What Does Meta Do? The Ads-Funded Hall Model
Meta converts advertising cash into captive AI halls: it gets paid when ads run, while the megawatts that would light those halls are still the stuck step.
Meta converts advertising cash into captive AI halls: it gets paid when ads run, while the megawatts that would light those halls are still the stuck step.
HYPERSCALERS · Data as at 29 Jul 2026 · Next update: after Meta Q3 2026 results (21 Oct 2026)
Affects: Hyperscalers · US · Neoclouds
It is a hyperscaler on this layer: a buyer large enough to pull chips, halls and power. Unlike Google, Amazon or Microsoft, Meta does not print a merchant-cloud remaining-performance-obligation. The earn is ads. The spend is halls. See AI capex and cash capex.
Education only, not advice to buy or sell any security. Figures are from company filings and company releases dated as shown. Outlook is labelled outlook. Secondhand reports are labelled reported.
Numbers that matter
| Figure | Date | What it tells us | |
|---|---|---|---|
| Advertising revenue | $59.363bn of $60.801bn | Q2 2026 · 29 Jul 2026 | Who gets paid |
| Family of Apps / Reality Labs | $60.370bn / $431m | Same 8-K | Mix |
| Operating income | $18.775bn (31% margin) | Same 8-K | Earn, after the spend |
| Capex incl. finance-lease principal | $31.08bn | Same 8-K | Cash that left |
| Free cash flow | $784m | Same 8-K | Almost all OCF went to halls |
| Construction in progress | $80.345bn | 10-Q, 30 Jun 2026 | Unfinished halls |
| Uncommenced leases | $278.99bn (+$68bn in July) | Same 10-Q | Clock, not this quarter’s cash |
| 2026 capex guide | $130–145bn | Outlook, 29 Jul 2026 | Not a print |
| El Paso campus | 1 GW · ~$14bn · online 2028 | Company, 28 Jul 2026 | Named hall clock |
Sources: Meta Q2 2026 results, Form 8-K exhibit 99.1, 29 Jul 2026; Form 10-Q for the quarter ended 30 June 2026; Q2 2026 earnings-call transcript (company), 29 Jul 2026. Status chips are Second Order’s read. The $130–145bn range is outlook.
Our arithmetic. Ads were 97.6% of sales (59,363 ÷ 60,801 = 0.976). Family of Apps was 99.3%. Capex including leases was $31.078bn (30,116 + 962). Operating cash flow $31.862bn less that capex is $784m, matching the release. Year-ago capex on the same basis was $17.012bn (16,538 + 474), so this quarter’s print is +82.7%. First-half capex was $50.918bn (49,113 + 1,805). A $137.5bn midpoint on the $130–145bn guide leaves about $86.6bn for the second half (137.5 − 50.918 = 86.582).
What does Meta actually do?
Three steps, in order.
Earn. Light. Use.
-
It earns on ads. Family of Apps is Facebook, Instagram, Messenger, WhatsApp and other services. In the quarter ended 30 June 2026, advertising was $59.363 billion, other Family revenue $1.007 billion, Reality Labs $431 million. Ad impressions rose 14% and average price per ad rose 12%. Daily active people averaged 3.60 billion in June, up 3%. Li said Advantage+ reached over $75 billion in annual revenue run-rate. That invoice does not wait for a hall to switch on. See AI capex.
-
It lights halls — on the balance sheet and off it. Q2 purchases of property and equipment were $30.116 billion; finance-lease principal $962 million. Servers and network assets were $119.683 billion; construction in progress $80.345 billion, up from $50.521 billion at year-end 2025. Uncommenced leases were $278.99 billion at 30 June, plus $68 billion signed in July for 2027–28 commencement. The Louisiana campus (Hyperion / Beignet in the 10-Q) is a 20% stake in about $27 billion of development, with leases from 2029, an initial commitment of about $12.31 billion, and residual-value guarantees of about $28 billion. On 28 July 2026 Meta and BlackRock announced a second 80/20 venture: a 1 GW El Paso campus, about $14 billion, Meta the initial sole occupant, online in 2028. See data-center leases and who finances GPU clusters.
-
It uses the compute itself — and may sell some later. Zuckerberg said a significant portion of compute goes to training, the core business, and personal agents, and that Meta is getting “a lot of offers for compute at a significant premium.” He also said there is “nowhere near enough compute for all the demand,” and that selling intelligence should earn more than selling the hours. Li is maximising 2026 and 2027 capacity and laying foundations for 2028 and beyond. No Q2 line is merchant GPU-as-a-service. See GPU as a service for the product Meta is not printing.
| Step | What it means | Where the risk sits |
|---|---|---|
| Earn | Advertising on Family of Apps | Price per ad, impressions, legal and policy hits |
| Light | Cash capex, leases, 80/20 campuses | Delivery versus the $130–145bn spend; watts |
| Use | Ranking, training, agents; optional compute sales | Halls that are not yet online |
How does Meta make money?
It gets paid when an advertiser is billed for an impression or a click. Three things decide how much of that cash becomes a live rack:
- The ad print — 97.6% of this quarter’s sales. Reality Labs lost $4.619 billion on $431 million of sales; Family of Apps operating income was $23.394 billion.
- How fast halls energise — Li: “You obviously are not getting value out of them until they’re online.”
- Whether compute is used inside or sold — spoken on the call, not this 8-K.
On 29 July 2026 Meta printed $60.801 billion of revenue, up 28%. Operating income was $18.775 billion, down 8%; the margin was 31%, from 43% a year earlier. Costs included $2.40 billion of legal charges and $1.18 billion of severance. Li said that excluding those two items, operating income would have increased 9% year over year. Our check: 18,775 + 2,400 + 1,180 = $22.355 billion, versus $20.441 billion a year earlier (22,355 ÷ 20,441 = 1.094). That add-back is ours, not a GAAP line.
Net income was $15.848 billion. Cash and marketable securities were $90.26 billion. Long-term debt was $83.664 billion after $24.910 billion of May notes. Restricted cash included $10.80 billion escrowed against infrastructure purchases, due for release between 2028 and 2030.
Who gets paid. Meta, on the ad. Chip, server, construction and power vendors, on the $30.116 billion of property-and-equipment purchases. Landlords and bondholders, on the lease and venture structures. Money sticks at energised, accepted megawatts — not at a capex headline and not at an unfinished $80.345 billion of construction in progress. See cash capex and which capex lines to watch.
Q3 revenue was guided to $61–64 billion. Full-year 2026 expenses $165–169 billion. Full-year operating income still expected above 2025. Those three are outlook.
What limits it?
Watts first. Unfinished halls second. A compute-sales story that has not printed third.
-
A spent dollar is not a live megawatt. Zuckerberg: “there’s just nowhere near enough compute for all the demand.” Li is maximising 2026–27 capacity because “existing capacity, including our own, [is] extremely valuable.” Construction in progress of $80.345 billion is the unfinished side of that bind. See speed-to-power and the US market door.
-
Leases signed are not commencement. Uncommenced leases of $278.99 billion, plus $68 billion in July, run through 2036. The Louisiana leases commence in 2029. El Paso is expected online in 2028. Meta is the initial sole occupant there; it is not this quarter’s advertising invoice. See data-center leases.
-
Power is billed before it flows. Amazon, Google, Meta and Microsoft signed Duke Energy’s North Carolina large-load terms: new loads of 50 MW or more fund their own substation and pay at least 75% of contracted demand, used or not. The commission is expected to rule by mid-November 2026. See the Duke Wire.
-
Off-balance-sheet halls still have a clock. The Louisiana venture is 20% of about $27 billion, with ~$12.31 billion of leases and ~$28 billion of residual-value guarantees. The Hyperion bond Wire is the market print; the 10-Q is the company print. Do not treat a bond price as Q2 cash.
-
Capex is the cash cost of the same bind. Free cash flow of $784 million on $31.862 billion of operating cash flow is the lag in one line. Ads fund the build. Watts gate it.
The honest uncertainty is conversion. A $59 billion ad print, an $80 billion unfinished-hall print, and a 2028 El Paso clock are different facts. Meta has no Cloud remaining-performance-obligation to convert; see RPO, explained for the peer mechanism.
Who else does this?
Roles, not a league table.
| Company | What it does | What makes it different |
|---|---|---|
| Meta | Ads-funded captive halls; optional future compute sales | Ads 97.6% of sales; no Cloud RPO; Q2 capex $31.08bn; FCF $784m; 1 GW El Paso online 2028 |
| Google / Alphabet | Broad cloud; converts Cloud remaining performance obligations; sells TPU systems | Cloud RPO $513.9bn; Cloud +82% to $24.8bn. See What Does Google Do? |
| Amazon | Broad cloud; converts AWS remaining performance obligations | AWS is the merchant line. See What Does Amazon Do? |
| Microsoft | Software plus Azure; converts commercial remaining performance obligations | Azure is the merchant line. See What Does Microsoft Do? |
| Oracle | Cloud infrastructure plus database; prints MW and GPU utilisation | Delivery of accepted capacity. See What Does Oracle Do? |
| CoreWeave and other neoclouds | Specialist AI compute, often debt-funded | One product; overflow when hyperscaler queues are long. See What Does CoreWeave Do? |
What separates them is who the customer is — an advertiser inside Meta’s apps, or a cloud tenant — and who funds the hall. Meta sits on the captive-capacity seat. See neocloud vs hyperscaler.
What would change this view?
The read on the model weakens if:
- Q3 2026 (21 Oct 2026) prints capex well inside the remaining ~$79–94bn implied for the second half, or cuts the $130–145bn year
- Construction in progress stops rising while Li’s “maximise 2026 and 2027 capacity” language disappears
- El Paso slips past a 2028 first-online, or the Louisiana 2029 lease-commencement clock moves out
- Duke’s North Carolina order (mid-November) is rejected or rewritten so the 75% floor does not hold
- A dated print of material merchant-compute revenue appears — the model would then be closer to Google Cloud
- Advertising growth rolls over while halls keep consuming cash — the funder, not the watt, would then be the bind
- Next GPU attach is late because packaging or HBM is short — the stuck step would then sit in semiconductors
It stays intact while ads remain ~98% of sales, free cash flow stays thin against $30bn-plus quarterly capex, construction in progress stays large, named campuses stay 2028–29 clocks, and watts remain the layer’s stuck step.
Practical takeaways
- Meta’s model on this layer is Earn → Light → Use: ads pay, halls are stood up, compute is used internally
- Q2 2026: $60.801bn sales, $59.363bn ads (97.6%), $18.775bn operating income, $31.08bn capex, $784m free cash flow
- Construction in progress $80.345bn; uncommenced leases $278.99bn plus $68bn in July
- 2026 capex guide $130–145bn. Our arithmetic: ~$86.6bn implied in H2 at the midpoint
- Named clocks: 1 GW El Paso, online 2028; Louisiana leases commence 2029
- Stuck step: watts and unfinished halls, not the ad invoice
- Next hard read: Q3 2026 results on 21 Oct 2026
FAQ
What does Meta do?
Meta runs Facebook, Instagram, Messenger and WhatsApp, earns almost all of its revenue from advertising, and spends that cash — plus debt and off-balance-sheet leases — to stand up AI data halls for ranking, training and agents. It is a hyperscaler on this layer because of that spend, not because it sells a merchant cloud.
What does Meta do in simple terms?
It sells ads, builds or leases giant computer halls, and runs its own models inside them. The ads print now. The halls light later.
Is Meta profitable?
Yes on a GAAP operating basis: $18.775 billion of operating income on $60.801 billion of revenue in the quarter ended 30 June 2026 (29 Jul 2026). Net income was $15.848 billion. Free cash flow was $784 million after $31.08 billion of capital expenditures including finance-lease principal. That is not a forecast.
How is Meta different from Google Cloud or AWS?
Google and Amazon convert remaining performance obligations into recognised cloud revenue as halls and chips light. Meta’s Q2 invoice is advertising — $59.363 billion of $60.801 billion. Zuckerberg has talked about selling compute later; that is not this quarter’s line. See What Does Google Do? and remaining performance obligation, explained.
What is Meta’s 2026 capex?
The company anticipates $130–145 billion of 2026 capital expenditures including principal payments on finance leases, narrowed from $125–145 billion (outlook, 29 Jul 2026). Q2 printed $31.08 billion. First-half printed $50.92 billion on the same basis.
Should I buy Meta shares?
We do not make buy or sell calls, and nothing here is a recommendation. What we can tell you is what the company does, how it gets paid, and which figures would change the picture. The decision is yours.
Keep reading on the map
- The layer: Hyperscalers — capex pull and the delivery bind
- The term: Hyperscaler, explained
- The mechanism: AI capex · Cash capex · Speed-to-power · Data-center leases
- The contrast: Remaining performance obligation · Neocloud vs hyperscaler
- The peers: What Does Google Do? · What Does Amazon Do? · What Does Microsoft Do? · What Does Oracle Do?
- The Wire: Duke’s 75% minimum bill · Hyperion bond at 91 cents
- Where it happens: US · PJM capacity short
The members cut on the long clock is already live: Where does the stuck step move?. This page stays the free Stack read of the Meta model.
How we check this
Figures come from company filings and company releases, each dated below. Capacity language, Advantage+ run-rate, and the compute-sales comments are the company’s spoken print on the 29 July 2026 call. El Paso figures are from the 28 July 2026 company release. Duke terms are from the named utility filing, via our Wire. The Hyperion bond price is a market print on the Wire, not a Meta 8-K line. Arithmetic is shown in full. Outlook is labelled outlook.
Last reviewed: 16 Oct 2026 · Next review: after Meta Q3 2026 results (company, 21 Oct 2026) print capex against the $130–145bn guide
Spot an error? Tell us; we will correct it here. Editorial standards →
Sources
- Meta, Q2 2026 results (Form 8-K exhibit 99.1), 29 Jul 2026 — $60,801m revenue; ads $59,363m; FoA $60,370m / RL $431m; operating income $18,775m (31%); net income $15,848m; capex incl. leases $31.08bn; FCF $784m; legal $2.40bn; severance $1.18bn; Q3 revenue guide $61–64bn; FY expenses $165–169bn; FY capex $130–145bn
- Meta, Form 10-Q for the quarter ended 30 June 2026 — PPE net $225.724bn; CIP $80.345bn; servers $119.683bn; uncommenced leases $278.99bn + $68bn in July; commitments $349.31bn; Louisiana 20% of ~$27bn, leases ~$12.31bn from 2029, RVG ~$28bn; restricted cash $10.80bn; held-for-sale $2.03bn for El Paso
- Meta, Q2 2026 earnings-call transcript (company), 29 Jul 2026 — “nowhere near enough compute”; offers at a premium; maximise 2026–27 capacity; Advantage+ >$75bn run-rate; 9 million SMBs on AI creative tools
- Meta / BlackRock, El Paso 1 GW venture (company), 28 Jul 2026 — ~$14bn; 80/20; Meta initial sole occupant; online 2028; Meta contributes ~$2.3bn CIP, receives ~$1bn distribution; BlackRock ~$4.9bn cash; $12.5bn debt financing
- Meta, Q3 2026 results date (company), 12 Oct 2026 — print 21 Oct 2026
- Duke Energy settlement, via Wire: 75% minimum bill, 6–7 Oct 2026
- Hyperion / Beignet market print, via Wire: bond at 91 cents
All external links rel="nofollow noopener".