META 2Q26 Earnings Update

See my writeup from last July where I laid the groundwork on this thesis.

Subscribe to Stock Thoughts for free to receive new posts

TLDR

Meta’s advertising revenue grew 27% in Q2, down from 33% in Q1 and up from 24% in Q4. Total revenue grew 28%, but operating income fell 8% as costs rose 55%. The operating margin went from 43% to 31%.

The core business is performing. Meta is using LLMs throughout its recommendation and advertising systems. The new business-agent product also has a logical path into WhatsApp paid messaging and the existing advertiser base.

The harder part is sizing the rest. Meta did not quantify the compute devoted to frontier-model training, the revenue opportunity for business agents, the timing of the LLM-native recommendation rebuild, or which new AI product should produce measurable returns first.

A year ago, I rated Meta’s five bets as advertising 10/10, engagement 9/10, business messaging 8/10, Meta AI 3/10, and AI devices 5/10. The first three continue to look like extensions of the existing business.

The quarter

The headline numbers were:

• Revenue of $61b, up 28%.

• Operating income of $19b, down 8%.

• Operating margin of 31%, down from 43%.

Costs and expenses rose 55% to $42b. That included a $2.4b legal charge and $1.2b of severance from the May headcount reduction. Operating income would have increased 9% excluding those items. On that basis, cost growth was about 42% and the operating margin was nearer 37%.

The remaining cost growth is still substantial. R&D rose 67% to $21.7b and accounted for 58% of the total increase in costs. Headcount was down 1% year over year and 3% from Q1. The increase occurred despite fewer employees.

Li listed employee compensation, infrastructure costs, legal costs, and third-party AI token costs as the drivers. Wells Fargo estimated token costs were about $2b of the year-over-year increase.

Capex including finance leases was $31.1b. Cash from operations was $31.9b, leaving free cash flow of $0.8b. Meta bought back no stock in the first half. Dividends were the only capital returned.

Long-term debt rose from $59b at year end to $84b, including $25b issued during Q2. Net cash fell from $23b to about $7b.

Meta narrowed its 2026 capex guide from between $125b and $145b to between $130b and $145b. Li said the change reflected three more months of visibility rather than a change in plan. Management did not provide 2027 capex guidance.

Advertising is performing

Advertising revenue was $59.4b, up 27%. Impressions rose 14% and price per ad rose 12%.

The recent progression:

• Q3 2025: +26%

• Q4 2025: +24%

• Q1 2026: +33%

• Q2 2026: +27%

Asked about the Q3 guide, Li named three factors behind the implied deceleration (~3-4pts at the midpoint, CC): lapping a quarter of accelerated impression growth, the fully rolled out less personalized ads offering in Europe, and continued integrity enforcement.

Meta added $12.8b of advertising revenue year over year. Alphabet added $10.3b of Google advertising revenue and Amazon added $4.1b of advertising-services revenue. Zuckerberg described Meta as producing faster dollar growth than any other reported advertising business.

Advantage+ end-to-end solutions passed a $75b annual revenue run rate. Nine million small businesses used at least one AI creative tool. Meta said its user-understanding models, sequence learning, and the GEM ads ranking model increased ad clicks by 8% and conversions by 16% on Facebook.

The new piece this quarter is Meta Generative Recommender, the first generative model deployed into ads retrieval. Li described it this way:

“Rather than scoring every possible ad individually, we are now using LLMs to reason about ad content and user preferences together, and predict the best ad for each person.”

Other revenue crossed $1b for the first time, growing 73%. Li said the increase was driven primarily by WhatsApp paid messaging and subscriptions.

LLMs in the recommendation stack

Meta now processes every public Reels and Feed post on Instagram through an LLM. The model analyzes topics, tone, and other characteristics. Those signals then feed ranking, recommendations, and content-policy enforcement.

Meta also shipped its largest single-release ranking improvement to date on Reels. The release increased Instagram sessions by 15 basis points. More than half of recommended Instagram Feed content is now less than one day old, more than double the level a year ago.

A typical recommendation stack works in stages. It retrieves candidate posts, scores them with several models, reranks the finalists, and then mixes in ads, safety, diversity, and freshness.

Meta is trying to give the model more information about what each piece of content means. Li explained the change:

“Part of this work involves developing an LLM native data infrastructure where we shift from using, for example, content IDs for ranking, which have a limited amount of information, to a more descriptive and meaningful semantic ID system that allows our models to just understand more about each piece of content and infer why it’s interesting to someone.”

A semantic ID can describe what a post is about and its tone, rather than just identifying it. I’d imagine this boils down to the model being able to decide that the next recommendations should be humorous beginner-golf instruction, detailed investing analysis, and a local restaurant video. Meta would then match those outputs to either organic or ads content in its inventory.

The longer-term proposal is a unified model:

“Eventually, we hope to get to a place where we can collapse our current multi-stage recommendation model to a simpler unified model that, similar to how LLMs do next-token prediction, can directly produce a set of tokens to show users, which we can then match to either organic or ads content within our inventory.”

Li separately described building foundation models “designed to power organic content and ads recommendations simultaneously,” another piece of the roadmap.

In the 2025 piece, I asked whether compute and talent devoted to recommendations could be moved into ads, Meta AI, or another workload. Meta effectively says it is designing the recommendation model so that the same training work improves engagement and advertising. This unified system is still in early validation. Management gave no date for replacing the existing multi-stage system.

There is also a separate use of compute for frontier-model training. Zuckerberg said:

“A substantial amount of the compute goes towards training models to be a leading lab, and I think that’s an important investment.”

Meta did not quantify “substantial.” That compute may eventually produce a consumer product, an API, developer tools, or another revenue source. No stated revenue line currently supports it. This remains closer to the 3/10 Meta AI bet from last year than the 10/10 advertising bet.

Business agents

Meta Business Agents became available globally on WhatsApp and Messenger during the quarter and are rolling out on Instagram Direct. More than one million businesses were active as of June, and Meta plans to start charging in 2H.

The product connects to catalogs, customer relationship systems, and inventory systems, then completes actions inside a conversation. Meta’s example was Movida, a Brazilian car-rental company with nearly 400 locations. Its WhatsApp agent handles vehicle selection, pricing, and payment.

Movida reported that over a one-month period its daily WhatsApp bookings increased 44% year over year and 85% of conversations were resolved without a human.

This is a direct extension of the business-messaging thesis from last year. Meta already has the advertiser relationships, the messaging surfaces, and the paid-message infrastructure.

Per-token pricing takes effect August 1, with one token charge covering both AI processing and message delivery. Meta is also testing the agent inside its Meta One subscription on a freemium basis: a limited number of free messages, then a subscription for higher limits.

The longer-term model could resemble advertising. Zuckerberg said Meta expects to move toward charging when the agent produces a result and eventually run an auction over compute. Meta would allocate agent capacity toward the customer interactions with the highest expected value, much as the ad system allocates impressions.

I cannot size it from the information provided. One million active businesses is a usage figure. Meta gave no price, revenue, take rate, or average usage.

Zuckerberg also acknowledged that enterprise software is “a somewhat different muscle than we have historically had.”

Capex, returns, and what 2028 is about

Meta’s 2026 capex guide is $130b to $145b. Capex including finance leases was $31.1b in Q2 and free cash flow was $0.8b. Over the last twelve months, capex including finance leases was 41% of revenue.

Management’s case is that the company remains supply constrained and still has “numerous ROI positive places” to deploy compute in the core business. The advertising and recommendation results support part of that claim. The unanswered question is how much of the spending goes toward those measurable uses and how much goes toward frontier-model training and products that have not yet produced scaled revenue.

What Meta did lay out is the shape of the commitment: maximize capacity in 2026 and 2027, secure land and power for 2028, and defer some chip decisions until it has more information.

“When we think about planning today for ‘28, we’re really focusing on flexibility. That’s just giving us kind of the ability to have land and power. But to really make the actual decisions about buying chips and other big-ticket items further in the future.”

Land, power, data-center shells, and network foundations take longer to secure. Chips can be ordered later. Meta is therefore committing to the long-lived infrastructure while delaying part of the shorter-lived equipment decision until it has more information about demand.

Management’s confidence declines with time. Li said Meta feels confident that it has uses for the compute being built in 2026 and 2027. For 2028 and beyond, she said Meta’s “hope” is that it is building enough capacity to absorb the expected high-return uses. By then, Li said, Meta will have “turned over a lot of cards,” and the new products and internal workloads will have more history.

The joint ventures are another part of that flexibility. Meta announced a venture with BlackRock to develop a one-gigawatt data center in El Paso, following a similar structure with Blue Owl for the Hyperion campus in Louisiana. Li said the partnerships allow Meta to reevaluate its compute needs every four years, while its backstop obligations decline over time.

They also move a meaningful portion of the commitment outside Meta’s reported long-term debt. The Louisiana venture has roughly $27b of estimated development costs, funded pro rata by its members, with Meta holding 20%. Meta has $12.3b of lease commitments beginning in 2029 (making the first evaluation point 2033) and residual-value guarantees with a maximum threshold of $28b. As of June 30, it disclosed maximum exposure to loss of $46b for that venture and $6.4b for its other variable-interest entities.

Li said operating cash flow remains the primary source of funding, but Meta is adding long-duration capital to lower its cost of capital.

What management quantified, and what it did not

Brian Nowak asked which opportunity would scale first and produce quantifiable, material ROIC in 2026 or 2027. Zuckerberg listed training, core optimization, consumer products, APIs, business agents, developer tools, and possible compute sales, then said he was “quite optimistic that we’re going to see meaningful growth in all of these areas.” He did not rank them or attach a size or a date. Li took the second half of the same question, on 2027 capex: “we aren’t providing a specific outlook for 2027 CapEx at this time. Infrastructure planning remains highly dynamic.”

Shweta Khajuria asked for a framework for the size of business AI. Management gave product detail and the one-million-business figure.

One new line did get a start date. Tom Champion asked how the business agent will be monetized, and Li gave the mechanics: subscriptions and volume-based token pricing, with per-token charging effective August 1.

Zuckerberg was direct on where personal agents stand. Mark Shmulik asked about consumer adoption. Zuckerberg said billions of people are likely to have personal agents within five years, that the product has to just work, then added:

“We’re going to ship this at some point soon and that’s going to be very exciting. And we haven’t done that yet.”

The multiple

In July 2025, I wrote that Meta traded at 27x LTM earnings, 22x earnings excluding Reality Labs losses, and 36x LTM free cash flow.

Reported earnings no longer support that comparison. The last twelve months carry a $15.9b one-time tax charge from the second half of 2025 and a $5.0b tax benefit from the first quarter of 2026, which puts the trailing effective rate at 22% against a 15% to 17% guide. Meta’s reported 2025 diluted EPS was $23.49. Wells Fargo and Morningstar both carry 2025 at $29.68 adjusted, and consensus is on the adjusted number.

Operating income avoids the problem, because it sits above tax and still carries the depreciation the build is creating (though that will continue to ramp as well). Over the last twelve months it was $86.9b, or $90.5b once the $2.4b legal charge and the $1.2b of severance are added back. Those are the only one-time operating items Meta disclosed in the period.

At $556.71 on July 31 that puts Meta at 15.8x adjusted operating income. On 2027 estimates it is 14.2x using Wells Fargo’s forecast and 12.4x using Morningstar’s. Adding back the Reality Labs loss as well takes the trailing figure to 13.0x.

Free cash flow is where the spending shows up, and it has gone the other way: 38x, against 36x last July, because capex including finance leases is now 41% of LTM revenue.

I cited a 15% to 25% historical range last year, and the bottom of that was already stale. The last year Meta ran at or below 15% was 2015, on $18b of revenue. Over 2017 to 2024 the range was 16% to 28%, with a mean and median of 21% and a decade low of 16% in 2021. On that basis, 2025 at 36% was already the highest capex intensity in Meta’s history as a public company, and the last twelve months ran higher still.

At a 21% intensity, holding revenue and everything else constant, free cash flow would be about $82b and the free-cash-flow multiple about 17x rather than 38x.

This changes the emphasis from last year. The Reality Labs operating loss was $16.1b in 2023, $17.7b in 2024, $19.2b in 2025, and $19.1b over the last twelve months. It grew 8% in 2025, all of it in the fourth quarter, and the first-half 2026 loss then fell 1% while revenue increased 7%.

Even more now than then, capex is a materially larger valuation variable (not to mention the R&D). Adding back the last twelve months of Reality Labs losses, tax-effected, is worth about $16b of earnings. Normalizing capex from 41% to 21% is worth about $44b of cash. In dollars, capex is roughly three times the lever Reality Labs is.

Some of that capex is improving the ads and recommendation systems today. Some preserves capacity for products that are not yet scaled. Some funds frontier-model training without a stated revenue path.

A year ago, Reality Labs was still a big piece of the discussion. Today the loss runs at about $19b a year and is disclosed every quarter, and the main uncertainty is infrastructure spending that crosses several businesses and is not separately disclosed. The gap between 15.8x on operating income and 38x on free cash flow is capex.

The quarter did not answer how much of that spending supports the proven core business, how much supports unscaled products and frontier models, or when the latter will produce revenue.

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

---