Meta: The Market Reads the Capex Bill and Ignores What It Is Buying

Meta Platforms reported first-quarter results that, on paper, should please any investor: revenue of $56.3 billion, up 33% year over year, and a 41% operating margin. The market sold the stock anyway. The reason fits in a single word c(apex) and in a figure that induces vertigo: between $125 and $145 billion of spending on chips, data centers and infrastructure in 2026 alone. The fear is legitimate. The conclusion the market draws from it, I think, is wrong.

The research house The Long Investor, which for months tagged Meta as "Accumulate," has softened its tone to "Hold" with a $730 price target — not because the business is doing badly, but because they believe there are better risk/reward profiles among the other big-tech names. It is a defensible stance. But note that even this sits below the average of analysts covering the stock, today around $828, with high estimates already brushing $1,000. The pessimism is really one of soft consensus, not of bearish conviction. This is where my thesis begins.

01 · The business: A money machine the market underrates

To understand Meta you have to separate two things that often get blended. On one side is the Family of Apps —Facebook, Instagram, WhatsApp, Messenger, Threads— which gathers nearly 4 billion users and generates almost all of the business. On the other, Reality Labs, the virtual-reality and smart-glasses division, which loses money by the bucketload. 98% of revenue comes from advertising in the first. Reality Labs, for context, contributed $2.2 billion of total revenue of $200.9 billion.

The key detail of the quarter is not growth itself, but how it grew. Ad revenue rose on two levers at once: impressions grew 19% and average price per ad grew 12%. Both rising at the same time is unusual and very powerful — normally, when an advertiser pays more per impact, you tend to serve fewer. That Meta manages to sell more ads and at a higher price means its targeting and recommendation ability has improved tangibly. And here AI already appears: Meta's automated tools are improving their clients' advertising ROI, and an ad that converts more is an ad that can be charged more.

This is the part the "runaway capex" narrative tends to obscure. Meta does not spend on AI as a fad; it spends because every point of improvement in ad ranking translates directly into revenue over a base of 3.56 billion daily users. The ad engine is not a cash cow to milk while new things are invented: it is, at once, the product that benefits most from the spending that scares investors.

The right question is not whether Meta spends too much, but what it buys with what it spends.

02 · The capex: what frightens and what it buys

Let's get to the heart of the fear. The company raised its infrastructure-spending guidance from $115–135 billion to $125–145 billion in a single quarter, and this year's free cash flow will be compressed to a ridiculous $2 billion, against a three-year average of around $46 billion. On paper, this is cash destruction. The market reads it, shudders, and punishes.

But a capex peak is not the same as a structural loss. It is an advance investment that either pays back or does not — and the entire thesis rests on that distinction. The same analysts who see $2 billion of FCF this year expect recovery toward $10 billion in 2027, $30 billion in 2028 and normalization around $52 billion thereafter, as the infrastructure gets amortized and begins to monetize. Recall, too, that Meta generates $135 billion of operating cash and that this figure should grow toward $197 billion in 2028. The cash is there; it is just that right now a huge part of it is diverted to construction.

And what does that construction buy? Until recently the answer was abstract — better targeting, the Meta AI assistant spread across WhatsApp, Instagram and Facebook, the AI glasses. This week a much more concrete and, for the thesis, much more interesting answer appeared: according to several outlets, Meta is studying renting its excess compute capacity to third parties, a move toward the compute-as-a-service business in the style of Azure or xAI. The stock reacted with a gain near 10%. If confirmed, it turns a cost center — the data centers — into a potential line of recurring revenue. It is exactly the kind of optionality that a market frightened by capex does not have in the price.

03 · The AI card: Wang, "Watermelon" and what nobody has in the price

On July 2, in an internal meeting, the head of Meta Superintelligence Labs, Alexandr Wang — the former Scale AI founder Zuckerberg hired to lead the superintelligence effort — told employees that the house's next model, in training and codenamed "Watermelon," has caught up to OpenAI's GPT-5.5 on the sector's most-watched benchmarks. Wang added that this model uses "an order of magnitude more compute" than Muse Spark, the April model that performed well but did not match OpenAI or Anthropic.

If Wang's reading is true, it would be the clearest sign yet that Meta's brutal spending is starting to turn into a frontier product. And it is precisely the piece the thesis was missing: capex stops being a hole of faith and gains a verifiable result. A model competing with GPT-5.5 feeds the assistant better, improves ad ranking and can be rented.

One has to be honest about the nuances, though, because there is as much narrative as data here. These are internal benchmarks, unnamed and still without independent verification; the model has not been released; and OpenAI has already shown GPT-5.6. What is more: that same week, Zuckerberg himself adopted a cautious tone and admitted that AI-agent development has not accelerated "the way we expected." So: treat "Watermelon" as a valuable option, not a done deal. But an option that costs nothing at today's price and can be worth a lot is precisely what a patient value investor is looking for.

04 · The numbers: what it should be worth, if the thesis works

Let's leave the narrative and go to the valuation. The method I use is simple and transparent: project free cash flow per share a few years out, apply a reasonable exit multiple to get a future price, and discount that price at the annual return one demands to invest. The beauty is that each assumption is explicit and you can move it.

The starting parameters are not aggressive by whim. The base FCF/share (~$19.3) reflects normalized cash — the ~$52 billion after the investment peak — not this year's cyclical trough. Growth of 18% is well below the historical FCF CAGR (~20.6% over eight years). And an exit multiple of 28× is demanding but defensible for a business with these margins and this user base. With all that, fair value today comes out at ~$843 and the projected price at five years, near $1,240.

Reference price: $582.90 (2 Jul 2026). Model and calculation: trimmtrack.com/explore. Results depend entirely on the assumptions chosen and are not a forecast.

With the default values, the implied return runs around 16% a year buying at today's price: enough to justify patience. And do not forget the multiple context: with the stock's recent weakness, Meta trades at a forward P/E of about 20.5× and an EV/EBITDA of 14× — below its own 16× average and below the rest of the majors. You are not paying up for one of the best businesses of its size.

05 · The counterpoint: what has to go right for it to work

Being bullish is not being naive. The thesis has clear dependencies and they are worth saying out loud. First and most important: if infrastructure spending is not monetized, $125–145 billion a year triggers a fast and painful re-rating downward. Second: Reality Labs will keep burning cash — the CFO expects losses similar to the ~$20 billion of 2025 — and it is expensive optionality that still bears no fruit. Third: regulation on minors is advancing — Australia already bans social media for under-16s, and Malaysia, France and Portugal are close behind — which pressures long-term usage. And fourth, structural: 98% of revenue depends on advertising, which is cyclical and sensitive to the macro cycle.

I would add a fifth warning, the most recent and most honest: Zuckerberg himself admits AI agents are not advancing at the expected pace. The AI card is real, but it has not been played yet. Whoever buys Meta today must be clear that they are partly buying a promise.

Still, when I add up the ad engine that already works, a multiple that is not expensive, a capex peak that should normalize, and an AI and compute optionality the market is not paying for, the risk/reward balance tilts clearly upward. The consensus sees an excellent company not worth chasing here. I see one of the best franchises of its size, on sale because it is paying in advance for its next decade. My target remains around $1,200 over five years. The next catalyst — July 29 earnings — will tell whether the market starts to look at what Meta is buying, and not just the bill.

Sources

Reuters / Business Insider reporting on Meta's town hall (Alexandr Wang and the "Watermelon" model, 2 Jul 2026) and on the exploration of the compute-rental business (1 Jul 2026). Valuation model: trimmtrack.com/explore.

This article reflects a personal opinion for analysis and informational purposes. It does not constitute investment advice or a buy/sell recommendation; the author is not a financial adviser. Each investor must do their own due diligence.