How are open-source foundation models ever going to make money?

Or, are we entering into a new era of not-for-profit VC?

There have been a slew of open source AI projects that have begun to proliferate over the last few months, but none of them have captured more attention than open source foundation models (e.g., Llama, Mistral, Falcon).

Unlike the closed source foundation model players (e.g., OpenAI, Anthropic, Cohere), open-source foundation models make their latest and greatest models available to the public, for free, despite the fact that they’ve often had to spend tens of millions of dollars to train them.

Why would they ever do that?

Are they simply being generous? Do they want to give back to the AI community that has been so good to them? Do they believe that they are doing their part to make modern AI as safe as possible, and ensure it’s made available to everyone?

Better yet, are the VCs and publicly traded companies funding these projects just trying to make the world a better place, with no expectation of receiving anything in return?

No. Of course not.

Both VCs and publicly traded corporations have a fiduciary responsibility to provide the best possible financial returns to their LPs or shareholders. Put simply, these open-source models need to make money… eventually.

How do they plan on doing so? Let’s dig into it.

History of loss leading products

Loss leading products are not new to the tech world. Unlike other industries, where scaling can take decades, a unique characteristic of technology companies is their scalability and ability to take dominant market share quickly. The velocity of scaling companies at warp speed often leads to “flywheel effects”, wherein tech companies can leverage their large market share and feedback loops to make their product better than anyone else’s, and therefore squeeze out and stifle all competition.

In the social media era, this was often realized via network effects: as more users joined a single platform, the inherent value of that platform increased for everyone. Once a platform reached some userbase tipping point, it became almost impossible to directly compete with (why join a social network when all your friends are already on a different one?). In addition to the product becoming more valuable to everyone as the userbase grows, the switching costs also increased meaningfully. These are some of the reasons that we’ve seen the emergence of social network monopolies (e.g., Facebook).

For a new competitor to truly take off, they often must have a radically different product offering than incumbent social networks (e.g., TikTok’s emergence in short term video). Often, incumbents are the first to pick up on this new engagement format, and they will try to quickly acquire startups before they become a substitute (e.g., Facebook acquiring Instagram).

As any investor will tell you, monopolies (or even oligopolies) make for great businesses. Knowing this, investors who can invest over long hold periods (e.g., VC’s) are willing to lose money in the short term to maximize the prospect of “getting in early” at a massive, profitable monopoly / oligopoly over the long term.

These high market share businesses often have high profit margins and trade at premium multiples in the public markets, and are often looked at as the holy grail outcomes in early-stage investing.

There are many, many, examples of VC’s backing these types of businesses. In recent years, Big Tech companies have caught on to this as well and are willing to also take on the short-term losses in the pursuit of long-term gains.

To be successful, these businesses must be able to capture huge amounts of market share in a space quickly, before eventually turning on monetization and cranking up prices. Some examples include:

• Facebook: lost money for years before activating Ads

• Uber: VC money heavily subsidized the cost of rides to grow the user base (has anyone noticed how much more expensive Uber has become now that it is profitable?)

• AWS / GCP / Azure: burned corporate funds for years as they offered dirt cheap cloud compute, but now operate as an oligopoly in one of the largest and most valuable markets in the world

The list goes on from there. This doesn’t always work. Business quality clearly matters: if you can’t scale or drive strong unit economics, this strategy can fail spectacularly (e.g., dot-com era internet bubble). However, if everything goes right, this strategy can lead to some of the largest financial returns possible.

In short, these money losing, market share expansion strategies are high risk, high reward bets. Funding open-source foundation models is no different.

But unlike the above examples, there’s currently no “price” for open-source models to raise in the future, and no obvious way for these companies to advertise on their platforms. So how will they go about making the math work? Below are a few of my hypotheses for how these companies will eventually attempt to monetize.

Option 1: Cross sell into adjacent products

The simplest way open-source model providers can make money is by cross-selling into adjacent, high margin products, using open source models as a loss leader. This makes the most sense for Big Tech companies who are trying to lock you into their ecosystem, where their other business lines are extremely profitable. One can imagine this making sense for the global cloud providers, who want to lock you into their highly profitable cloud environments, but I also expect Meta to take this route with Llama.

The Llama open-source license already locks you into sticking with Llama models for the foreseeable future: you can only use a Llama model to train or fine tune other Llama-based models, effectively locking you into Meta’s ecosystem. Once your company is “locked in” to the Llama ecosystem, you will likely be forced to work with Meta’s other, high profit margin products (is this how they finally force people to pay for the metaverse?).

Option 2: Bifurcate your offerings

Right now, model providers generally fall into two camps: closed or open. But there’s nothing that says you can’t actually be both. OpenAI has already open-sourced older GPT models to the public for free, but charges money to use their latest and greatest offerings. This works as a developer acquisition tool, where devs can get used to the GPT offering when it’s free, so that when they flip the switch to activating more powerful, paid models, it’s simplest just to stick with OpenAI.

In the world of AI, you can experiment with less performant, free models, but ultimately you need to use the latest and greatest to compete with your competitors. OpenAI is betting that customers stay with them as they upgrade.

I would bet that this is exactly how players like Mistral will operate in the long run. Today, Mistral actually doesn’t lock you into their model ecosystem as Llama does: you can use Mistral models however you like, as they have a particularly flexible license. Once Mistral is able to prove that they can become the provider of choice for open-source models, they will almost certainly launch a paid offering, similar to Open AI.

As long as they continue to provide great (but not their very best) models to the open-source community for free, they can leverage their market share and distribution base to launch paid offerings. While I suspect this will include their most performant models, they might also keep enterprise features (e.g., security, PII management, etc.) behind a paywall.

While this all sounds like a sound strategy, maintaining a whole slew of open and closed sourced models is bound to be prohibitively expensive (hence the massive amounts of capital that they’ve already raised).

Option 3: Charge a markup on inference costs

While open-source models like Mistral, Llama, and Falcon don’t require users to spend money training a model, there is no way to get around paying to serve the model (also known as inference). If training a model represents the “fixed costs” of working with foundation models, serving it represents the “variable costs” or “COGS” of using working with them. Serving costs are volume based and dependent on the number of times a user calls an API and uses cloud resources to create the output.

Today, even if you use an open-source model for free, you need to pay for your own inference costs. If you’re only calling an API a few times per day these costs can be minimal, but as usage increases significantly, so will costs. In a future in which all software products call foundation model APIs constantly, many have theorized that inference costs will outpace training costs, meaning that the bills could add up quickly.

If you’re already using an open-source model and that company has a relationship with a cloud provider that makes it easy to use them for your inference, they can act as a middleman and take a spread on that transaction (whether they markup the cost to the user, or the reduce the cloud providers take-rate). While the above is targeted at inference usage, one could easily construct a similar argument regarding fine tuning spend.

Hugging Face has done an excellent job of leveraging this markup / spread on cloud spend with their users already, as they can facilitate models and cloud providers to their users directly and simply. Hugging Face, however, doesn’t have to train their own models, and as such doesn’t have the same capital requirements as model providers do. For this reason (not to mention having to compete with Hugging Face itself), I don’t expect many open-source providers to use this as their primary business model: I just don’t think there’s enough juice to justify the squeeze.

Option 4: Monetize your user data

Another possible monetization strategy for open-source model providers is to take a lesson from the social networks themselves: sell your user’s data, without ever charging them directly.

One advantage of open-source models is that they can generate an incredible volume of usage & prompt data very quickly, as a huge number of users flock to their models. Today, however, I don’t think they can track customer prompts and data on the other side. If the model is hosted on the customer’s private cloud / on prem, that usage data is likely never leaving that customer’s environment without the customers permission. This data sharing is a clear hurdle that needs to be crossed.

In the future, providers could explore data sharing agreements in which they get access to the anonymized usage data (perhaps just the embeddings), which they could then leverage to retrain & fine tune their own models, or even just sell directly to enterprises (after all: the real key to having the best AI is having access to the best data).

Takeaways

As expected, there’s no free lunch here. All open-source models will have to find a way to monetize eventually, but most are likely trying to drive up their market share as much as possible before turning on monetization.

This doesn’t mean that these companies can’t one day become great businesses: it just means that their monetization strategy has yet to be truly validated yet.

I’d love to hear from anyone who has different ideas and perspectives on how these companies will until turn into great businesses.

As always, if anyone is building in the space and would love to share ideas (especially if you think I’m wrong!), I’d love to chat.