Understanding the GenAI Tech Stack : Part 3 — Foundation Models
Make no mistake, GenAI is still very hot and continues to evolve at an astounding pace. Despite the critics and the collective hysteria on the topic — fueling the question of whether or not this is hype — GenAI has already had a more successful debut than software-as-a-service (SaaS), with over 1B$ in start-up revenue alone. In June, we delved into essential knowledge for investors regarding the generative AI tech stack. We highlighted three fundamental elements within it — the Foundation Model, MLOps Model, and Application Model — including the unique opportunities and risks associated with each of these components. This month, we will explore Foundation Models in greater detail.
These businesses employ AI systems trained on extensive sets of unlabeled data, allowing for versatile adaptation to various tasks with minimal fine-tuning. One of the most publicly visible examples of this category is Microsoft-backed Chat GPT; while GenAI had been in development for years, the release of Chat GPT 3.5 last fall represented a substantial leap in innovation and quality. Chat GPT has quickly become the fastest-growing application.
Understanding Foundation Models
Foundation models — as the name suggests — serve as the current foundation of GenAI development and technology. They process natural language patterns at scale to comprehend and generate human-like text, making them invaluable for applications like natural language understanding, conversation generation, and content generation.
Foundation Models in text (GPT-4, Claude, Cohere), image (Midjourney, Stable Diffusion), Video (Stable Diffusion) and audio (Bark, ElevenLabs) serve as the entry point for most businesses into the realm of generative AI. Consequently, the market potential for these models is immense.
Currently, Foundation Models are primarily being developed by tech giants like Google, Meta, and Microsoft, or by highly funded private companies such as OpenAI (which received $13B funding from Microsoft in April), Anthropic or Inflection. Anthropic is currently in the process of raising an estimated $4 billion of venture funding from Amazon.com to accelerate the development of the company’s future foundation models and make them widely accessible to AWS customers, thus reinforcing the weight of tech giants. This presents a formidable barrier to entry for smaller AI companies, though there are other LLMs available such as those built by Aleph Alpha, Cohere, AI 21 Labs or the one just released by French start-up Mistral.ai
There are three core pillars of the Foundation Model tech stack: open source, closed source, and platforms + hubs. Using a Hybrid Growth Diligence (HGD) lens, this article will explore the revenue, growth, and margin quality aspects of each model, with the understanding that all of these are brand new technologies with many unknown variables.
Closed-Source vs. Open-Source
Closed-source models are currently the most favored by investors (just looking at where fundraising has gone to date) because they have the clearest value proposition; these are models that have been trained on large datasets and that need extensive compute capacity. The underlying source code and architecture are proprietary and not openly accessible to the public or the broader developer community. That said, a distinction should be made between fully closed models (e.g. PaLM (Google), Gopher (DeepMind)) and ones gated to the public through cloud-based/API access (e.g. GPT-4 (Open AI).
These companies have typically invested heavily in their technology and retain strict control over their code. There are some differences nonetheless; for example, Open.ai’s GPT-4 is meant to be a general model that is used as-is, while the Cohere model will be augmented with additional training data from the user. While closed-source models may offer advantages in terms of intellectual property protection and controlled quality, they can limit transparency, hinder collaborative innovation, and raise questions about data privacy and dependency on specific data providers.
Unlike closed-source models, open-source models such as Llama 2 by Meta, Stability AI, Mistral AI rely on a public community of developers collectively, transparently building on a shared foundation. Benefits to open-source models include greater knowledge-sharing, and improved innovation and agility.
From an investor standpoint, these companies allow investors to stay at the forefront of AI advancements while minimizing the risks associated with closed, proprietary solutions. Open-source AI models often have lower upfront costs (some are trained on smaller datasets) and can be customized to meet specific needs, making them an attractive choice for investors seeking flexibility and long-term scalability. That said, monetizing these models is a considerable challenge; success relies on the model’s ability to attract subscription interest because it is considered the default tool for its purpose. Reaching that “default tool” status means demonstrating a strong ability attract a wide user base, as well as a wide community of developers. Additionally, open-source projects generate revenue on enterprise support, hosting cloud and other indirect sources of revenue.
Key questions to understanding Foundation Models’ intrinsic value are common to both closed-source and open-source models, though the answers strongly diverge.
In assessing Quality of Revenue, one of the key questions is that of the revenue model. Pricing of LLMs is still at the beginning of its evolution. Today most closed-source models are based on tokens. Input and output tokens can be priced similarly or differently. But as enterprise use cases emerge (vs. consumer and developer first more basic cost metrics), pricing based on the use case is currently growing (e.g. per classification and per search unit for Cohere). This is probably a natural path that will be taken with open-source models like Mistral.ai — whose first model is entirely free to date — but could develop paid enterprise versions in the future. The question of the monetization of Foundation models is still unresolved to a large extent, but answers will emerge as successful use cases (notably at enterprise level) are built.
In assessing Quality of Growth, a critical question is that of user retention. GenAI is not lacking in use cases or customer demand to date, but per recent Sequoia data, one-month retention of AI-first companies such as ChatGPT (42%) are behind the usual stats of month one mobile app retention. A second critical question, also similar to mobile apps, is that of user engagement measured through DAU/MAU. Per Sequoia, except for character.ai (due to the conversational nature of the app itself), AI-first companies have a much lower ratio than some of the best customer companies. This ties back to the quality of the final use case (as for pricing above) that will be the ultimate success factor in our view for foundational models.
In fact, many of these businesses follow the familiar SaaS playbook and mobile apps playbook, where their growth and success are strongly correlated to the quality of their customer (and developer) books, i.e. their number of customers or logos, and their user retention and engagement. Ensuring that product-market fit is okay is a first critical factor in making the investment decision. But where it gets tricky is that these companies are also “deep-tech” companies with regard to their initial upfront development costs, and large investments are often made at pre-revenue stage (think Mistral.ai) — where there is no product-market fit yet by definition (at least when the revenue model is still being figured out, as value is progressively demonstrated to the end customer). This is what should be playing out in the upcoming year of GenAI development, in which a clear framework for assessing Quality of Revenue and Quality of Growth will emerge.
Last, in assessing Quality of Margins, a critical question is the cost of computation (and energy) which has created a barrier to entry for start-ups wanting to build foundation models. But given the recent advancements in AI-first infrastructure, we expect these costs to significantly come down in the short-medium term. A second significant element to consider is that of the ballooning cost of AI data scientists. This costs is critical in understanding the margin computation of these companies.
Platforms and Hubs
Like open-source models, the success of AI-powered platforms and hubs relies on their ability to attract enough users to become the de facto industry standard. ML-focused examples here include Hugging Face, which recently secured $235 million of Series D funding in a deal led by Salesforce Ventures, or LangChain, which is still at an earlier-stage. The revenue model differs here from closed- and open-source foundation models, as Hugging Face for example does not charge for access to the models themselves (this is provided for free) but for computing power and customer support. Though this marks a difference in the nature of the revenue stream itself, questions around Quality of Growth, and the need for an investor to build a conviction on the retention and engagement of the user base, are identical to the ones presented infra. This similarity in the frameworks applied also works for understanding the cost structure of these companies, for which workforce costs are a key hurdle in scaling these AI platforms and hubs.
From a cost perspective, while Closed-Source and Open-Source companies require a lot of computing power, platforms and hubs are not as expensive to build. For instance, infrastructure startups like LangChain aren’t training models and, therefore, have smaller needs.
While there are a lot of open questions and few answers when it comes to analyzing Foundation Models within the GenAI tech stack, they play a pivotal role in generative AI development and the market potential for these models is substantial. Although this pillar of the GenAI tech stack is still evolving in real time, knowing which questions to ask when it comes to assessing the intrinsic value, resiliency and growth of these disruptive businesses is half the battle. Part of our work is to identify them, creating a benchmark for future Foundation Model businesses.
