Appen (ASX:APX) is one of Australia’s high-flying tech stocks – which collectively go under the WAAAX acronym. Its share price has almost tripled over the past year as many investors ride the Oz tech story. But to my mind, the company’s valuation looks extremely stretched.
According to its own description, Appen is a global leader in providing human-annotated training data for machine learning applications. With its share price up more than 3,600 per cent in a little over four years, it’s worth exploring whether the gains are likely to continue.
An investment in Appen is pitched as investing in high tech and the continued growth of Artificial Intelligence and machine learning across a spectrum of applications. At this point however these applications are dominated by social media ad and newsfeed relevance, and search relevance.
A little background here may help
Since 2007 one of the big problems in machine learning has been access to the training data needed to be fed into the machines. And since about 2013 there has been an explosion of image labeling requests. In order for a search engine to deliver an image of a shoe when you ask for it, the robot needs to know what a shoe is, and to recognize a shoe, the machine needs to be taught. The way to teach a machine it to feed it data sets that have the images with shoes differentiated with tags.
As you can imagine it’s a monumental task because you may also want to tag ‘leather’ or ‘yellow’, or ‘adidas’, or ‘high heels.’ And that’s just the tip of the iceberg for shoes. What about everything else!
You can see the opportunity
Appen has two divisions. Content Relevance provides the annotated data for search and social media, while Language Resources provides speech data. The company essentially offers a cent or two per tag to more than a million crowd-sourced, low-paid and home-based workers in 130 countries but generally throughout Asia.
It’s really a low-tech business feeding the machines of high-tech customers like Google, Facebook and Microsoft with datasets for their algorithms to learn from.
In recognizing Appen’s workforce as a leader in providing high-quality data relative to competitors, it is important to acknowledge the 2017 acquisition, by Appen, of Leapforce for US$80 million, which added 800,000 at-home workers to Appen’s 400,000, as well as 50 per cent of Appen’s 2018 revenues.
The runway for further revenue growth currently appears very long and projections for growth of AI, automation and machine learning are exponential but unless an analyst has a MS in Computer Science, quantifying the opportunity is almost impossible – and it’s a guess even with a MS in Computer Science.
And despite having a strong position in the markets it serves, the core Appen business model appears to have commodity-like elements. By that we mean to suggest that their arbitrage on manually-intensive labour is susceptible to a decline in unit economics either as competitors emerge or as technology improves.
There is little doubt that huge investment dollars are being ploughed into the opportunity by big-tech companies so the need for greater volumes of higher quality training data is a trend that appears somewhat reliable. For now, at least. If deep learning – machines teaching themselves rather than requiring human annotated datasets – arrives, the trend could end abruptly.
Big Tech is, without doubt, working on the problem (and quite frankly, I don’t know how AI will evolve). But the question they are trying to answer is; why pay for millions of annotated images if humans might be able to manually annotate just a few hundred and the machine does the rest itself? Or what if we didn’t need the humans at all?
Appen’s acquisition of Figure8
And that brings us to Appen’s acquisition of Figure8. Appen recently announced the intention to purchase Figure 8 for its “Leading Tech Platform.” That’s good to hear because the Figure8 lost US$12 million at the EBITDA level in FY18 and might have required a capital injection within six months based on its cash burn rate. Moreover, the large jump in revenue between 2016 and 2017 is due almost entirely to one large government contract.
Despite these reservations, Appen paid US$175 million upfront, which is circa 5.7 times 2018 revenue, and will pay up to another US$125 million in March 2020 on outperformance of incremental 2019 subscription revenue targets.
To pay for the acquisition, Appen conducted a placement to raise A$285 million and has entered arrangements for new debt facilities of US$125 million expected to be drawn down in 2020 for the ‘earn-out’.
Figure 8’s tech speeds up the annotation process and presumably lowers its costs, through prediction.
But the founder of Figure8 stepped down as CEO some time ago to start a second machine learning company and in a blog post written in May 2018, wrote that he had “spent the last six months heads down building a new machine learning tool called Weights & Biases with my longtime co-founder, my new cofounder and friend and brave early users at OpenAI, Toyota Research, Uber and others.”
Clearly AI and how datasets are produced is evolving rapidly, which makes sensible investing incredibly challenging simply because the earnings profile lacks visibility.
When examining most broker reports covering Appen it appears their growth assumptions are merely an extrapolation of recent compounded average annual growth rates. At circa $25, the share price implies revenue growth of 15-20 per cent compounded for at least five years with no diminution in per unit margins. Nobody can have clarity on whether this is reasonable so we must conclude that much of the company’s share price appreciation is due to momentum.
Appen currently trades on 6.2 times FY19 revenue and 5 times projected FY20 revenue. This equates to 36 times FY19 EV/EBITDA and 27 times FY20. They are ambitious numbers in anyone’s language.
Between August 2018 and February 2019, the company upgraded its EBITDA by US$13-20 million. But in that time, the share price has seen the market capitalisation increase by A$1.8 billion. If you make the Figure8 acquisition valuation ‘neutral’, the market has effectively valued the US$13-$20 million earnings increase A$1.55 billion.
While it’s entirely possible the combined Appen/F8 entity is more valuable in the medium to long term, it’s difficult to reconcile the A$1.55 billion re-rate of earnings and the projected near-term losses for the F8 acquisition.
That doesn’t seem to be a problem for the US-based Private Equity owners of Appen’s US-based peer Lionbridge. According to recent reports, Morgan Stanley, UBS and Taylor Collison have been appointed to list Lionbridge on the ASX of all places. I wonder why?
Momentum is an incredibly rewarding investment strategy until it’s not.
Good read Roger. They're calling Straker Translations the next Appen now so who knows what momentum will do to that small cap.
Thanks Roger. Did you attend their tech demonstration? It was quite worth the while as it helps understanding how Deep Learning works and where the importance of human annotated data lies. Sure, Deep Learning with minimum human help could happen (thinking the Terminator...) 5, 10, 20 years later, but same as this current phase of Deep Learning and it's adoption in the tech and commercial world, it will be a phased process . Essentially plenty of time for investors and indeed the company itself to re-position. So should one pass a KNOWN strongly growing company for the fear of that down slope (not a cliff in my opinion) in an still UNKNOWN future? worth a thought... not a short...
Thank You very much for your article, Roger. I have just learnt more about my stock! And thank you for recommending Altium years ago (another story on Livewire, I think). I bought them at $4.74 thinking they were expensive! Still both holds for me.