# Venture Capital in Machine Learning Startups

John Spindler, Capital Enterprise | MLOps Meetup | Episode 16 | 1:05:41
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=v9OvXxTUBtg
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/venture-capital-in-machine-learning-startups
Published: 2020-06-06
Tags: finance, monitoring, startups

## TL;DR
- Machine learning startups face technology and product risks that generalist investors often do not understand, so John Spindler evaluates whether the team can build a model that works better than a rules-based system.
- Early-stage investors have to judge companies with little data by assessing execution, market, product and technology, business model, and their own investment risk.
- A machine learning company must move from experimental technology to a usable product, then build a business around the data and market it can reach.

## Summary
John Spindler explains how he evaluates machine learning startups as an investor at Capital Enterprise and AI Seed. He compares machine learning deals with ordinary software investments, where market risk often matters most. In machine learning, the team may still have to prove that the technology works, that the data can be captured and structured, and that the model performs on live data. He assesses execution, market, product and technology, business model, and investor risk. He also describes the stages from an idea through commitment, a first signal, data and product-market fit, and scale. Spindler is candid about failures, including passing on Revolut and backing a data-annotation team that returned its funding after finding the problem too difficult. He argues that companies should find specific problems where machine learning creates value beyond rules-based software, then turn their data and models into products that customers will keep using.

## Key ideas
### Machine learning investing requires a different risk assessment from ordinary software investing
[08:38](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=518s)
Spindler says generalist investors often assess a machine learning company like any other startup. They ask how quickly it can reach market and gain traction, while paying little attention to what is under the hood. He created AI Seed because these companies have a different risk profile. In ordinary software, market risk may dominate. In machine learning, product and technology risk can remain unresolved for years. The team must show that it can build a model that works consistently, use messy real-world data, and create something better than a rules-based alternative. The investment decision therefore depends on technical evidence as well as customer demand.

### Investment skill comes from repeated experimentation and learning from failure
[05:08](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=308s)
Spindler describes becoming an angel investor after receiving the final payment from an earlier business exit. He invested nearly £200,000 across eight companies and says six died, one produced a small exit, and one remained without a clear outcome. He contrasts those results with the flat he and his partner considered buying in London, whose price later rose substantially. His conclusion is that investment involves luck, opportunity, experimentation, and accumulated experience. The same applies to entrepreneurship. Being in the right place at the right time can help, but perseverance and learning from mistakes are what improve judgment over time.

### Early-stage investors must judge companies before the usual business data exists
[23:41](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=1421s)
Spindler says investors usually examine five risks, though they weight them differently. Execution risk asks whether the team is credible, capable, and able to carry out the plan. Market risk asks whether customers have a real problem and whether the company can reach them more effectively than competitors or substitutes. Product and technology risk asks whether the team can build a system that works and gives a meaningful performance gain over rules-based software. Business risk asks whether the company can make money at scale. Investor risk covers how the fund can make money and whether the expected return fits the risk.

### A startup moves through stages from an idea to a company that can scale
[28:38](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=1718s)
Spindler describes five stages. An idea becomes commitment when the founders start building, telling people about it, or forming a team. The next step is a first signal from customers or technology that the idea may work. After that comes data and product-market fit, when there is evidence that the product works, customers value it, and the business could make money at scale. The final stage is scaling and dominating a market. Investors often fund the transition between stages. Seed investors move a company from commitment to its first signal, or from an early signal toward usable data and product-market fit.

### A machine learning product must prove value beyond a rules-based system
[30:45](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=1845s)
Spindler says his investment thesis focused on teams that could build strong technology and solve difficult engineering problems. For a machine learning startup, the question is whether the model works better than a rules-based approach. A model can be analytical, predictive, prescriptive, or used for automation, but it must achieve a useful result for the customer. He also warns that a complex model is not automatically better. If linear regression solves the problem, it may be easier to explain and maintain than a neural network. The product's value has to justify the added cost and complexity.

### Moving from a laboratory model to a working product takes longer and costs more
[35:07](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=2107s)
Spindler compares a machine learning startup's early technology to a tadpole that must become a frog. The first task may take one or two years, or longer, because the team has to make the technology work, productize it, and then pursue customers. Machine learning products also require more than ordinary software iteration. Teams must deal with real-world data, model performance, data capture, and ongoing maintenance. Compute can be cheap through early cloud credits, but costs rise sharply at scale. Customers may also demand lower latency, which can require more expensive GPUs and compute.

### Niche data problems can give a startup room to build a business
[53:40](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=3220s)
Spindler describes investing in Observe, which uses computer vision in murky water to count fish and predict their weight, helping fish farms manage feed and health. He also mentions a company using infrared cameras to count and predict the health of chickens in large sheds. These companies begin with difficult, specific problems and data sources that already exist. He does not expect every such company to become a billion-dollar business. He does expect a company to become good if it can solve a problem that the industry cannot easily solve itself. The data can then support further products beyond the initial tool.

### Model monitoring matters because live data and conditions change
[1:02:56](https://www.youtube.com/watch?v=v9OvXxTUBtg&t=3776s)
Near the end, Spindler discusses the problem of models becoming unreliable after deployment. A model trained on one data distribution may behave differently when the live data changes, especially during an event such as COVID-19. Teams need to ask whether the model still does what it was trained to do and understand the trade-offs between false positives and false negatives. He also agrees that human involvement will remain useful. Machine learning systems can become bespoke projects that depend on one dataset and one customer, which makes them difficult to maintain. He describes this as software that can go wild after deployment.

## Notable quotes
- John Spindler: "The one big fact about machine learning and AI which is often overlooked in all the press is it is bloody difficult." (17:43)
- John Spindler: "If you can get linear regression working, yes, it is simple, it is explainable, it is a lot easier to do." (18:07)
- John Spindler: "The biggest hurdle to build them are even worse to maintain. It is feral software, it is software, it goes wild." (49:22)
- John Spindler: "The real important stuff with that is, can your model then work on live data?" (1:00:31)

## Tools & references mentioned
- Capital Enterprise
- AI Seed
- Entrepreneur First
- Techstars
- London Co-Investment Fund
- Revolut
- DeepMind
- Google
- Imperial College London
- Observe
- Palantir
- COVID-19

## Who should watch
- You are building a machine learning startup and need to explain why the technology creates value beyond a rules-based product.
- You invest at seed stage and want a practical framework for judging teams, data, market risk, and technical risk before there is much evidence.
- You are taking a model from research into production and want to understand the cost, data, monitoring, and maintenance problems that appear after deployment.

## Related talks

- [Investing in MLOps](https://mlopstalks.com/talks/investing-in-mlops) (Leigh Marie Braswell, Founders Fund & Davis Treybig, Innovation Endeavors, 48:52)
- [ML Stepping Stones: Challenges & Opportunities for Companies](https://mlopstalks.com/talks/ml-stepping-stones-challenges-opportunities-for-companies) (John Crousse, 47:47)
- [ML Unicorn Start-up Investor Tells-IT-All](https://mlopstalks.com/talks/ml-unicorn-start-up-investor-tells-it-all) (George Mathew, Insight Partners, 51:01)
- [LLM Valley](https://mlopstalks.com/talks/llm-valley) (Lilly Chen, Contenda, 11:57)
- [Managing Machine Learning Projects](https://mlopstalks.com/talks/managing-machine-learning-projects) (Simon Thompson, GFT, 45:02)