Understanding the GenAI Tech Stack — Part 1: MLops
Last month, we discussed what every investor should know about the generative AI tech stack, identifying three core components — the Foundation Model, MLOps Model, and Application Model — as well as the opportunities and risks that are specific to each. In this month’s post, we dig deeper into machine learning operations (MLOps and LLMops) that are a combined category encompassing machine learning, development and operations that serve as quality control across a number of business functions.
MLOps is now particularly in the spotlight following the recent billion-dollar Databricks acquisition of MosaicML, an open-source startup with neural networks expertise that has built a platform for organizations to train large language models and deploy GenAI tools based on those models. This acquisition is part of a larger theme, as MLOps infrastructure vendors have been the biggest winners in the market to date.
That said, gone are the halcyon days of inexpensive capital favoring reckless investing; the new paradigm for investors requires a sharper assessment of companies focused on assessing intrinsic value — taking a hard look at a company’s revenue, growth, and margins — and long-term resilience.
What is driving MLops?
MLOps is an expansive category of businesses that encompass a variety of organizational infrastructure activities with a critical role in the tech stack, such as model selection, data preparation and training, model configuration, performance, and deployment, as well as maintenance and monitoring. Broadly, these companies address the operational challenges of deploying artificial intelligence (AI) and machine learning (ML) models with customers.
Despite recent enthusiasm for AI, MLOps is not new; it has helped deploy AI in enterprise use cases for years. While tech giants have a large stake in MLOps, smaller companies like Dataiku are building a reputation for user-friendliness. While application at scale may still be a long way off, recent advances in GenAI have created profound implications for MLOps businesses to cover the gap. The two primary forces driving this development are new use cases and a more acute need for tools:
ü New use cases
An influx of new market players and broad topic interest have added momentum to GenAI adoption in MLOps. GenAI “playgrounds” enable less sophisticated users to explore different models, which in turn leads to the creation of more prosumer use cases and. GenAI “Alignment” helps build confidence in the product, supporting better management of the safety of LLM models. New use cases will continue to evolve, driving further development.
ü More acute need of tools
The rise of AI models, specifically LLMs, has revealed numerous gaps in the current AI + ML infrastructure stack. Additional tools are needed to support the training, deployment, and performance management of models. Many of these needs relate to quality and availability of data, which is a core domain in which we see the creation and development of start-ups.
How should investors measure the intrinsic value and resilience of MLOps companies?
When we look at intrinsic value, we assess the quality of a company’s revenue, its growth rate and potential, as well as its margins. Across these metrics, MLOps has the potential to create tremendous value within the GenAI tech stack, specifically with respect to revenue and growth.
From a revenue model standpoint, MLOps follows a fairly standard, familiar SaaS business model with clear benchmarks. This is contrary to open-source models, where establishing revenue quality is much more challenging. Most MLOps businesses employ a fixed subscription model, where customers pay a consistent fee. Some combine this with usage-based ARR, dependent on the amount of data used. The combination of fixed fees and usage-based fees provides customers with flexibility and strong potential to scale. Service revenue can also play an important role in the overall revenue model, with an associated impact on the company’s margins. It’s important to keep in mind that the “humans in the loop” factor in AI + ML implies lower margins in these types of companies compared with pure SaaS players.
From a growth standpoint, MLOps has built-inpotential stemming from the overall explosion of GenAI. But simply establishing growth is not enough; investors must assess the quality of that growth. And growth quality depends on the target’s growth playbook. For example, winning companies in the MLOps space will be those that can successfully establish themselves as a full-service platform. This means they will be able to sell additional products into existing customers, which is generally much less expensive from a sales and marketing perspective than acquiring new customers. Successful platforming will be reflected in expansion at a logo level, with greater ARR per logo year-over-year as the same customer becomes a client for X number of tools. This kind of high-quality growth also has longer-term effects than upsells that come from price increases.
Existing players have already covered significant territory in this journey, which is especially critical ahead of an IPO. In Databricks’ case, acquiring MosaicML — which allowed Databricks to provide GenAI tooling alongside the company’ existing multicloud offerings — was a key business decision that leaves it advantageously positioned going forward.
The important takeaway here is that we anticipate a flurry of consolidation as MLOps players seek to build out their platform offerings. Many companies will not make it, especially considering the challenging fundraising environment. The most successful platforms will be those whose growth creates an ecosystem moat, with customers relying on their tooling.
Another avenue for growth will come from companies that prove their ability to demonstrate high stickiness, especially those that are associated with a particular niche. The ability to expand clients bases through new logo wins and keep existing customers happy is vital; customer churn in a niche environment is not an option with this playbook. This is particularly critical if the company’ growth playbook is not that of a platform. Companies in this category include startups building tools for new use cases that are too specific to become a broadly used platform. These companies’ products must be integrated into their customers’ day-to-day operations, proving stickiness.
Across both growth cases, startups will need to demonstrate their ability to scale and raise funds faster than competitors. They must be key players in their markets, with the best go-to-market plan, best-in-class sales teams, and the largest customers, etc.
Also, when we look at resilience, an important thing to consider is a company’s carbon footprint as a function of model configuration and the volume of treated data. This will become increasingly relevant as companies scale, treating more data and generating more use cases.
What does it mean?
Right now, investors are paying heavily for potential. We are still at the beginning of the genAI wave, and many of these loftily valued companies have little or no revenue, much less high-quality revenue. To put things in perspective, the $1.3B paid by Databricks for MosaicML equates to $21m per employee. In this environment, technical expertise and the ability to attract talent becomes paramount, from both an investor’s and acquirer’s perspective. Speculative investment strategies are no longer viable in the current capital environment, creating a need to assess MLOps companies on the basis of intrinsic value and resilience, with specific considerations and benchmarks for each.
