Meta Platforms (META) Q3 2026 Earnings: Record $34.9B Quarterly Capex, Negative Free Cash Flow, and Muse Pushes Compute from Training to Inference

12 hours ago

9 min read

1. Executive Takeaway

・Core business is solid; the headline was dragged down by a settlement charge: Revenue was $64.60B, up 26.1% YoY, with ad impressions up 10% and average price per ad up 14%, both volume and price rising. GAAP operating margin was 20.1%; the company disclosed that excluding the $9.89B settlement charge, operating income would have been $22.88B, up about 11% YoY.

・Capex is still accelerating, and it is now eating into cash flow: Q3 capex (including finance lease principal) was $34.91B, and free cash flow was negative $2.12B. The full-year guide narrowed to $135–145B; with about $85.8B spent in the first three quarters, Q4 implies $49–59B in a single quarter.

・Muse changes the shape of compute demand: Muse is an extremely compute-hungry inference product. Management aims to cut unit compute cost by more than 10x next year while building capacity to serve over 1 billion people. Training is a fixed cost; inference is a variable cost that grows with users, and it is the latter that keeps consuming data center network bandwidth.

2. Revenue and Financials

・Revenue: Q3 2026 $64.60B; YoY +26.1%; QoQ +6.3%

・Gross margin: Q3 2026 79.3%; YoY -2.7pp

・Operating income / operating margin: Q3 2026 $12.99B / 20.1%; YoY -36.8%

・Net income: Q3 2026 $9.86B; YoY +264.0%

・Diluted EPS (GAAP): Q3 2026 $3.83; YoY +264.8%

Segment revenue: Family of Apps $64.16B, up 26% YoY; within it, advertising was $63.06B, up 26%, and other revenue was $1.10B, up 59% (mainly WhatsApp paid messaging and subscriptions). Reality Labs revenue was $443M, down 6%.

What the numbers actually mean:

・The 264% jump in net income and EPS is a base effect. A year ago Meta booked a one-time $15.93B valuation allowance on deferred tax assets, leaving net income at just $2.71B and EPS at $1.05. This year's $3.83 is actually depressed by the settlement charge; the company disclosed EPS of $7.40 excluding it. Both years carry one-off items, so take the YoY percentages with a grain of salt.

・The 2.7pp gross margin decline is infrastructure cost. Cost of revenue rose about 45% YoY, and depreciation and amortization jumped from $4.96B to $7.15B (about +44%). Management cited third-party cloud spend, depreciation, data center operating costs and "third-party AI token costs." That last item is worth noting: Meta trains its own models but is still buying inference from outside.

・The tax rate was 24% this quarter, above guidance, because part of the settlement charge is non-deductible.

Meta Platforms Q3 2026 earnings data card: revenue, gross margin, operating margin, net income, EPS and Q4 guidance

3. Capex Breakdown: Q4 Implies $49–59B in a Single Quarter

The statement most worth unpacking this quarter is not the income statement but the cash flow statement.

・Q3 alone: purchases of property and equipment of $33.94B plus finance lease principal of $0.96B, totaling $34.91B, a quarterly record. Operating cash flow of $32.79B no longer covered capex, and free cash flow was negative $2.12B (versus positive $10.63B a year earlier).

・Full-year guide: narrowed from $130–145B to $135–145B, raising the low end by $5B while keeping the top end unchanged.

・First three quarters: equipment purchases of $83.06B plus finance lease principal of $2.77B, about $85.82B in total. Q3 alone accounted for 40%.

・Q4 implied: $135B − $85.82B ≈ $49.18B; $145B − $85.82B ≈ $59.18B. Against $22.14B in Q4 2025, that implies roughly +122% to +167% YoY.

・Full year: 2025 capex was $72.22B; the 2026 guide implies roughly +87% to +101% YoY, nearly double.

Management is not hitting the brakes; it is setting Q4 up to be the biggest quarter of the year.

Where is the money going? Management describes it as servers, data centers and network infrastructure. Meta will not break out the networking share on an earnings call, but its public talks at OCP point in a clear direction. In Omar Baldonado on Meta's full-stack, open AI networking we covered why Meta treats the network as a standalone bet rather than an accessory to its GPU purchases.

With capex doubling and free cash flow turning negative, the next question is where the money comes from. That lines up exactly with the thesis of our Optical stocks fell on funding cost, not orders (2026 W42 Industry Review): demand has not cooled, and the real variable is whether funding costs start to bite into shipments once 2027 capex guidance comes out. Meta ended the quarter with $82.45B in cash and marketable securities against $80.92B of long-term debt; the two are now nearly even.

4. Muse: Compute Demand Shifts from Training to Inference

Muse is this quarter's headliner. It is a personal agent running on Meta's in-house Muse model family: the free tier defaults to Muse Spark, while the paid tier unlocks the flagship Muse Wonder.

The numbers management shared:

・About six weeks after launch, more than 10 million people in the US and Canada have set it up, and it has ranked No. 1 in the App Store almost every day.

・About 25% of people who set it up use it daily, more than twice where Threads started, according to management.

・The most common use case is shopping: more than a quarter of users engage in it, with an average purchase above $200 (including travel). Personal productivity (email and calendar connectors) and creativity follow.

But the cost structure behind those numbers is the real story. Zuckerberg was blunt: Muse is very compute-intensive, and compute efficiency is the main reason it has not launched globally. He set two goals:

1. Cut unit compute cost by more than 10x within a year, by using off-peak spare capacity to run smaller models, optimizing inference, and distilling Muse Wonder into smaller product models.

2. Build capacity to serve more than 1 billion people over the next few years.

These two points need to be read separately. A 10x cost reduction does not mean a 10x drop in compute demand. Management's logic is that lower costs are what allow the user base to scale, and once it scales, inference traffic grows linearly or even faster with the number of users. CFO Susan Li said it plainly: expanding the user base will require significant investment, and total serving costs will rise.

Training is a one-time fixed cost; inference is a variable cost that runs every day. Muse is pushing Meta's compute curve from the former toward the latter.

What this means for the optical supply chain: training clusters consume a small number of very large, tightly synchronized scale-up networks, while inference clusters need large volumes of distributed, flexibly scheduled scale-out capacity, plus the ability to shift training compute to inference during off-peak hours. The two workloads place different demands on the network, but both need more bandwidth.

5. In-House Models and the Enterprise Platform: Amortizing Training Costs Through Inference

Zuckerberg made a point this quarter with real supply-chain implications: training is a large fixed cost, so now that Meta has leading models, it wants to serve as much high-margin inference demand as possible. That is the rationale for the Meta Enterprise platform led by CJ Desai, whose near-term products are the model API, Muse Code and enterprise agents.

A few internal numbers also show the payoff of in-house models:

・After moving content understanding onto its own Muse model family, Meta's offline evaluations show better performance than the prior production models, with per-unit content processing costs down more than 80%.

・Engagement per person in Meta AI has more than doubled since April, when the new models started rolling out.

・Meta is using Muse Wonder in evals to let models autonomously improve its recommendation and ad systems.

Management said it has "not yet found the limit" of how effectively compute can be applied to the core business. In other words, Meta's compute has three destinations: the consumer Muse network, external enterprise inference, and internal recommendation and ad systems. If any one of them outperforms, compute will fall short, which is why management is comfortable promising aggressive investment without giving 2027 revenue guidance.

6. Reality Labs Shrinks, Class C Shares Arrive: Where the Money Comes From

Reality Labs gives up budget. Reality Labs posted a $4.13B operating loss this quarter (vs. $4.43B a year earlier). The full-year loss guide was lowered to $18–19B, with a target of about $15B in 2027. Zuckerberg said new VR devices will ship only when there is a major breakthrough, and the glasses business model will center on Muse subscriptions rather than hardware margins. The savings are explicitly going "primarily to AI."

Headcount is shrinking too. Meta ended the quarter with 66,775 employees, down 15% YoY, reflecting the roughly 8,000-person reduction in May.

A new non-voting Class C stock. The board approved a publicly listed, non-voting Class C common stock: each Class A or Class B share receives one Class C share, effectively a two-for-one split, with trading expected to begin in late January 2027. Stated uses include capital raising, employee compensation and acquisitions. The CFO said there are no specific financing plans, but establishing a framework for diverse funding sources was prudent.

A company whose free cash flow just turned negative, with a capex guide still ratcheting up, chose this moment to build an equity financing tool. It is preparing for 2027.

7. Management Outlook and Wording

Hard numbers:

・Q4 2026 revenue: $73–76B, implying roughly +22% to +27% YoY; about 1% FX headwind

・Full-year 2026 total expenses: $175–179B (raised to include the settlement charge)

・Full-year 2026 capex: $135–145B (previously $130–145B)

・Reality Labs 2026 loss: $18–19B

・Q4 tax rate: 15%–17%

・2027: no revenue or expense guidance until the Q4 call

Reading the wording:

・On 2027 revenue, management said its investments "position us to sustain strong overall revenue growth": an expectation, not a number.

・On 2027 investment, the phrase was "continue investing aggressively," while the budget was described as "very dynamic, still pretty early." Analyst Doug Anmuth asked directly whether earlier comments about maximizing compute in 2026 and 2027 and having more flexibility in 2028 still hold. The CFO did not answer directly and said to wait for Q4: no denial, no confirmation.

・The tone on monetizing Muse is noticeably more urgent than for any previous new product. Answering Evercore's Mark Mahaney, Zuckerberg said Meta will work on monetization sooner than its normal playbook, because both the costs and the opportunity are so large.

・Excluding the settlement charge, management still expects 2026 operating income to exceed 2025. That is a confirmation, and the highest-confidence statement on the call.

8. Supply Chain Signals: What It Means for Optics and Networking

Meta does not name suppliers on earnings calls, but several threads from this quarter connect to the supply chain map we track:

・Custom accelerators and scale-up networks. Meta's training and inference do not run on GPUs alone; its in-house MTIA accelerators need their own scale-up switching fabric. We broke down that path in Meta Tahan: an open, Ethernet-based scale-up switch for MTIA: Ethernet, open, and a rack design Meta defines itself. With capex doubling, the volume on this path scales up too.

・The procurement logic of optical interconnects. How Meta chooses between LPO, NPO and CPO is not only about technology, but about whether the supply chain can be multi-sourced and serviced. For context, see Meta's Nan Ge on building the NPO ecosystem, where Meta frames its NPO decision as procurement logic rather than technology preference, and Meta's LPO production reliability data from OCP.

・Third-party cloud and token costs. Management listed third-party cloud spend and third-party AI token costs as expense growth drivers, meaning Meta's own capacity has not caught up with demand and some workload is spilling over to cloud providers, which also supports their capex.

・Personal Agent Protocol. Meta is developing an open standard for interactions between personal agents and business agents with partners including Sierra, Walmart, Stripe and Shopify. If agent-to-agent transactions scale, they become a new source of machine-to-machine traffic, a longer-term demand variable.

9. Risks and Counterarguments

・Negative free cash flow: operating cash flow no longer covers capex. If 2027 capex grows sharply again, Meta will have to rely on debt or equity financing, and funding costs will directly affect the pace of investment.

・Legal and regulatory: beyond the $9.89B already recognized (the agreement provides for payments of up to about $18.59B), management noted several youth-related trials scheduled in the US in the coming months.

・Muse unit economics: the 10x cost reduction is a target, not an accomplished fact. If efficiency cannot keep pace with user growth, compute spending will surge ahead of revenue.

・User metric definitions: 10 million refers to people who set up Muse, not monthly actives; 25% is the daily-use share among those who set it up. Retention needs more time to prove out.

10. Summary

Meta's income statement looks ugly this quarter because of one settlement charge, but volume-and-price growth in the core business and Q4 revenue guidance of $73–76B show the ad engine is fine. What has really changed is where the cash goes: a $34.9B quarter of capex, negative free cash flow, a Class C stock that paves the way for equity financing, and an inference product meant to serve a billion people.

Three metrics to track next quarter:

1. Whether actual Q4 capex lands in the $49–59B range, and the 2027 capex figure given on the Q4 call.

2. Whether Muse expands beyond the US and Canada: the most direct evidence of improved compute efficiency.

3. Whether free cash flow stays negative, and whether any actual capital raise follows once Class C shares list.

Bottom line: Meta's ad machine is still printing money, but it can no longer pay the AI bill on its own, and 2027 capex guidance will be the single most important number for the entire optical supply chain.

This article is for technology and industry trend analysis only and does not constitute investment advice.