In recent years, investment in ‘pure-play’ machine learning (ML) has taken off.
Four facts stand out:
1. Twenty-five times more ML businesses in the UK raised Seed and Series A funds in 2017 than in 2013.
2. The median amount raised by these businesses increased significantly.
3. During the same period, the mean valuation of Seed and Series A pure-play ML businesses - enterprises that only develop ML solutions - increased at a compound annual growth rate (CAGR) of 16.6%.
4. Revenues for ML-as-a-service (MLaaS) are anticipated to grow at a CAGR of in excess of 40% through the next the five years.
These high-level numbers, although compelling, are simply the aggregate result of the answer to two main questions:
- How are ML businesses valuable?
- What is the source of their value?
This report provides a response to both.
By considering such insight, and its conclusions, those running ML businesses can adjust their strategy to maximise shareholder returns, and those investing in these enterprises can conduct commercial due diligence and negotiations with confidence. Lastly, potential victims of ML solutions can reflect on how their businesses and industries should respond.
A machine learning solution is valuable, first and foremost, because it has sufficient predictive power. Predictive power is the ability to anticipate future events and is the consequence of machine learning. Without it, there is no ML solution, and therefore nothing of value.
To create predictive power an ML business needs three key resources - the right people, adequate training data (from which an ML algorithm learns), and significant computing power - all focusing on applying the appropriate method.
Data scientists are the most important people in the process, and serve three primary purposes:
1. Selecting the appropriate ML method.
2. Establishing and improving predictive power.
3. Building the initial solution, based on data appropriate to the problem.
An ML solution becomes truly valuable when its predictive power exceeds 90%. Every gain made in the 90-100% bracket can be extremely challenging to achieve, but will serve to cement trustworthiness and uniqueness. Data scientists are responsible for this.
Two other groups of people are involved at the creation phase: the creator (usually the founder), and software engineers. The creator is mostly responsible for ensuring that the solution continues to match the problem.
Software engineers concentrate on creating the medium for delivering the solution - the platform, user experience (UX) and ability of the users to integrate/utilise its predictive power.
Reflecting the focus on people, the dominant investor in UK ML businesses at the Seed stage, Entrepreneur First, makes its investments about providing the right, high-quality individuals with the agency and time to create ML solutions.
The richness, maturity and predictive power of an ML solution cannot be easily replicated. Nor can capital invested in a competitor necessarily recreate it. Which is why predictive power is central to value; it’s what makes the solution unique, and why ongoing development to ensure consistent accuracy and relevance is key.
Although people are central to product creation, it is because of improvements in the availability of training data and computing power that ML solutions have proliferated over recent years. Data inputs from the Internet of Things (IoT) and the ever-increasing digitisation of life mean that 2.5 exabytes of data are now created daily, equivalent to 90 years of HD video. A further stimulus for solution development has been the release of ML best practice, example algorithms and modelling frameworks to the public by leading technology companies such as Google, Facebook, and IBM (the first to do so in 2015).
As the barriers have come down, pure-play ML businesses have become more common. No longer are solutions and expertise concentrated in the hands of a few major technology companies.
Potential money multiples in excess of 7x would not be an unusual expectation
To secure the necessary capital to create an ML solution, businesses typically raise funds through the sale of equity. Based on the limited available data, it would not be unusual for those that invest at the first Seed phase to expect a 7x return on their money. This expectation will influence how aggressive the business needs to be, and is also the expectations bar that the business will need to reach to secure funds in the first place. In simple terms, the product/market fit, future value opportunity, leadership team, and so on, must convince the investor that this level of return is possible.
Traditionally, private equity (PE) or venture capital (VC) funding will look for internal rates of return of 30% plus.
This infers that, if they made any early stage investment matching the 7x money multiple expectation, they would be willing to wait seven years until an exit. However, we would expect investment horizons to be shorter as the ML industry is fast moving - a CAGR for MLaaS revenues of in excess of 40% over the next five years is anticipated.
So it would be realistic to expect institutional investors to be comfortable with less than a 7x return in exchange for committing capital for between three and five years. In short, later phases of investment will be less patient, will expect returns sooner, and are likely to have lower money multiple expectations.
We are cautious about using discounted cash flows (DCFs) to calculate potential returns for ML businesses, as future cash flows are hard to predict and other valuation methods are significantly more appropriate. Our experience is that the ML investment community share this view.
1 Executive summary
- A machine learning solution is valuable, first and foremost, because it has sufficient predictive power
- Potential money multiples in excess of 7x would not be an unusual expectation
- The six benefits of ML solutions
- Cash and people requirements change markedly by the time an ML solution is commercialised
- An exit orientated strategy could help maximise returns and make the most of creators, data scientists and software engineers
- Creating value via machine learning
- 10 key lessons
3 Defining machine learning
3.1 Types of machine learning
3.2 Typical classes and uses of machine learning algorithms
3.3 Driving development and interest in machine learning
3.4 What is a machine learning business?
4 Employees are a key creator of value
4.1 Data scientists are the most critical for success
5 Machine learning trends and holes in the market
5.1 The upstream/downstream concept
5.2 General trends
- Contract values
- It is more common for machine learning solutions to be implemented where
- technical expertise requirements are low and training data availability is high
5.3 The six applications of machine learning
- Increasing resource yield
- Manufacturing efficiency
- Paper-based task automation
- Promoting services/products
- Service/product delivery
5.4 The application-benefit-value (ABV) nexus
5.5 Cash requirements for machine learning businesses
- Upstream - Increasing resource yield and improving manufacturing efficiency
- Midstream - Network optimisation
- Midstream - Paper-based task automation
- Downstream - Promoting and delivering services and products
5.6 The impact of analytics on the value of machine learning
- Poor utilisation of analytics (within industry) increases machine learning value
5.7 Machine learning autonomy during training and/or deployment
5.8 Why machine learning is valuable - the value ecosystem
5.9 The owner, their ethics and their strategy influence the value of a business
- The further upstream, the more valuable a partnership becomes
- An overview of machine learning business stages
- Commercial expertise becomes increasingly important over time
- Cash value changes, the best owner principle does not
- Shareholder expectations
5.10 What growth to expect: lessons from the analogous SaaS space
- Other lessons from SaaS
5.11 Other market trends: the shift towards generalised learning
6 Machine learning in 2017
6.1 Amazon bolsters cyber security with a machine learning twist
6.2 IBM aims to bring machine learning to the corporate masses
6.3 Google goes into direct competition with Amazon’s Echo speaker
6.4 Honda teams up with artificial intelligence research team
6.5 Apple spends big on data structuring
6.6 Spotify to improve recommendations with the help of Niland
6.7 Splunk consolidates data collection and machine learning capabilities
6.8 NVidia release machine learning-optimised hardware for ‘prosumers’ GDPR - a critical event Less data will make existing data more valuable ‘Right to erasure’ shouldn’t impact training or implementation
- The ‘right to explanation’ could undermine the core value of unsupervised and reinforcement machine learning
7 The valuation landscape
7.1 Fundraising in Seed to Series A stage businesses
7.2 Fundraising in post-Series A businesses
- Machine learning businesses focused on upstream applications have far higher valuations
7.3 Acquisitions in the machine learning industry
- Technology firms are the dominant bidder community
- The 7x money multiple is one reference point
7.4 GAFAM investment teardown
- Alphabet Inc.
- Apple Inc.
- Facebook Inc.
- Amadeus Capital
8 Driving the return on capital invested
8.1 Distribution and retention key to IP profits
8.2 Possible business models for machine learning delivery
8.3 Why churn rate is so important for machine learning businesses
8.4 Investor return expectations
8.5 What exit multiples should MLaaS aim for?
8.6 Not all contracts are created equal
8.7 The rise of the chief commercial officer
8.8 Success stories
8.9 War stories demonstrate the need to train adequately and ensure ML solution scalability
9 Employment landscape
9.1 Salary review
- Data scientists
- Software engineers
9.2 High demand is breeding more consultants
9.3 Salary is only a part of the deal
9.4 Supply and demand of employees
- Supply and demand of data scientists
- Supply and demand of data engineers
9.5 Solving the people shortage can have major implications for capital availability and the future agility of the business
- Software engineers come from a range of educational backgrounds
- Data scientists have a strong numerical background, in addition to an understanding of computer science
9.6 ‘Buzzword abuse’ makes candidate screening essential
10 Tax considerations for machine learning businesses
10.1 Intellectual property holding companies
10.2 How transfer pricing affects machine learning solutions
10.3 Licensing and withholding tax
- Indirect taxes
- R&D tax credits
10.4 Share incentive plans
10.5 Tax obligations facing ML businesses will be heavily scrutinised moving forward
- End notes
- Companies/organisations mentioned