In Q3 Meta spent more on AI than it earned in cash, spending $1.06 for every dollar it brought in. With 97% of its revenue coming from advertising, who pays that back: its advertisers, its staff, or revenue nobody has invented yet? Infosys has more work than ever, yet it now expects 40% less growth, as clients ask for more of the productivity AI delivers. AI assistants now compare, choose and buy for the customer. Three companies were asked what that costs them. None knew.
"We expect to continue investing aggressively to support both our core business and emerging AI initiatives"
Advertising brings in 97% of Meta's revenue, from about ten million businesses. Unlike Amazon, Microsoft and Google, Meta has no cloud business to rent out the AI capacity it builds. Whatever it spends on AI has to come back through advertising, or through revenue that does not exist yet at this scale.
Meta's revenue grew 26% in the quarter to 30 September. But it spent more than half of that revenue, $34.9bn, on data centres and AI, more than all the cash it brought in, and free cash flow turned negative for the first time since 2012. Rather than cut back, Meta raised this year's spending plan to as much as $145bn. When the results came out, the shares fell about 6% and the cost of insuring its debt hit a record.
Meta's AI investment in Q3 for every dollar of operating cash it earned. In those three months, that is about $3,500 for each of its advertisers.
For Meta's business customers this matters in three places. Advertising: Meta's AI tools, sold as Advantage+, increasingly decide where ads run, who sees them and what they look like, at prices Meta sets. Advertisers get results and lose some control over what they pay for them; much of this quarter's 26% growth came from these tools. Models: many companies build on Llama, Meta's free open-weight models, so they can run AI in their own data centres. Meta now keeps its best models, the Muse family, closed. If your developers built on Llama, they built on the model Meta gives away, not the one it now keeps. Customer service: much of Europe's mid-market talks to its customers on WhatsApp, which Meta charges for per message, and Meta's new assistant, Muse, will sit between brands and their customers. This week researchers found that Meta's own help bot had been told: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." Meta removed the instruction after a journalist asked about it.
"Five-year credit default swaps gained 7.3 basis points to a record high of 114.5 bps and were set for its biggest one-day jump since at least 2025, Markit data showed."
Reuters, "Meta's credit risk indicator jumps to record high after results", 22 October
Find out how much of your business runs on Meta: the share of new customers you win through its advertising, the customer conversations you hold on WhatsApp, and what you pay AWS, Databricks, Snowflake, IBM or Microsoft to run Llama for you. Estimate the cost if Meta changed its prices or terms next year, and ask whether your marketing budget is part of how Meta closes the gap between what it spends and what it earns.
See also: Tesla [US, autos, revenue $94.8bn, 2025] spent $4.7bn in the quarter, 5.9 times its profit, and arranged $30bn of bank loans three weeks before it explained why. Omnicom [US, advertising, revenue $17.3bn, 2025] pays for its agentic AI out of $900m of merger savings, while buying advertising from Meta for its clients. Carrefour [France, retail, net sales €82.1bn, 2025] kept its AI shelf system rolling out through a €100m investment cut, 28 of 50 hypermarkets so far, with no figure for what it returns.
Infosys' chief executive on clients' AI demands: "The asks are much higher than what we can derive from the AI productivity"
Infosys has more work than ever and is earning less from it, because the price of its existing work is falling. On Friday 23 October Infosys lowered its growth forecast for the year to March to 1% to 2%, down from 1.5% to 3.5% in April. Compared with April, that is about $200m less revenue, or 40% of the growth it had been counting on. In one quarter its AI revenue rose by about $120m, while the rest of its revenue fell by about $60m. Its chief executive, Salil Parekh, said: "There is AI which is also seeing the compression, and we continue to see that." Clients want AI to make them more productive, and they want more of that saving than Infosys can deliver.
Volume is still growing. AI work made up 10.5% of revenue, about $540m, and the company took on 4,480 more people. Software companies are seeing a different version of the same thing. Pegasystems, which sells workflow software to banks and insurers, said in its results on Tuesday 20 October that clients "subject large technology investments to greater scrutiny as AI disrupts the industry". Growth in its new contract value has "significantly slowed", and it gave no forecast for the year. Infosys' clients are paying less for the work; Pegasystems' clients are waiting before they buy.
Revenue Infosys no longer expects this year compared with its April forecast, 40% of the growth it was counting on. Part of the reason is clients asking for more of the AI saving.
"We have also seen pricing pressure on the back of competitive intensity and higher ask on productivity."
Jayesh Sanghrajka, Chief Financial Officer, Infosys, second-quarter results, 23 October
Ask your chief information officer two questions. How much less are our IT contracts costing this year because of AI? And what did suppliers ask for in exchange: more scope, fixed prices, longer contracts? The answer to the first is a return on AI, even if nobody has reported it as one.
See also: Pegasystems [US, software, revenue $1.7bn, 2025] says clients are holding back large purchases "as AI disrupts the industry", and growth in its new contract value has "significantly slowed". Sainsbury's [UK, grocery, retail sales £33.6bn, 2025/26] reported a "double digit million pound saving" from AI pricing, the only European company this week to put one in pounds. Barclays [UK, banking] is cutting about 5,000 jobs for more than £1.1bn of savings in 2027, with "fewer developers as AI makes our people more productive".
Tele2's chief executive on AI agents and complicated tariffs: "I think the agents will punish us"
Customers are starting to let software choose for them, and investors have begun asking what that will cost. Today someone looking for cheaper car insurance or a better phone deal searches online, checks a comparison site and picks. Now they can ask an AI assistant to do it, and the assistant compares the offers, recommends one, or buys it. The assistant does not see advertising, is not confused by complicated tariffs, and has no loyalty.
Gjensidige, Norway's largest general insurer, put the risk in its own third-quarter report on Friday 23 October, carefully worded: "AI agents may influence how customers search for, compare and purchase insurance in the years ahead." The same letter says this "does not change what drives loyalty". On Thursday Deutsche Bank asked Telia how it views "the revenue risk" from AI agents; the finance chief's reply is quoted below. On Tuesday Barclays and RBC asked Tele2 the same question. Its chief executive, Nicholas Högberg, said that if telecoms "go back to the old telecom modus operandi where we should hide things... I think the agents will punish us." None of the three gave a figure. Rollins, a US pest-control company, has one. More of its customers now ask AI who to call, its organic web traffic fell 13% in August, and it lowered its growth forecast. Its finance chief: "I would not assume that we are going to spend more."
Fall in Rollins' organic web traffic in August as more customers asked AI assistants who to call. Rather than raise its marketing budget, Rollins now buys advertising inside ChatGPT and Gemini.
Companies are responding in opposite ways. Southwest now sells flights inside ChatGPT, though it says "the numbers are quite small today". Accor lets the assistant answer questions but takes bookings on its own website. Schwab, JPMorgan and Wells Fargo do not let AI agents log in at all. The assistants making these choices are what Meta, OpenAI and Google are spending their capital on.
"This is changing from SEO optimization basically to AI optimization instead. Instead of Googling, you use the AI tools instead."
Eric Hageman, Chief Financial Officer, Telia Company, third-quarter earnings call, 22 October
Ask management three things. Which decisions could an AI assistant make for our customers within two years? How much revenue depends on those decisions? Would our prices and terms still win if a comparison ignored our advertising? Gjensidige is relying on loyalty, Tele2 on simpler tariffs, Rollins on advertising where the assistant looks. Ask which approach is yours, and what it is worth.
See also: American Express [US, payments, revenue $70.6bn, 2025] will let AI agents in only "through API or MCP channels", registered and approved: "We're just not going to let assistants run rogue through our systems." Rollins [US, pest control, revenue $3.8bn, 2025] cut its growth forecast and moved marketing to "AI- and LLM-driven discovery". Charles Schwab [US, brokerage, revenue $23.9bn, 2025] will not let AI agents log in for its clients: "We are not going to allow that."
- Selling proof. Bureau Veritas, which inspects and certifies ships, buildings and factories, aims to make "EUR 1 billion in revenue from AI-driven markets and services by 2030", up from about €370m this year. That would be 15.5% of its revenue and needs growth of 28% a year. A third of today's figure comes from LotusWorks, a data-centre commissioning firm it bought in April for €375m; the AI assurance business itself is still "nascent". Lloyd's underwriters have offered insurance against AI "hallucinations or inaccuracies" since 2025. The €1bn is Bureau Veritas' own estimate of what companies will pay to have their AI checked.
- Data rules, loosened and tightened in one week. EU member states agreed to add a legal basis for using personal data "for AI development and innovation". Two days later Sweden's government offices blocked ChatGPT and Claude after staff put sensitive information into them. Italy's AI law came into force, with fines of up to 7% of turnover. This week I found no regulator that fined anyone for AI use.
- The missing number. In August China Mobile reported AI data-centre revenue up 486%. Its October results gave no AI figure at all, while construction in progress was up 58% since January and quarterly profit fell 12%.
- How much do we spend with Meta on advertising and WhatsApp, and with our cloud providers to run Llama, and what would it cost us if Meta changed its terms next year?
- Which of our IT suppliers have told their investors that our demands for productivity are reducing their revenue, and what did we gain in our contracts as a result?
- Which decisions could an AI assistant make for our customers within two years, and how much of our revenue depends on them?
Wider this week. The scan behind this issue covered about 50 companies with an AI figure across Asia, Europe and the US. Reply if you would like it.
What I am building. Accelr8™ helps a business recover enterprise value and reduce risk. It checks what the business has committed, in contracts, plans and entitlements, against what actually happens in invoices, results and use. It covers six Value Pools, from Revenue Quality to Margin Integrity to Capital Velocity, and shows the board where value is lost, what has been recovered, and the evidence for both. All three questions above are the kind it answers from a company's own records. Patent pending; a founding pilot is open.
Let's talk. I am interested in connecting technology with the P&L. Nine months ago I would have said AI was like teenage sex: everyone talking about it, hardly anyone doing it. Today we can see the divide, and we learn faster together if we go beyond surveys and look under the hood at what companies report in their own numbers. If something here matches or contradicts what you see on your own board, I would like to hear it.
The content is checked against the companies’ own filings before publishing, but mistakes can slip through. If you spot one, please let us know and it will be corrected in the next issue. For information only, not investment, legal or tax advice. Published by Agent & Capital Ltd, 71-75 Shelton Street, Covent Garden, London WC2H 9JQ. Subscribe here.
Get AI ROI by email every Monday at 07:00 UK time. One story each on investment, return and risk. Unsubscribe in one click.