Might Meta step back from foundation model development?

Last week, The New York Times reported that Meta will delay the rollout of its next frontier foundation model, Avocado, and might license Gemini from Google for use in Meta AI. From the article:

Meta’s new foundational A.I. model, which the company has been working on for months, has fallen short of the performance of leading A.I. models from rivals like Google, OpenAI and Anthropic on internal tests for reasoning, coding and writing, said the people, who were not authorized to speak publicly about confidential matters …

As a result, Meta has delayed Avocado’s release to at least May from this month, the people said. They added that the leaders of Meta’s A.I. division had instead discussed temporarily licensing Gemini to power the company’s A.I. products, though no decisions have been reached.

Meta announced last December that its TBD lab, under the Meta Superintelligence Labs (MSL) umbrella, would develop two new frontier foundation models: Avocado, a large-language model, and Mango, an image and video generation model. Meta had hired aggressively to staff this team, led by the former CEO of Scale AI, which Meta became a minority owner of in June 2025 through an acqui-hire arrangement.

Meta released its Llama 4 family of models, Scout and Maverick, in April 2025. At the time, they outperformed several leading models, including closed models, although Llama currently does not appear among the top-ranked models on LM Arena’s overall text leaderboard. Among open models, Alibaba’s Qwen overtook Llama in adoption in October 2025 and is now the most popular family of open models (note that Qwen uses the more permissive Apache 2.0 license).

Public market investors may wonder if Meta will step back from the CapEx cliff and slow its investments into foundation models. But I believe this question does not logically follow from the reporting and probably overstates the strategic severity of delaying a model release. First, if Meta is indeed committed to building frontier foundation models, then it would delay any release until it outperforms the current leaders. If anything, the delay reinforces Meta’s commitment. While I remain skeptical that Meta is, in actuality, considering licensing Gemini in the interim, as the reporting suggests, to do so would be neither a capitulation nor a significant deviation from its current operating approach: it partnered with Midjourney and Black Forest Labs for its Vibes app. So I don’t read too much into any rumored intention to license Gemini for Meta AI.

But second, and more fundamentally, I believe Meta is committed to the development of frontier foundation models as a means of establishing the kind of platform defensibility it has never really enjoyed. My closest approximation of a theory of mind for Mark Zuckerberg on this point comes from a memo he wrote to top lieutenants in 2015 as the company (then called Facebook) considered acquiring the game development platform Unity3d (disclosure: I’m an advisor to Unity on its AI initiatives). Relevant excerpts from that memo are presented below (emphasis mine), although I recommend reading it in its entirety:

The strategic goal is clearest. We are vulnerable on mobile to Google and Apple because they make major mobile platforms. We would like a stronger strategic position in the next wave of computing. We can achieve this only by building both a major platform as well as key apps.

I will discuss the main elements of the platform and key apps further below, but for now keep in mind that we need to succeed in building both a major platform and key apps to improve our strategic position on the next platform. If we only build key apps but not the platform, we will remain in our current position. If we only build the platform but not the key apps, we may be in a worse position. We need to build both …

The brand goal is also simple. The weakest element of our brand is innovation, which is a vulnerable position for us as a technology company dependent on recruiting the best engineers to build the future. Having an innovative brand will pay dividends not only in recruiting but therefore across all of our products and other efforts as well …

The key apps are what you’d expect: social communication and media consumption, especially immersive video. Gaming is critical but is more hits driven and ephemeral, so owning the key games seems less important than simply making sure they exist on our platform. I expect everyone will use social communication and media consumption tools, and that we’ll build a large business if we are successful in these spaces. We will need a large investment and dedicated strategy to build the best services in these spaces. For now though, I’ll just assert that building social services is our core competence, so I’ll save elaborating further on that for another day …

It’s worth noting a few things. First, these platform services should be cross platform. Most of the services can be offered on iOS, Android, on desktop, etc. On Android, we can bother offer an app store and offer many of these services to apps distributed through Play – if we app switch to our preloaded marketplace for purchases, we won’t even have to pay Google’s 30% rev share. Second, this platform definition is not actually an OS in the technical sense. In modern OSes, however, most of the value comes from advantaging the OS provider’s own platform services on the devices where its OS is installed. I think our primary platform strategy should not be focusing on building a fully independent OS, but owning these core platform services across all systems. This will be challenging as OS providers will try to push us out, but if we build superior services and provide things OSes need (eg Unity support), then we have a good shot at success …

One important question is that if our strategy is to win key apps and platform services, then why do we need to make such a big investment in hardware and systems? This is an especially important question when we’re considering investing billions of dollars into Unity over the next decade … At some level, it’s important to own the core technology you depend on to achieve your mission. Even if there is potentially a path forward with it, owning it increases integration opportunities and decreases risk …

If we own Unity, then Android, Windows and iOS will all need to support them or larger portions of their ecosystems won’t work. While we wouldn’t reject them outright, we will have options for how deeply we support them.

For open platforms like Android and Windows, this helps level the playing field and helps to ensure we can continue offering our app stores and other key services. For iOS, this will not influence Apple to let us offer an app store, but it could give us other important bargaining chips as part of our VR / AR strategy or otherwise.

Note here that I’m not comparing an investment in a game development platform to an investment in building foundation models. I’m investigating something more fundamental: the recognition that owning the services layer for a cross-platform consumer interaction paradigm is likely seen internally at Meta as the company’s prevailing strategic priority, especially following its experience with Apple’s App Tracking Transparency (ATT) privacy policy in the 2021-2022 period.

In the current context, that can mean developing not only a frontier foundation model but also offering a suite of services built on top of it. By providing a core service to developers, Meta establishes indispensability on existing platforms like iOS and Android (which are seeing their grip on things like payment processing and app store exclusivity weakened) and future-proofs its position on any emergent consumer platform.

At first blush, this might seem at odds with the point I made about Meta’s superintelligence efforts in Natura non facit saltus:

Meanwhile, ‘superintelligence’ is, at best, an abstract concept that is impossible to truly evaluate progress for. At worst, superintelligence is either a marketing gimmick or simply bad science, like ML alchemy. Superintelligence, from the outside, features no falsifiable milestones or recognized benchmarks. It takes the shape of an empty metaphor.

But most importantly, superintelligence represents a leap: a discontinuous progression that is not only not guaranteed but might not support Meta’s core business to a degree that justifies its pursuit. Meta is most productive, across a number of dimensions, when its focus is absorbed by the core technologies that empower its humanity-scale advertising platform. The company flounders and flails when it pursues sensational, abstract goals that aren’t anchored to its core business in obvious ways. From my vantage point, superintelligence is one such pursuit.

But these aren’t truly incongruent. First, Meta already builds highly performant foundation models: GEM is, at its heart, a large-scale deep ranking model that functions as a foundation model within Meta’s advertising stack. And second, any LLM developments also support Meta’s commercial advertising efforts, as I note in Can LLMs be used for ads ranking?. If Meta’s Avocado and Mango (and rumored Watermelon) models also advance its advertising and social businesses, then that alone could justify further investment, even if Meta could conceivably continue to operate those businesses with other open models (they use Qwen in the paper cited in the LLM ad ranking piece).

Second, owning a leading foundation model puts Meta into a critical role in the consumer-focused AI supply chain that, again, protects its position on platforms it doesn’t own. Frontier models function as the lowest common dependency layer for the emerging AI software stack, which makes them structurally similar to operating systems as conceived in the memo. That’s more valuable for the company than tethering its own apps to its own models because it creates the kind of dependence that renders it immune from platform enforcement in the way I describe in The App Store has a ‘Too Big To Fail’ problem:

The presence of certain apps is so imperative to the App Store’s broader consumer appeal that Apple has no choice but to allow those apps to defy its various platform rules and restrictions. These apps are too big to fail: if Apple were to enforce its rules against these apps and block them from the App Store, it would suffer extreme commercial consequences.

Meta, as the service layer for various forms of AI-enabled functionality across the consumer landscape, could provide that, and it’s evident, at least as far back as the memo excerpted above, that Meta sees it as strategically vital.