Can Anyone Compete With Meta’s New AI Model?

Was this newsletter forwarded to you? Sign up to get it in your inbox.

What would you do if your competitor started offering a product that did exactly what yours did—for free? 

This isn’t a hypothetical scenario or founder’s nightmare. It is the cold, capitalist reality for AI model providers today and one with a long history in tech. The Internet Explorer web browser killed Marc Andreesen’s Netscape browser in 1998. Microsoft took Slack out behind the woodshed with Teams in 2019 and has far surpassed the former’s growth since then. And now, Mark Zuckerberg might do the same to OpenAI.

This week, Meta released the Llama 3.1 series of large language models. According to various benchmarks in areas such as math and reasoning, these models are sometimes superior and always competitive with offerings from OpenAI, Google, and Anthropic. To make it worse (at least, for OpenAI employee’s stock options), the model is open-source, making an important part of the AI tech stack free to use for enterprises.  

Llama’s new release is indicative of three truths:

1. An AI model is a remarkably poor product.
2. Open-source AI isn’t truly open-source if it relies on Meta’s good will.
3. If GPT-5 isn’t a banger, we are going to enter a rather extended AI winter.

Llama 3.1 is the most important AI model—open-source or otherwise—of the last year. I can think of no other category of technology, in the history of humanity, that has received AI’s level of funding, only to be potentially undercut by an entirely free offering.

Models are not products

At its most fundamental level, every technology is created by a producer combining components in novel ways. A pencil, for example, is just wood and graphite combined in a specialized manufacturing process. Similarly, when a startup claims that it is an AI company, it needs to be specific about the technical components it’s using as well as the “manufacturing” techniques that are combining those components.

So, when a major tech company calls its large language model “open-source,” what it’s actually saying is that it combined thousands of GPUs, enough electricity to power the state of Nebraska, and an entire internet’s worth of data using some fancy math to create model weights. Only the weights are open-source. They are themselves the lone end product. When a tech company says it’s made an open-source model, it typically means open-source weights.

Think of these weights as a map of an artificial mind. They determine how strongly different parts of an AI model—like its layers, neurons, and how it chooses to pay attention—are connected to each other. And these weights are given away for the community to use.

No closed-source AI company—not OpenAI, not Anthropic, not Google—is in the business of selling weights. All of them are in the business of selling access to the weights their large language model uses—a crucial difference. To illustrate what I mean, look at this chart below from the Meta blog post announcing Llama 3.1:

Each of the rows in this chart represents a technical capability that is necessary for the 3.1 weights to be performant—like fine-tuning, which helps an AI perform better in a specific domain. Without them, the model wouldn't be useful.

So if you are building a product that is reliant on large language model capabilities, the weights from Meta aren’t all that helpful on their own. You’ll still end up paying some of the providers in the columns to do a task contained in the rows.

You’ll note that OpenAI and Anthropic, the two leading AI model vendors, aren’t listed. They have a different business model—one against which Zuck is explicitly betting.

Unlike Meta’s free proposition, OpenAI and Anthropic don’t offer the raw weights. Customers pay for access to LLM capabilities through their API and consumer applications like ChatGPT. The model is only part of the proposition. In the case of the API, you are paying for a host of other features that matter for enterprise use cases, like security, continuous upgrades, and data analysis; with the consumer applications, you are paying for features like memory or data analysis tools neatly packaged in a chatbot. The average user needs both of these things more than they need the weights. The LLM is simply table stakes to accomplish the thing the developer is trying to have the application do.

Even companies in the large language model business make money by selling features surrounding the weights themselves. Mark Zuckerberg is aware of this. From the model release announcement:

“To ensure that we have access to the best technology and aren’t locked into a closed ecosystem over the long term, Llama needs to develop into a full ecosystem of tools, efficiency improvements, silicon optimizations, and other integrations. If we were the only company using Llama, this ecosystem wouldn’t develop and we’d fare no better than the closed variants of Unix.” [Emphasis added]

Zuck believes it will be more valuable in the long term to enable a “full ecosystem” to develop around Llama by opening its weights. Developers can modify the weights for specific tasks, and researchers can use them to advance AI technology. He also believes it’ll lead to a more thriving ecosystem—for Meta and AI as a whole—which isn’t the case with closed-door services like ChatGPT or Anthropic.

Which strategy is right? One question can answer that: Which components (i.e., rows in the table) are more valuable when directly integrated by the model provider? If the majority of value comes from component integration instead of supplier specialization, model providers like OpenAI and Anthropic are going to be just fine. Meta’s Llama models are only destructive in the long term if the ecosystem of suppliers (i.e., the columns in the chart) can outperform components integrated into the model. 

Beyond capabilities, there are also cost considerations. Llama is spectacular relative to closed-source competition. Again, from the blog post about the release, “Developers can run inference on Llama 3.1 405B on their own infra[structure] at roughly 50 percent the cost of using closed models like GPT-4o, for both user-facing and offline inference tasks.” [Emphasis added] These supposed savings are almost certainly made without the cost of the expertise to implement Llama. The benefit of using an integrated provider is that it does the optimizations that your own developers don’t have the specialization to do—but still, a 50 percent savings is significant. 

If anything, Llama 3.1 is mostly good news for its ecosystem partners. It is a compelling enough offering that some companies will be looking to swap out their reliance on OpenAI—and they’ll need help to do it.

How open is Meta’s open-source model?

In the same post, Zuckerberg discussed his reasoning for launching Llama as open-source:

“A key difference between Meta and closed model providers is that selling access to AI models isn’t our business model. That means openly releasing Llama doesn’t undercut our revenue, sustainability, or ability to invest in research like it does for closed providers. (This is one reason several closed providers consistently lobby governments against open-source.)”

Meta runs ads—it doesn't sell models. For that reason, Meta argues that open-source is beneficial for them. AI models are only revenue-threatening if people spend more time using them than scrolling Facebook and Instagram. 

I’m unconvinced by his argument. While Meta does not sell AI model access, it does monetize the time that AI models could theoretically be taking from Meta’s apps. Publishing Llama’s weights as an open-source offering makes sense if the LLM itself becomes integrated into the Meta universe of apps. But so far, integrating LLMs into the Meta-verse of apps (pun intended) has mostly consisted of haphazardly shoving AI into Instagram search (I have yet to meet anyone who has enjoyed that feature) and launching a destination chatbot—both deeply underwhelming experiences.

Meta has been successful in integrating an AI assistant into its Ray-Ban smart glasses, so perhaps that is where the Meta team sees these models being used. In an interview with Bloomberg, Zuckerberg talked about allowing creators to train AI duplicates of themselves as “artistic artifacts” that allow them to communicate with their fans. Sure! Maybe? When it comes to AI products, we are still mostly dealing with hypotheticals.

That leaves the hardest road—making something new—which, based on Meta’s history of innovation, means ripping off competitors’ products. 

The most likely candidate for the company to copy is Character.AI, an app that enables users to have conversations with AI chatbots mimicking celebrities like Elon Musk or popular franchise figures. Users spend a staggering amount of time on it—last October, the company claimed that its 3.5 million users spent two hours a day talking with AIs. (Eek. Gross. Ahhhh. Weird.) For Meta, this product would present such an extreme reputational risk that it is likely a long time before it’s willing to copy and paste the offering. As demonstrated by my own reaction, I’m not sure society is ready for our biggest social media company to sell us a product where we spend all day role-playing with chatbots. 

There is another possibility. Let’s say Zuck really is doing all of this altruistically. Maybe his bone-deep hatred of Apple and its platform restrictions has convinced him to incinerate billions to ensure that he isn’t held hostage to someone else’s platform ever again, even if there is a small likelihood that the platform would benefit his company in the long term. Why do I think he hates Apple this much? Because he said so:

“One of my formative experiences has been building our services constrained by what Apple will let us build on their platforms. Between the way they tax developers, the arbitrary rules they apply, and all the product innovations they block from shipping, it’s clear that Meta and many other companies would be freed up to build much better services for people if we could build the best versions of our products and competitors were not able to constrain what we could build. On a philosophical level, this is a major reason why I believe so strongly in building open ecosystems in AI and AR/VR for the next generation of computing.”

We have the numbers to back this theory up. Since the fourth quarter of 2020, Meta has lost $46 billion in its VR division, a number so large that I struggle to contextualize it. By comparison, the few billion he is spending on AI likely feels relatively easy for Meta’s executive team to stomach. It could be that LLMs end up being the next great consumer app, but it could also be that Zuckerberg is doing an act of collective tech community philanthropy by ensuring we are not ruled by God Emperor Sam Altman.

I do need to give additional credit to the Meta team: Open-sourcing the weights are useful, but they went a step further. Meta’s research paper about Llama 3 is more detailed than anything the AI scientific community has seen in years. To enable companies to customize their models, you have to outsource the scientific progress that went into building the open-source one (i.e., how you combine the components that I described earlier). Doing so means other companies can recreate models from scratch faster and more cheaply than they could on their own. And that’s just what Meta enabled.

Still, when I consider how much of the open-source AI community is reliant on Meta’s goodwill—or its patience to fund open-source until it finds an LLM use case—I grow concerned. While a few billion is pocket change to Meta, it is an unfathomable number to every other startup. An open-source AI movement that depends on Meta to release frontier models for everyone to use is just closed-source AI with extra steps. Frontier models are enormously expensive, and if we all need Meta to foot the bill, we should be worried. Unless a more obviously material revenue lift becomes clear, I’m not sure how much stomach Meta has to sustain long-term losses.

This industry hinges on GPT-5 being a banger

Over the past year, I have talked to dozens of AI investors and executives at over 50 software startups, all of whom have been trying to figure out how to incorporate AI into their applications. Almost universally, the feedback is that the technology of this current generation of models isn’t quite good enough. While there are some use cases that work well—things like call centers, coding, and text generation—hallucination and error rates are still too high for many of the hoped-for ones to emerge.

Llama 3.1 only came out two days ago. So far, I have not been wowed by its capability. It feels competitive with other models of its vintage, but it is not in a league of its own. 

The potential of AI has driven extraordinary levels of investment. Sequoia Capital estimates that in the fourth quarter of this year there will be approximately $300 billion in AI data center spending. It also estimated that there would need to be $500 billion generated to cover the gap between estimated spend and AI revenue.

The current generation of models simply isn’t capable of reaching that revenue number. Perhaps people will figure out new ways of using them, adding new rows in our earlier chart that will unlock some huge revenue breakthroughs. But I doubt this will be the case. I think we will need to make all those component and implementation improvements, plus get a new generation of models with dramatically improved capabilities.

The incremental gains from the last year are insufficient to justify the current level of spend. I am increasingly concerned that even several years more of optimization won’t be enough to deliver on the grand promises of LLMs as an entirely new way of interacting with computers. The science fiction we are hoping for will be downgraded into another wave of price increases for B2B SaaS products.

This is the privately held opinion of many industry operators: Current models are pretty good, next models should be great. OpenAI’s GPT-5 should be really good. Meta signaled that this is also its position by naming this release Llama 3.1 instead of Llama 4. Microsoft CTO Kevin Scott went on the record with Every CEO Dan Shipper about how GPT-5 will be that much better than its predecessor. (Microsoft is the lead investor in OpenAI.) All spending and conversations are being conducted under that assumption. If they are right—great! If they are wrong, be prepared for an unbearable amount of think pieces about Silicon Valley’s hubris. My fears may be unfounded, but given Llama’s 3.1 only marginally market-leading capabilities, we should probably be discussing the possibility.

This is exactly what the middle of a bubble looks and sounds like. If GPT-5 is only an incremental improvement, expect the bubble to pop.

Executives are aware of this risk but have considered the investment worth it. In the same interview with Bloomberg, Zuckerberg said that there is a “meaningful chance a lot of companies are over-building now and we’ll look back and they’ve spent some number of billions more than they had to, but they’re all making rational decisions because the downside of being behind leaves you out of position for the most important technology over the next 10-15 years vs. over-investing you lose some amount of money that an affordable amount to lose.” [Emphasis added]

Right or wrong, this is still the belief in Silicon Valley: It’s better to be wrong to the tune of a few billion than to miss out on the next trillion. As always, much of the assumptions of this industry depend on scaling timelines.

Llama is a perfect case study of what matters most in business strategy—power. Increasingly, it is looking less and less like power lies with the models and more with the ecosystem surrounding them.