For years, Zuckerberg’s superpower and weakness were the same thing: he could bet Meta on an idea no one else believed in.
But since 2025, something has changed.
The idealist who charmed developers and users is suddenly gone. In its place, a capitalist CEO prioritizes shareholders and fund managers.
In this one, I’ll cover:
Where did Llama go? Is this the beginning of the end for open-weight AI?
Is Meta really preparing to sell compute? How serious is Zuckerberg about this?
Is enterprise AI becoming the next Meta business?
What should we look for in the next earnings call?
But let’s start with a little history, so you know why this is such a big deal.
Why was Meta a cornerstone of software?
To understand why this shift matters, you have to remember how Meta used to earn loyalty.
Meta was much more than just the social media empire that sells ads.
It was the company behind many serious open-source projects, not an exaggeration to say it's a cornerstone of today’s software. Because the world w/o open-source will be slower, costlier, and have fewer innovations.
React, for example, a default interface language, initiated by Facebook. PyTorch, the language that powers flexible, efficient modern AI, originated at Meta.
In the Q2 2024 earnings call, Zuckerberg believed the same approach would bring Llama as the foundation of AI:
This approach has consistently worked for us, and I expect it will work here (Llama) too.
In his dream, Llama should not be locked to a single cloud, and developers should be able to use it anywhere.
Though that same founder logic also showed up in a pricey way: for example, Meta's CFO warned investors in late 2021 that Reality Labs alone would reduce operating profit by approximately $10 billion. A bet that still hasn’t paid off.
And unfortunately, Llama is the AI version of this loss bet.
By 2025 and 2026, Meta’s vocabulary changed. Less “open,” “community-forward,” and “ecosystem.” But more “returns,” “grow,” and “superintelligence”.
Where’s Llama?
The word Llama was mentioned 19 times in the 2023 earnings transcript, and now to 0 in the Q1 2026 earnings transcript.
What happened?
Meta had already proven this model with software time and again, so why stop now?
For this, you need to understand some fundamental differences between AI and all other deterministic software.
Code libraries like React, open source, create a real feedback loop: anyone can inspect the code, fix bugs, build libraries around it, and teach other developers how to use it. The community distributes and improves the product.
When React becomes the default framework for any professional web project, Meta benefits three times:
The ecosystem gets better as more people build tools around it
Meta is in control of the standard
And engineers arrive already familiar with tools that came out of Meta.
But why did this model fail in AI?
In essence, if open-source is about collaboration and community, the nature of AI hinders both.
What do I mean by this?
To start with, very few people can meaningfully retrain or improve the base model, apart from the company that owns it.
Which means there’s no way to treat an AI model (especially the frontier ones) like a standard library where a developer can submit a patch from their devices. Improving the model itself requires enormous resources that only the hyperscalers like Meta, Google, and Alibaba have.
Secondly, red-teaming can expose problems, but realistically, no one, not even the AI labs, has the means to fix them; similarly, inference optimization can make deployment cheaper, but again, the base model remains controlled by whoever has the compute, data, and capital to train the next one.
Basically, with AI, the community isn’t in control and doesn’t have the means to improve AI as it did with open-source libraries.
They can only do things around the edges of a model, such as tooling, fine-tuning methods, and developer adoption, think OpenClaw and Cursor, or webUI to provide an easy chat interface.
Those things still matter; however, they are not the same as improving the core model.
So the open-source dynamic is weak, if not impossible for AI.
Interchangeable nature
Now zoom out to the AI economic level.
AI models are becoming increasingly interchangeable.
Just think of AI as broadband, and the labs are the provider. As long as you didn’t sign a two- to three-year contract, you can switch at any time to a provider that promises cheaper, faster access.
Some companies have already slowly learned not to build too tightly around a single model, because of all the obvious downsides I pointed out earlier.
So with the existing infrastructure, developers can route through tools like OpenRouter, which allows them to switch from one model to another for better quality, lower latency, or lower token prices.
The competition is fierce, and the ecosystem drives the commoditization hard.
Where model routing makes sense for users and companies, it certainly weakens the strategic value of open-sourcing Llama.
Given there’s no incentive for anyone to commit to locking themselves into Llama, Meta will never get the same ecosystem advantage it had with React.
Put simply, there’s no path to a return or long-term benefit to justify maintaining public models. Safety debates, developer relations, open-source positioning, and so on, all take real effort. And effort means people, which ultimately means money spent.
This explains why I can no longer find open-weight or Llama in the latest earnings transcript.
Because cutting back is the only rational decision on an investment without return.
If Llama is the first clue that the old philosophy is no longer valid, the real pressure comes from CapEx.
Once infrastructure spending reaches this scale, compute stops being just something Meta consumes. It becomes something Meta has to justify.
Compute Stops Being Just Zuckerberg’s Bet
Before discussing compute, we should address a shift in mindset.
A public company is ultimately run for shareholders. Even with 61% of the voting power, shareholder value is still not decorative and can’t be ignored; having every major fund owning a piece of Meta, Zuckerberg eventually has to bend to shareholder pressure.
If investors stop believing, they sell.
If the stock falls far enough, it does not just make Zuckerberg poorer on paper. It also makes the business harder to run. Stock-based compensation becomes less attractive. Acquisitions become more expensive. Raising capital gets harder, and so on. If it gets worse, ultimately, the company’s strategic flexibility shrinks.
So there’s a limit to how many “trust me” cards Zuckerberg could play.
He has made huge bets and failed hard, such as the metaverse, AR glasses, and, as we just talked about, Llama.
Some of these are still available as products, but none of them show enough traction to justify the CapEx. So after enough giant bets with delayed payoffs, the market demands a simpler answer to this one: when will we see returns?
And Meta’s spending is now too large for that question to be optional, its CapEx growing from $15.7B in 2020 to a 2026 guide of $125-145B driven by investments in servers, data centers, and network infrastructure.
Now, if someone outside Meta is willing to pay for GPU capacity, that creates a point of reference.
Then all internal uses have to compete against it. We’re talking ads ranking, recommendations, Meta AI, business agents, frontier model training, and internal productivity tools as mentioned.
So Zuckerberg is now forced to answer: Does this internal use of compute create more value than selling the same capacity to someone else?
This dynamic changes the nature of the spending.
Compute is no longer only a cost center at Meta. It becomes an asset that can be monetized and assigned a market-based price.
What did Meta originally build the capacity for?
From 2022 onward, Meta’s stated rationale for the data-center build-out was not to serve external needs. But the
ranking and recommendations
ads systems
generative AI training
frontier models… and more
Until the end of 2025, the story always was: Meta needs enormous compute because its own products are becoming AI systems.
How the cloud strategy has changed over the last 3 years
More than once, Zuckerberg has been explicit that he doesn’t plan to run a cloud business.
For example, in Q2 2023, Zuckerberg stated:
We partnered with Microsoft specifically because we don’t have a public cloud offering. So this isn’t about us getting into that. It’s actually the opposite.
Even in mid 2025, Susan Li (CFO) was still drawing a line around internal use:
So at present, we’re not really thinking about external use cases on the infrastructure…
Then three months later, in the earnings transcript, October 2025, Zuckerberg said something very different:
almost every week, people come to us from outside the company asking us to stand up an API service or asking if we have different compute that they could get from us.
with no definite rejection, hinting that they were considering cloud service as their offering.
While Zuckerberg himself introduced the downside case. In the Q3 2025 earnings call, he said it was “if you got to a point where you overbuilt…” followed by:
in which case of course there would be some loss and depreciation, but we’d grow into that and use it over time.
But in reality, if internal demand is slower than expected, to avoid this expensive loss, selling access to outsiders becomes the obvious way to earn something on the asset before Meta needs it back.
So the strategic question now becomes:
Is this a cloud business or just optionality?
If Meta wants to become a real cloud/infrastructure provider, it needs more than spare GPUs. It needs pricing, support, reliability commitments, sales motion, developer tooling, and some version of enterprise account management.
You do not become AWS because you have extra capacity for a few quarters.
Yet, if Meta only wants to monetize temporary overbuild, the logic is the opposite; it’d want minimum investment and max the tenancy rate. So when internal compute demand catches up, Meta can pull capacity back in the shortest possible time.
To properly analyze this, we need to quickly look at its cash cow and what it lacks.
From Distribution Channel to Cloud Business
Meta’s old business was an attention hub and a distribution platform. It owned consumer attention, then businesses paid Meta to reach users through ads.
Meta never went inside a customer’s company and rewired how the business worked. It sold reach, targeting, measurement, and conversion. The customer did the rest.
But cloud and AI are different.
As I said, if Meta overbuilds, the simplest version is to rent capacity.
That could mean raw GPU access, hosted model access, API access, or some bundle of all three.
GPU access only
The easiest path is the SpaceX/Anthropic model: find one huge AI lab tenant and let them absorb a large block of capacity.
That is the landlord's approach. It’s more than just clean, because in this scenario, Meta has limited clients to manage, and most importantly, it is much more flexible if it still expects internal needs to catch up with the estimated capacity.
Hosted model access and API
Here’s a more complex and less flexible option: to sell both cloud and AI, essentially, to compete with the likes of Google and AWS.
Why is this one more complex?
Because you cannot just show an enterprise customer a cloud or AI product and assume adoption happens. Enterprise AI needs platforms, models, tools, governance, support, migration help, and integration.
If Meta wants companies to use its AI tools, it needs so much more than a product page.
On top of this, the company’s language has shifted from Llama (facing the developer community) toward Meta Superintelligence Labs, Muse/Spark, business agents, and more controlled AI products.
This creates a new problem: someone has to buy and use them. In AI, that increasingly means forward-deployed engineers: people who sit with customers and turn the vague “we need AI” ambition into real workflows.
Therefore, if they go with this route, it also creates usage.
Because the customers would need AI inside CRM, support, commerce, finance, compliance, and internal tools. Which translates to token usage. Token usage creates compute demand. Compute demand justifies the data centers.
That said, the evidence from earnings calls still points more toward optionality than full cloud commitment. Zuckerberg said:
And we haven’t done that yet, but obviously if you got to a point where you overbuilt, you could have that as an option.
However, a few weeks later, some media reported that they had already allocated significant personnel to oversee this division, more than just Meta is building Meta Compute as a landlord; they also named the people appointed to oversee the Enterprise Solutions division.
If you put these two together: the Meta Compute + Enterprise Solutions, this becomes an initiative a whole lot more serious and long-term than it sounds.
Either way, this is a very different Zuckerberg.
The old version made giant internal bets and asked investors to wait.
The new one seems to stop everything that doesn’t generate returns; instead, it focuses more on justifying CapEx and on leveraging unused compute in an unseen-by-Meta way.
All of which are signs of Zuckerberg being a responsible CEO rather than a hot-headed founder.
What to Look Out for?
As we talked about in the last video, the model layer alone has no clean earnings path and keeps getting pushed toward commoditization.
Same as the other established players, Meta has a real edge.
It has ad revenue. It arguably has the best LLM AI use case on earth: AI-tailored ad experiences, something I talked about a year back. But even that, Zuckerberg, like all other CEOs, understands Meta can’t just live on its past glory. I admired his attempt with Llama, but the only play that won't lose investors' hearts is a clear revenue path.
With the current circumstances, it means selling infrastructure, distribution, or workflow lock-in around the model.
In the next few earnings calls, I’ll keep an eye on the following questions:
The language used to indicate whether Meta still says it is capacity-constrained, or starts talking openly about utilization and monetizing surplus? In other words, is this moving to either “selling excess capacity” or “we now build a real cloud business”?
Which route are they taking? Is Meta selling raw GPU capacity, access to hosted models, API access, or bundled AI services? Keep an eye on which rumor Meta confirms in the next few earnings calls.
Most importantly, it will take a year or two to see how this business plays out. If it’s just selling raw GPU capacity, the prep work is minimal, while Meta is still going to be a year behind their opponents, eg. SpaceX. But if Meta is going after what Azure and AWS already do, the game is far more complex: more clients, more services, more infrastructure (not limited to just hardware) to build, which puts it 10 years behind the established players.
Which path do you think Meta is more likely to take?
