Drew Angerer
Twilio (NYSE:TWLO) has been one of the hardest hit victims in software amid this year’s market rout. In addition to the macroeconomic overhang that has driven a market-wide correction, Twilio has also been facing internal operating inefficiencies and “mis-executions” on growth initiatives, causing further erosion of investors’ confidence in the stock.
The latest frenzy on OpenAI’s ChatGPT chatbot, which we have discussed in detail in a recent coverage, also threatens to upend Twilio’s of scaling its current market leadership in CPaaS (communications platform as a service) towards profitability. ChatGPT, which is built based on a “generative pre-trained transformer” (“GPT”) language model dubbed “GPT-3”, is capable of delivering far more complex output beyond foundational natural language processing (“NLP”) like speech recognition and natural language understanding deployed by Twilio.
And with the company’s latest revision of its longer-term growth and profitability targets already implying heightened vulnerability to “competition and mis-execution” of its business plan, OpenAI’s ChatGPT and continued innovations to improving state-of-the-art (“SOTA”) AI-powered language models serve as additional factors that could potentially push Twilio into obsolescence and derail the current CPaaS leader’s planned road to recovery.
Twilio’s Rise and Fall
The swift demographic transition within the labour force today, “driven by a wave of retirements” with the aging population dominating birth rates at a faster rate than “at any other time in history”, has made it increasingly difficult to hire within low-paying jobs across sectors such as healthcare, retail and food service. And Twilio’s portfolio of customer communications, customer data management, and customer engagement solutions have stepped up to the role of addressing said shortfall.
On one hand, consumers are usually satisfied with the fast response times and 24/7 availability aspect of automated communications solutions. Meanwhile, on the other hand, direct enterprise customers are favouring the cost efficiency aspects realizable via the automated solutions, as it eliminates the high demand for human labour while saving as much as 40% on annual service costs. And Twilio enables just that for both its enterprise customers and consumer end-users alike. Last year, Twilio capitalized on more than 50% of the $5+ billion CPaaS market, which is expected to expand rapidly towards a $60 billion opportunity by the end of the decade.
While the market trends point to favourable tailwinds for Twilio’s business model, they are also drawing greater competition to the industry. Competition becomes an even more gruelling challenge for Twilio when there is no weight on brand reputation among consumer end-users, which are often ignorant to whether they are “speaking with AI-driven assistants or humans”, as long as the matter at hand can be resolved in a convenient and efficient way:
Automating customer service is the new normal to the companies, as more than 50% of the customers do not care whether they are speaking with AI-driven assistants or humans unless their problems are solved. AI chatbots are gradually taking over the customer service game as they are filling the gap of the whole customer support system.
Source: “Customer Service Automation – Pros, Cons & Best Practices”, Trupp Global
The real caveat is that many customer service chatbots and automated communications solutions like those offered by Twilio today are not yet able to address more complex requests proposed by said consumer end-users, serving as nothing more than an automated FAQ generator that often needs to be redirected to a human representative. This accordingly docks a point off Twilio’s core customer communications solutions offered when being put against new innovations in the field, such as recent achievements in AI language models – which can be just as cost efficient for implementation by enterprise customers, and more effective for use in consumer-facing end-markets.
In addition to mounting macroeconomic headwinds that have roiled valuations this year, Twilio is also dealing with a sprawling list of shortfalls across its day-to-day operations, which paired with growing competition from both existing rivals and emerging innovations, paint a bleak outlook for its already battered valuation. Investors’ sentiment for the stock has largely soured as a result of rising uncertainties to the company’s trajectory to profitability, especially under the current market climate where the “growth-at-all-costs” narrative previously boasted across tech is no longer preferred.
Specifically, Twilio management has recently decided to curtail its multi-year annual organic growth target of at least 30% over the next three- to five-year horizon to the 15% to 25% range – a sharp cut that has increased investors’ wariness of internal execution inefficiencies beyond just “current volatility from the macro environment” which management had attributed the slashed guidance to. Management’s commitment to bringing the company to profitability is also showing some flaws. Instead of focusing on expanding its higher-margin (~70% range) software business to improve sales mix, the company has taken preference for expanding its largest, and faster growth but less profitable messaging business in communications. And to justify that shift, management has alluded to the segment’s increasing gross profit dollars, despite lagging gross margin percentage:
Okay. First, messaging. I know there’s a lot of discussion about gross margins and look, I get it. I acknowledge that historically we’ve guided to a 60%-plus gross margin in aggregate for Twilio and that we’d achieve that level by mixing more and more software into our revenue mix. But the telecommunications ecosystem and our business has changed since we set that goal. So I want to shed some light on what’s driving our recent decline in gross margins and why as a business we’re more focused on driving gross profit dollars.
Messaging is the biggest and fastest growing portion of our communications portfolio and has been the key contributor to our recent gross margin decline. Carriers charge a fee to access their customers and we pass that fee along both in the U.S. and internationally. And internationally, it’s generally higher. Those fees have gone up over time as this space has matured and taken off over the years…[and] we passed these through to customers at no additional margin, [hence, messaging] gross margin percentage has dropped. Furthermore, our international messaging traffic has seen strong growth, which has put further pressure on our gross margin percentage.
…but here’s the important part. Our gross profit per message has actually increased over the last few years and we believe that’s what really matters. So net-net, we believe that gross profit per message is a more meaningful signal for Twilio, even more so than gross margin percentage…So off of those aggregate gross profit dollars, we intend to continue driving leverage, which means that more and more of those dollars will drop to the bottom line over time and then drive the whole business with positive and increasing operating profit.
As much as management’s intuition for favouring gross margin dollar increase over gross margin percentage expansion seems like a sound strategy still, the anticipated results remain underwhelming when compared against the broader software industry. Specifically, management has guided a “medium-term operating margin framework [with] 100-300 bps of expansion annually”. Based on Twilio’s actual third quarter non-GAAP operating margin of -3.6%, the guidance would warrant an equivalent of 12% by 2027 at best. And management’s commitment to dialling share-based compensation levels down from the current 22% to the 15% to 20% range over the next three to five years is also disappointing considering the increasing urgency to not only achieve profitability and positive free cash flows, but also restore credibility on the stock from soured investors’ confidence.
Circling back to management’s decision to double-down on Twilio’s messaging business, the strategy also increases the company’s vulnerability to obsolescence. Messaging currently accounts for more than 60% of Twilio’s communications segment sales and more than 50% of consolidated revenue (the communications segment represents 84% of Twilio’s consolidated sales). Further expansion of this mix to drive Twilio’s profitability strategy would only dial up revenue concentration in one segment (i.e. messaging) that also happens to be facing the emerging threat of being replaced by better performing language models already available today over the longer-term.
The Threat of Next-Generation Language Models
Language models in AI are essentially a type of transformer models capable of learning and improving from a base data set over time to enable human-like speech and understanding output, like “translating text and speech in near real-time, [and/or] making online recommendations”. A basic iteration of transformer models used commonly in our daily life settings include Google’s “BERT” (Bidirectional Encoder Representations from Transformers), which is deployed in Google Search (GOOG / GOOGL) to help the search engine better understand queries. Transformer models like BERT can also perform “sentiment analysis” by combing through data like emails to gauge opinion and emotion – a NLP feature largely found across customer relationship management tools like Microsoft’s “Viva Sales” embedded in Dynamics 365 and Salesforce’s (CRM) “Social Studio” offered as part of its Marketing Cloud solutions. Twilio’s messaging solutions, like the APIs used by vendors to enable customized customer service chatbots, is also a transformer that can learn, understand, and process queries.
But the capability of language models have moved on by large margins from what we are most familiar with today in the day-to-day setting. OpenAI’s ChatGPT bot may be talk of the town in recent weeks, but its underlying language model, GPT-3, is not the only SOTA AI algorithm of its kind on the market today. Remember a few months ago when Google faced the whole debacle on whether its chatbot is sentient? Yea, that was another language model, known as LaMDA (Language Model of Dialogue Applications).
And ChatGPT is impressive. Consisting of 175 billion parameters, the underlying GPT-3 language model on which OpenAI’s latest chatbot is built on is more than 100x better-performing than its predecessor, “GPT-2”, and 10x better-performing than Microsoft’s (MSFT) “Turing NLG” language model introduced in 2020. And over the past week-and-a-half, we have learned that the newest chatbot in town can do a lot more than spitting out search results like Google Search or FAQ-based responses like Twilio API-enabled automated customer service reps.
The next-generation model has been pre-trained to comb through its base data to generate query-specific responses that can range from simple fact-based answers (i.e. Q&A style query and response) to well-versed opinion-based essays and poems with personality. This capability can easily address the shortfall experienced by automated customer service chatbots available today – let’s face it…we’ve all been frustrated with generic chatbot answers that often need to be redirected to a human representative, not to mention the eventually long wait times which defeat the purpose of automation. Next-generation language models like GPT-3 can also be deployed in simpler use cases, such as enabling “relevant, lightning-fast semantic search”, which can potentially address the headache we face when trying to search for solutions in typically overwhelming FAQ / troubleshooting pages, say delivery and return policies, especially with accelerating e-commerce adoption.
What this essentially means is that there is now a very real possibility for Twilio’s core messaging offering – on which its medium- to longer-term business plan largely depends on – to become disrupted, or worse, replaced over the same period. From a consumer end-user’s perspective, next-generation language models like GPT-3 could enable a better experience compared to online chatbots powered by Twilio and employed by vendors today. Meanwhile, from an enterprise customer’s perspective, the implementation of similar next-generation language model APIs to enable performance of “a wide variety of natural language tasks” can be just as efficient and scalable as what Twilio already offers today, not to mention the improved customer service experience delivered.
As impressive as ChatGPT is today (discussed in detail here) though, there are still shortfalls that researchers and engineers are still trying to address. But refinements to the model that are already well underway could easily derail Twilio’s medium-term growth plans. As discussed in our recent analysis on OpenAI’s ChatGPT chatbot, significant improvements have already been made over a short span of about two years since the closed beta release of its predecessor, “InstructGPT”:
ChatGPT has improved significantly from the GPT-3 API closed beta launched in 2020, which only few had access to. Just merely two years ago, the language model did not know how to reject nonsense questions or say “I don’t know”. Fast-forward to today, not only can the model reject nonsense questions and inappropriate requests, ChatGPT can also make suggestions spanning simple one-liner responses to prose-form essays.
Source: “OpenAI Impact Analysis: Microsoft, Google and Nvidia”
Now think about how much more can be improved from ChatGPT today over the span of the next three to five years – the same timeline on which Twilio seeks to expand its automated messaging business. This is further corroborated by OpenAI’s development of “WebGPT”, an enhanced GPT-3 model, that is already well underway. WebGPT has been trained to comb through data available on the internet in real-time to generate more accurate responses, addressing the GPT-3 model’s current limitation of being pre-trained by data dated only up to 2021:
ChatGPT has been trained on millions of websites to glean not only the skill of holding a humanlike conversation, but information itself, so long as it was published on the internet before late 2021.
Source: Bloomberg
Browsing the live web allows the model to give up-to-date answers, even though GPT-3 was trained before 2021.
Source: OpenAI
WebGPT can also cite sources in its response, addressing concerns over the rate of accuracy in current responses that ChatGPT spits out. Meanwhile, researchers and engineers are still trying to better refine the capability, such that the model could comb through and “cherry-pick” sources that are most reliable and accurate:
Even though we discourage it, the model sometimes quotes from unreliable sources. This is a pitfall we hope to mitigate with adversarial training…Eventually, having models cite their sources will not be enough to evaluate factual accuracy. A sufficiently capable model would cherry-pick sources it expects humans to find convincing, even if they do not reflect a fair assessment of the evidence. There are already signs of this happening. We hope to mitigate this using methods like debate.
Source: OpenAI
While Twilio remains fixed on expanding its messaging business as a key strategy to shoring up growth and profit margins over the medium-term, there appears to be little focus on the R&D aspect of its technologies to better take on competition coming from both existing competitors (e.g. NICE, FIVN, etc.) or emerging innovations like OpenAI’s GPT-3 developments in the years ahead. This drives worries that the company might be too focused on monetizing its current market leadership in the segment and overlooking competition stemming from future disruptors, which makes it difficult to defend any hopes of a near-term comeback for the stock.
We are investing strategically into our core CPaaS products to improve margins and maintain our market leadership position. To reduce costs, we’re adding additional self-service capabilities and modernizing our infrastructure.
Risks to Twilio’s Bear Case
Twilio’s recent acquisition of Segment, a warehouse of first party data collected across “websites, applications, and even data warehouse”, could be a potential solution to improving its current NLP APIs and drive better user experiences for end-customers of its automated communications solutions. However, training its existing NLP models with the newly acquired data will take time and will be a capital-intensive endeavour. Twilio currently strives to provide its enterprise customers with access to the data via a flexible API that can help them train their respective communications solutions, like customer service chatbots, instead:
We are focusing on speeding up development in Segment, Flex and Engage where we can deliver differentiated solutions by combining our data foundation and communication channels…
Adding Segment provides deep insights about a customers’ journey and preferences to better use our messaging products or Flex when engaging with users.
If you look at just our messaging customer base, the use cases fall into roughly three categories, marketing, verification, and conversational sales support. In the marketing use case, a customer might use Twilio’s messaging products to promote special offers to existing customers, but without the right data to build their target profiles and map their customer journeys, our customer can’t consistently deliver the right message to their customer on the right channel at the right times.
That’s where Segment comes in. And if the customer adds Engage on top of Segment, they can not only trigger the right interactions in real time, but they also know when not to engage with a customer.
Yet, two years following the transaction’s close, Twilio continues to face mounting inefficiencies on the integration of Segment, which puts any hopes of bolstering performance of its existing NLP models to better compete against next-generation language models like GPT-3 further out into the future. Not to mention, Twilio’s “Flex” contact center solution on which it hopes to cross-sell Segment solutions continues to struggle from mediocre market share gains despite “being in market for nearly 5 years”. The business currently represents only 22% of Twilio’s product sales and a nominal contribution to consolidated revenue (the software segment represents only 11% of consolidated sales), a far cry from the strength observed in messaging. This continues to underscore the challenges ahead for Twilio on not only building / sustaining its current market leadership in CPaaS, CCaaS (contact center as a service), and CDP (customer data platform), but also matching the incoming prowess of disruptive technologies in its field.
Alternatively, there is potential for Twilio to integrate the GPT-3 API in the future instead to bolster its solutions’ performance. It would be a more cost-effective way to address competition arising from emerging innovation, despite potentially enabling a “revenue-sharing” model with competing companies like OpenAI, thus ultimately diluting Twilio’s market share.
Final Thoughts
Twilio currently trades at 1.5x forward EV/sales on consensus CY/2023 revenue growth estimate of 17% versus the SaaS peer group mean of 6.4x forward EV/sales on consensus CY/2023 revenue growth estimate of 18%. While Twilio’s valuation today after an 80%+ selloff this year looks relatively attractive to peers with a similar growth profile, significant uncertainties remain on whether it can even achieve the aspirations outlined in its latest medium-term business growth plan, let alone weather out existing and emerging competition. With ChatGPT unleashing new expectations for next-generation language models that far exceed what Twilio is capable of today, the company’s recovery prospects likely face a bumpier ride ahead.
Editor’s Note: This article discusses one or more securities that do not trade on a major U.S. exchange. Please be aware of the risks associated with these stocks.
Be the first to comment