What Matters at Meta (2025) — a Diligence Blueprint

Jul 07, 2025

TLDR

Meta’s super-profitable ads engine now funds a portfolio of AI bets. The edge lies in mapping today’s CapEx/OpEx to those five bets, judging fungibility of the spend, and handicapping payoff odds.

Quick Background

For the uninitiated, my view on the relevant context is as follows:

Meta has proven itself a dominant force in the ads industry. In 1Q25, Family of Apps revenue was up 16% with a 51% operating margin. Ad impressions were up 5% and price was up 10%.
Zuck has proven himself one of the best operators out there. He has been running this company since he was 19. He has made it through the web to mobile transition, feed to stories, stories to short-form video, Cambridge Analytica, and completed arguably two of the best acquisitions ever in Instagram (certainly) and WhatsApp (probably).
Meta is currently guiding to 2025 capex of $68b at the midpoint. This is 36% of consensus ‘25 revenue, up from ~15-25% historically.
Reality Labs is still generating operating losses of ~$4-5b per quarter.
Valuation background: On an LTM basis, META is trading for 27x earnings (22x ex RL losses). However, it trades for 36x LTM FCF and probably more like ~45x this year’s FCF (which will be down significantly yoy due to the capex jump).

Meta’s Culture and Capital Allocation

Meta’s approach to capital allocation philosophy is driven by its culture, which can be described as builder-first.

Understanding Meta requires an understanding of Zuckerberg’s builder philosophy (note that Zuck has 61% of voting power).

I think this is captured in the below excerpt of Zuck’s interview with Theo Von (edited for clarity). At first, I was a bit confused when Theo introduced Zuck as, among other things, an inventor. But now it makes sense:

Mark: …our seven-year-old is just purely generative, constantly just creating things. We’ll do 3D printing. She'll create 3D worlds in this Horizon Metaverse system that we're building. She codes, she writes books, she makes music… Seeing her has really helped me understand myself in a way because it's like - why do I just keep building stuff? Why do I care so much about creating stuff? And I think some people just have a thing in them where they have to create stuff. The stuff just comes out of them. They're constantly generating things. And I notice that in her. She's like that… Before she started growing up and doing that, Priscilla would [say to me], “why don't you just relax? You've built enough things. Your company is good. You can just chill.” And now I think after seeing this one, she's just like, “OK, no, I get it. You have the same thing. You have what she has, right?” It's like you're just constantly creating stuff.
Theo: Dude, yeah, you're Thomas Metason.

This is where capital allocation starts with Meta. It starts with Zuck asking himself, “What do I want to build?” Everything else follows.

What’s the function of a finance organization in a culture like this? CFO Susan Li paints the picture in her interview with John Collison (edited for clarity):

There's stuff that we can rigorously measure. So that's a lot of the core family of apps work in terms of the impact on engagement, the impact on monetization. There's a lot of that stuff that is really finely tuned.
And then there's a set of things which we constrain more. There's some envelope of investment that we're willing to make that's not in this really ROI-driven bucket. [For example], it is very difficult to pencil out what the annual revenue forecast for Reality Labs is going to look like over the next 20 years.
And so for bets like that, we sort of invert the problem. But when we talk about the return on the investment, the question that we pose as a finance organization to Mark, and make sure that Mark and the board understand, is, what does this have to be worth to pencil out at the end? And does that pass the sanity check, the intuition about what the size of these markets can be, based on maybe some comparisons to markets that exist today, but of course in another 10, 20 years, you expect that the world will look different, and maybe those markets should be bigger or smaller for whatever reason. And that's kind of the guide, which is like, hey, for this thing to succeed at the rate at which we're investing, it needs to be worth this at the end, and does that make sense?
And we're only building because we think that [the upside case] not only exists, but it's compelling, and it's compelling for financial reasons but also strategic reasons why we want that version of the world to exist.
This is a place where I've got to be honest with you. I was one of the last people at the company to hand my BlackBerry over for an iPhone… I am not a tech visionary. There are many things I'm good at, but envisioning the future of the world and what I want it to be like is not one of them… But Mark very much has a vision for what he wants that world to be. And for him, I think the strategic imperative is that we have to be building these next states of the world for us to, again, be a good business, but also just be a compelling company that builds technology and puts it out in the world and builds incredible experiences for people.
I remind people in the finance organization all the time, we are very good at skeptically evaluating each bet. But the point is not that we have to look at every bet and be like, "This bet is going to work." The point is there is a portfolio of bets, and some of them are going to pay off massively beyond, in fact, what the case on paper looks like when you make the bet, and many of them are going to not work out, but the ones that pay off are going to more than justify the overall investment strategy or the overall roadmap that you're building toward. And if we just allowed ourselves to nix everything that the paper case didn't seem high-confidence, then we would never make a lot of the important bets that have been really important over the history of the company.

So, we can think of Meta’s spending in two distinct buckets:

Core FOA Optimization – incremental investments in the ads engine, ranking systems, and engagement tools that reliably throw off high-ROI cash. It’s easy for them to model internally. There’s a long history, clear KPIs, and tight feedback loops.
Long-Horizon Venture-Like Bets – Reality Labs, frontier AI models, and whatever else springs from Zuckerberg’s builder instinct. Finance’s role here is to set the guardrails: “If we’re pouring $X billion in, the end state needs to be worth $Y billion; is that plausible?” Mark’s answer is almost always going to be, “Yes, let’s build.”

I’m uneasy with the second bucket. The payoffs are opaque, so sentiment (not fundamentals) often drives the multiple, and the sheer size of the spend can swing the stock hard. Investors punished Meta for its VR splurge in 2022, then rewarded the aggressive AI ramp in 2025, and today the stock sits near record levels while the company reportedly dangles nine-figure pay packages to lure top talent away from OpenAI.

Still, if I have to back anyone on moon-shot R&D, it’s probably Zuckerberg. He’s earned the benefit of the doubt.

Five Bets

On the 1Q25 call, Zuckerberg laid out five major AI-related opportunities that META is focusing on. I will list them below including a 1-10 rating of how core/predictable vs moonshot/speculative I view them to be:

Improved advertising
- 10/10, core. This a fine tuned machine.
- The fully evolved vision of this is advertising as “an AI agent that delivers measurable business results at scale”. The business owner tells Meta what they want to achieve and how much they are willing to spend, and Meta handles the rest. This would likely “make advertising a meaningfully larger share of global GDP than it is today.”
- Examples from the last quarter
  - Testing a new ads recommendation model for Reels that is driving conversion rates up by 5%
  - Seeing 30% more advertisers use AI creative tools
More engaging experiences
- 9/10, core. They are also very good here.
- Better recommendations for existing content
- Better new content from generative and interactive AI (this is more ambiguous to me)
- Example from the last six months
  - Improvements to their recommendation systems has led to a 7% increase in time spent on Facebook, 6% on Insta, and 35% on Threads (which now has 350m+ MAUs).
Business messaging
- 8/10, core. Newer but still a clear and growing business case.
- WhatsApp now has 3b+ MAUs, including 100m+ in the US “and growing quickly there”.
- Messenger has 1b+ MAUs.
- There are now as many messages sent each day on Insta as there are on Messenger.
- Thailand and Vietnam - due to low cost of labor, many businesses conduct commerce through messaging apps. As a result, those countries are both in META’s top 10 or 11 by revenue despite being ranked in the 30s in global GDP. This phenomenon will spread to developed countries with the advent of AI agents to handle messaging.
- A couple interesting notes from a June 2023 interview with Ankur Prasad, who was Global Director for Business Messaging Product Marketing at the time:
  - “About six or seven years ago, we saw that businesses in Southeast Asia, especially Thailand, and Vietnam, were using messaging. Our team took a trip down here, and we talked to a bunch of businesses, and we recognized so much commerce and so much conversation was happening on our Messenger app. So, that gave us an idea for what we call click-to-message ads.”
  - Business messaging (click-to-messaging ads) was a $10b run-rate business at the time
  - Marketing Messages - reengage customers with a messaging notification. 50-60% more efficient than email and SMS in terms of open rate and conversion. Even though you can opt-out of these messages, ~85% of people opt-in.
  - ~40% of advertisers in Vietnam use Meta, and ~80-90% of them use Click-to-message ads.
  - One in three people say they message a business each week
Meta AI
- 3/10, moonshot. Highly competitive, monetization questions, etc.
- Now 1b+ MAUs across apps
- Management mentioned that they are first focusing on scaling this before thinking about monetization, but that they will likely go down the path of trying ads and/or premium subscriptions.
  - Ads - I think it is still an open question as to whether people will be okay with ads in their chat bots and what the appropriate ad load would be. My personal opinion is that having ads start to show up would be detrimental to my use of these, and I would pay to avoid them (I already pay for three of them anyway).
  - Subscriptions - while I think knowledge workers who consider LLMs a second brain might subscribe to three, I’d be surprised if the general population does. So the market size here is tough to evaluate, and there are at least 3 major, well-funded competitors in Claude, Gemini, and GPT that could make this a race to the bottom in terms of margins.
AI devices
- 5/10, as I am increasingly convinced AR glasses will be a big market, but still skeptical on the VR headsets.
- Glasses “enable you to let an AI see what you see, hear what you hear, and talk to you throughout the day”.
- 1b+ people in the world wear glasses, “and it seems highly likely that these will become AI glasses over the next 5-10 years”.
  - This would be an incredible user platform to monetize with services/subscriptions and maybe ads. No one has ever monetized glasses like this. Forget screen time, you are essentially monetizing sight. That’s an exciting opportunity.
- Ray-Ban Meta AI glasses have tripled in sales yoy. “People who have them are using them a lot.”
- Quest - seeing deeper engagement with Quest 3S, and more people creating experience in Horizon with AI tools.

Diligence Framework

I think there are two groups of people worth talking to.

Formers from the finance organization
- 2-5 calls - the questions for this group are relatively tangible/knowable
- Profile
  - FP&A
  - ~Mid-level or higher
    - Finance Manager
    - Director, Finance
    - Senior Director
    - VP, Finance
  - As recently displaced as possible (Jan 2025 or earlier)
- Questions
  - What percentage of total capex/opex does the 5 bets account for?
  - Split of 2025 capex/opex across the 5 bets?
  - How ROI thresholds are set and reviewed, how this varies across the bets
  - What percentage of servers/GPU pools (and opex) are truly repurposable?
    - And does it map to certain use cases e.g. maybe you can repurpose “improved advertising” talent and/or compute infrastructure to “more engaging experiences”, but can it be used for Meta AI as well? And vice versa?
- Example candidates (honestly, these candidates are not that great either due to small scope of responsibilities and/or time since leaving, so you’d hope to maybe piece together a picture from talking to a few of them and/or continue looking for someone who has more relevant experience - including asking them who they think you should speak with)
Formers and/or other experts from infrastructure/AI engineering
- 5-10 calls - more ambiguous questions
- Profile
  - Preferably Staff or Principal Engineers in Production and/or ML Systems, or Data Center experts and other infra owners
  - As recently displaced as possible (Jan 2025 or earlier)
- Questions - more dependent on the candidate but some of these can include
  - Fungibility of both capex and opex
    - Opex - e.g. - if they hire someone to work on Llama 4, could (and would) they pivot to work on ads improvement or content recommendations?
    - What does the reallocation/repurposing process for GPUs look like? What needs to be done, if anything, to go from using a GPU for ads ranking to using it for glasses inference or Llama training or whatever other application has the highest transferability given the return profiles, hardware requirements/optimizations, etc.?
  - Custom silicon outlook, and to what extent that can be fungible
  - Useful life of servers (see an interesting analysis about this issue here)
  - Given Llama’s shortcomings, the aggressive recruiting from OpenAI, and the creation of Superintelligence Labs, how is morale in the META AI talent base? Are different factions arising?
- Example candidates (same note from before applies here, this was just who I found after a quick scan but surely there are better people out there):

My Guesses On Some Key Questions

Percent of total capex attributable to the five initiatives
- Probably 80-90%+. Maintenance capex likely very low and these five initiatives should be the substantial majority of the growth capex.
  - Opex is trickier. Obviously RL is operating at a $4-5b a quarter loss, but FOA is tough to read into in terms of how much is growth vs maintenance spend.
  - Additionally, useful life of servers (likely one of the more rapidly depreciating assets in the base) is supposedly 5.5 years.
Breakdown of spend across the different initiatives
- Meta has noted that the majority of the capex is directed toward the core business. While I could be misinterpreting this, I believe this means it is attributable to the first three initiatives (ads, content, messaging). I get the sense that most of the rest probably leans toward Meta AI given how compute-intensive it is to train foundational models like Llama. And then while RL accounts for a significant R&D/opex expense, capex is relatively low (this was a point Meta repeatedly emphasized when the stock was getting slammed for the RL investments). Perhaps the breakdown of growth capex is something like:
  - Ads - 25%
  - Content - 25%
  - Messaging - 10%
  - Meta AI - 30%
  - Reality Labs - 10%
Fungibility
- I don’t have a strong view on this from a technical perspective yet.
- Certainly, buildings etc. are fungible over time. Note that buildings accounted for 29% of total gross PPE in 2024, and 26% of the change in gross PPE (servers and network assets were 42% of the total and 62% of the increase). It’s possible they could hit an air pocket of underutilized facilities for a few years, but the buildings will be filled up eventually. So that isn’t material from a DCF perspective. The market might not like it (recall when AMZN “overbuilt” after COVID) but it doesn’t change the value much.
- GPUs and custom silicon seem like more of an open question. I don’t have conviction here yet. The useful life is shorter (apparently 5.5 years) and I am not sure about the fungibility across workloads. My base case it is reasonably fungible between core apps and gen AI work. This could lead to a similar air pocket if foundational model investment turns out to be too aggressive. The bigger concern, though, is if they just keep pouring billions in this indefinitely in a race to the bottom on ROI. Reality Labs has some of that same concern but at least a) we can clearly see the size of the investment on the income statement, and b) glasses are starting to show some level of promise.

Disclosure: none of this is investment advice, I may own positions in securities mentioned.

Stock Thoughts

Discussion about this post