the models let non-technical users get much farther before they need traditional software development.
AI coworkers require reliable reasoning and could be priced against labor spend
Aditya distinguishes current copilots from future coworkers by the amount of work the system can perform independently. He wants models to reason more reliably, reflect on their outputs, and prune a search space intelligently. He mentions reinforcement learning and Monte Carlo research, with AlphaZero's chess performance as an example of techniques worth considering. If these systems can perform useful knowledge work, products could be priced as a fraction of wage spend. Aditya says today's successful AI applications are still SaaS businesses with familiar multiples, so much larger economic value depends on reaching the coworker model.
Inference may become a larger infrastructure market than training
Aditya says GPU clusters remain expensive because training is inefficient, networks are poorly suited to these workloads, and GPUs can fail during runs spread across thousands of devices. He sees networking as interesting because hyperscalers provision separate networks for LLM training, yet current hardware and network fabrics leave room for improvement. Sales would be difficult because the customer set is narrow and established companies are active. He still thinks concentration can be acceptable if a few customers pay large amounts. Inference is especially attractive because every AI application will need it, making the workload potentially much larger than training.
The next investment target is an agent that automates a complete unit of work
When Demetrios asks what kinds of companies should contact him, Aditya points to agents that can automate useful work rather than provide another narrow assistant. He gives security operations analysts, business process outsourcing workers, and site reliability engineers as examples. These applications need to reason reliably across large, multimodal data sets. The investment case depends on both technical difficulty and a tangible value proposition. Aditya is interested in teams pushing the frontier of agent reasoning where the resulting product could own an entire workflow.
"The whole dream for this AI movement is can you get these things to be able to reason well enough such that they can resemble a human being and as a result be a coworker versus a copilot."Aditya Naganath27:20
Who should watch
You are deciding whether an MLOps or LLM Ops product can become a broad enterprise business rather than a narrow tool.
You are building AI infrastructure and need to think through GPU utilization, networking, reliability, customer concentration, and inference economics.
You are exploring agents and want a candid view of the technical bar for automating a complete unit of knowledge work.
