Two ad businesses of the same size
For most of the past decade, Google Search was the bigger advertising machine by a comfortable margin. That gap has closed. Meta's Q3 2026 results put ad revenue at $63.1 billion, up 26% year over year. Alphabet's Q3 2026 results put Google Search at $63.4 billion, up 12%. The difference in absolute terms is a rounding error on a $63 billion line item.
The difference in growth is not. Fourteen percentage points separates the two growth rates, and that gap matters more than the totals. A business growing at 12% and a business growing at 26% can post the same total today and look very different four quarters from now.
What ten million advertisers actually used
Meta does not sell tokens. It sells impressions, and the AI tools it shipped into Ads Manager this year are there to make more impressions easier to buy. The company says nearly 10 million advertisers tried at least one of them in the quarter, which covers generative creative variants, campaign copy and automated audience work.
That figure deserves a second look. "Tested at least one tool" is a low bar, and Meta has not published retention, spend lift or how many of those ten million kept using the tools after the first campaign. The number proves distribution, not value. What it does show is that Meta can push a new model into an existing sales channel overnight, which is something no standalone AI vendor can do.
A sharp market reaction, and Meta's October slide
Investors read the quarter as a search-monetisation problem, and Alphabet's shares fell sharply in a single session. Meta did not escape either. Its shares also fell during October as investors went through the capital expenditure lines. The same AI that improves targeting is expensive to run, and the market is no longer willing to treat that spending as free.
The capex arithmetic, in euros
Full-year 2026 AI spending is projected at up to $205 billion for Google and up to $145 billion for Meta. At roughly $1.16 per euro, the annual figures are about €177 billion and €125 billion; per quarter that is about $51 billion and $36 billion (≈€44 billion and €31 billion). Comparing annual capex with annualised ad revenue gives an uncomfortable ratio:
| Company | 2026 AI capex (up to) | Q3 ad revenue | Annualised ad revenue | Capex as share of annualised ad revenue |
|---|---|---|---|---|
| Alphabet / Google | $205bn (≈€177bn) | $63.4bn | ≈$254bn | ≈81% |
| Meta | $145bn (≈€125bn) | $63.1bn | ≈$252bn | ≈58% |
That ratio is rough. Alphabet's capex funds Cloud, Search, YouTube and DeepMind research, not just the ad auction, so 81% overstates the burden on search specifically. The direction is still clear: Meta is spending about 58 cents of every annualised ad dollar on infrastructure, and advertising accounts for 98% of its total revenue. There is no cloud business to absorb a bad quarter.
Europe: labels on AI-made ad creative
This is where the European picture differs from the American one. Since August 2026, Article 50 of the EU AI Act applies across the single market: synthetic content must be disclosed, and machine-readable marking is expected to travel with generated images, video and audio.
In practice, an agency that generates a product visual with Meta's tools and runs it as a paid ad inside the EU has to label it and keep the provenance metadata intact. That is not a marketing problem, it is an engineering problem. Watermarks and metadata routinely break when an asset is cropped, resized or re-encoded for a placement. Anyone running a generative image pipeline for EU campaigns has to test that survival, asset by asset, before the creative goes live. The rules do not shrink the ad market. They add a compliance step per file, and they apply to Meta's internal tools exactly as they apply to a developer calling an API.
Meta gives the model away
The strategy behind the numbers is unusual. As European Business Review has described, Meta's open-source strategy is not purely altruistic: by releasing its LLaMA models for the open-source community, it fosters an ecosystem of innovation that feeds back into its own products and advertising business. The licence is not unlimited — companies with more than 700 million monthly active users must obtain special permission — but Meta does not primarily aim to monetise access to the model itself.
For anyone paying API bills, that asymmetry is visible: Meta's openly released LLaMA models carry no per-token price; you host them under the licence. The return arrives through the auction, not through a per-token invoice. If you want to know what hosting an open-weight model costs on real hardware rather than on a price sheet, our AI Arena benchmarks cover exactly that on an RTX 5060 Ti 16 GB.
The quarter settled one question and left another open. Meta has caught Google Search in scale and is growing faster, according to the reported figures. Whether that growth covers $145 billion a year in infrastructure is a calculation the next few quarters will have to make. Advertising was 98% of Meta's revenue in Q3, so there is no large non-advertising cushion to absorb a bad quarter.
Does Meta charge separately for its AI advertising tools?
No. Meta has not published a per-token or per-asset price for the generative tools inside Ads Manager. The cost appears as ad spend, not as a line item on an invoice, which is why the tools are hard to compare directly with an API.
Do AI-generated ads need a label in the EU?
Yes. Article 50 transparency obligations have applied since August 2026. Synthetic image, video and audio content shown to EU users has to be disclosed, and machine-readable marking is expected to stay attached to the file through editing and delivery.
Can a company self-host Meta's open-weight models?
Meta releases its LLaMA models under a licence with restrictions; companies with more than 700 million monthly active users need special permission. Self-hosting the model does not replace the ad system or the EU labelling duty.
