Part one: what goes away
Most AI data will sit there like microns in an ocean droplet. Trillions of data points, generated, scraped, summarised and stored, each one technically available and practically useless: no source anyone can open, no method anyone can read, no record of ever being right, and no one accountable for it. Most of it will be made only to be sold to another AI, and most of the time it will fail even at that, because the buyer can generate the same summary for nothing. It yields nothing to the economy, nothing to the people who own it, and nothing to the data centre it inhabits except a bill for the power, the cooling and the disks that keep it alive. It answers everything and helps no one decide anything. A brand cannot build a program on it, a banker cannot put it in a model, and an assistant cannot cite it, because there is nothing behind it to cite. It is not an asset. It is a cost with a storage address.
That changes what an answer is worth. When every assistant can produce one in a second, an answer on its own is worth nothing, and the value moves to the line that can be checked: the fact with its source, its date and a link anyone can open. The same pressure falls on every method kept behind a curtain. A proprietary score no outsider can test was a moat when the data was hard to get; it becomes a liability the day a published method, run on the same public filings, gives the same answer for free. Google's own guidance already tells publishers that the way into its AI answers is to be indexed, plain and quotable, and the largest research houses now list, in their own annual reports, the risk that third parties will use their data with AI to make their own insights.
Reach goes the same way. For twenty years brands bought traffic, impressions and "people reached", counted by the platform that sold them. Most counted traffic is now automated, and the house's own counters show it plainly: about one in four counted desk views is a reader at a home or an office. Once buyers can see that, they stop paying for numbers the seller grades itself. And when the numbers can be seen, so can the seams in the old way of serving a brand: the research firm that does not build, the agency that does not measure, the broker that never meets the brand and the event producer hired separately, each handoff leaking and each markup now visible. Personal data, the fuel of much of that machine, is under steady legal pressure, and measurement is moving from what people can be tracked doing to what companies actually report.
The proof, in their own numbers
None of this rests on opinion. The companies, the regulators and the grid operators have already written it down.
Follow the money first. Amazon spent $215.2 billion on capital in the twelve months to September 2026, against $120.1 billion a year earlier. Alphabet raised $80 billion of new equity in June 2026 to pay for "AI compute infrastructure" and expects to spend $180 to $190 billion this year. Meta spent $72.22 billion in 2025 and now expects $135 to $145 billion for 2026. In the annual reports filed with the SEC, the phrase "artificial intelligence" appeared in 558 documents in 2020 and in 3,914 so far in 2026; "generative AI" went from 69 in 2023 to 1,207. Inside the companies, the money comes out of the old lines: Klarna cut its external agency spend by a quarter and credited AI with 37% of its marketing savings.
The agencies' own accounts show the other side. WPP ended 2025 with 98,655 people against 108,044 a year earlier, revenue down 5.4% like for like and a £641 million impairment, £393 million of it on Ogilvy, and a plan to fold itself into a "single company". Dentsu wrote off ¥396.1 billion of goodwill, lost ¥327.6 billion and paid no dividend. Omnicom bought Interpublic and is consolidating to survive, on track for $900 million of savings this year. The giants that are still growing are growing on AI, data and technology, not on the old agency trade. Publicis grew 5% this quarter by combining "data, technology and talent" and expects the industry to settle around "three global players"; Omnicom's growth runs through its Omni platform, built on "comprehensive identity and data infrastructure", the very asset this piece shows turning into a liability, while it exits real estate; and Estée Lauder cut 31 agencies to one. The traditional agency is not being saved by this. It is being absorbed into three data companies, or closed.
The tangible brands still growing tell the same story from the other side: they are the ones people consume, wear or carry out of a store the same day. Coca-Cola's volume rose 5% and McDonald's plans about 2,500 openings this year; Tapestry and Ralph Lauren raised their outlooks on store sales, and Estée Lauder now calls the store its media. What is going away is the tangible that sold nothing on its own: the printed piece, the mailed flyer, the branded giveaway that could never show what it earned.
The data glut has a bill, and it is landing on the grid. Data centres used 4.7% of US electricity in 2024 and are projected to use up to 11.8% by 2030, by the Lawrence Berkeley National Laboratory's report for the Department of Energy. In Virginia, Dominion Energy's contracted data-centre capacity went from about 16.5 gigawatts in July 2023 to about 53.8 gigawatts in July 2026, and on the PJM grid nearly all of the rise in forecast peak demand now comes from data centres. Every data point stored and never used is paid for here, and much of it is paid for with borrowed money: Moody's says its fastest-growing ratings line this quarter came from "AI-related financing and hyperscaler investments", Meta has created a new class of stock to raise capital, and Snowflake sold $3.5 billion of convertible notes. What is borrowed has to be earned back.
And success is already measured in moments. Walmart's advertising grew 38% in its latest quarter, Walmart Connect 43%; Amazon took $19.6 billion of advertising in a single quarter, at the shelf, in the moment of purchase. The checkout itself is moving to agents: Stripe and OpenAI put Instant Checkout inside ChatGPT on the open Agentic Commerce Protocol, Google built its Universal Commerce Protocol with Shopify, Etsy, Wayfair, Target and Walmart, and Visa says agent-initiated payments will reach "mainstream adoption in 2026". Underneath it, nine federal agencies have named the Legal Entity Identifier the standard for who a company is, from 1 October 2026.
Who falls from the top
Palantir said it plainest, in its letter to shareholders this quarter. Every industry, from banks and insurers to hospitals and the old law firms, will be "remade or set briskly aside". The outcomes will be binary, winners and losers with nothing in between. The middle becomes a wasteland of former leaders who did not listen, and the bottom fills with those who once had an insight or an edge and let it go, their fate already sealed. Palantir can say it that harshly because its own numbers show it happening: in the third quarter of 2026 its US commercial revenue reached $984 million, up 148% on a year earlier, and passed its US government revenue for the first time.
Follow that same pressure through the industries this note touches and the verdict is the same. A business is remade around what can be checked and delivered, or it is set aside. The middle empties first, because it sold the layers a buyer no longer needs, and the bottom fills with those still selling the cheapest copy of what a machine now does for nothing.
The first to feel it are the businesses that sold access to public facts. Seat-licensed research and data terminals were built when filings were hard to read at scale; now the facts are public, the arithmetic is published, and an assistant will not cite a paywall. S&P Global, FactSet, Morningstar and Gartner each name AI, and third parties using their data, as competitive risks in their latest 10-Ks. Next to them sits analyst opinion sold as a ranking: a placement decided by opinion loses ground to a placement decided by filed numbers anyone can check, and Gartner's own report expects "more competition with increased adoption of AI services".
Content built only to catch traffic goes next. Assistants answer the question before the click, and a page that only rephrases what is already known is never the one cited; Anthropic's citations, for instance, return the exact passage that supports a claim, so the page with the passage wins and the page with the paraphrase disappears.
Then the traditional agency, in every form it took. That is long gone, and none of the quality workers want anything to do with agencies that gouged clients an hourly rate for "services" that never resulted in anything. The list sellers are already gone: their rented names fed the advertising mail that has fallen by almost half since 2007, and their trade in personal data ran into the privacy laws state by state. The marketing-services shop that executed campaigns by hand is going the same way, its work now drafted, scheduled and reported by a model in an afternoon. So is the design studio that sold layouts, decks and templates a model now produces in seconds, in a hundred variations, for a few cents, and the media buyer whose margin lived in the gap between what it paid and what it billed, a gap now closed by automated buying and by clients who can see the auction. So are the expensive CGI firms and advertising sleuths billing hundreds of dollars an hour for a thirty-second clip that a model now renders in minutes. Above them the holding companies are consolidating to survive: Omnicom completed its merger with Interpublic in November 2025, and it measures itself by its own revenue growth and expenses, not by what happens to its clients' filed results. A few will be remade around measured results and accountable delivery. The rest are closing, or already closed.
Success is now measured in microns of moments, within trillions of data points engaging with each other for perceived success. Traditional methods are gone.
Commerce moves too. Gifting and promotional platforms, social listening tools and AI platforms all grew in the gray area of the budget, the line nobody had to defend. That gray area is gone completely. Gifting and promotional platforms that grew by holding the buyer's data now meet a new kind of buyer: an AI agent acting under a signed mandate, which wants a structured feed and a quote, not a marketplace that keeps its customer. And the data they held has turned from an asset into a liability. Anyone holding a buyer's names, addresses, orders and gift lists now carries the breach, the lawsuit and the headline that come with them, and the old agencies and networks that passed that data from hand to hand, with no one checking who held it next, are exactly where the damage starts. The house's own papers on the House Edge have traced it: the small supplier who sinks a billion-dollar mark over a weekend, the vendor who becomes the client's headline, the agencies that bill the handoffs, and the chain of custody most houses cannot even name. That is dangerous, and buyers now know it. Social listening sold as brand success measures what people say rather than what the brand and its peers actually earned, and that gap is now visible beside the filed numbers. And AI platforms sold without proof of results are the most exposed of all; C3.ai's latest 10-K, with revenue down, a restructuring and a change of chief executive, shows what happens when a platform has no public record to point to.
Part two: what the future keeps
What survives all of this is what can be checked and what can be done.
It keeps the checkable line: one fact, its source, its date and a link anyone can open, the line an assistant cites, a banker pastes into a model and a brand builds on. It keeps published methods, the measures the top investment houses and VCs already trust, run in the open on the actual audited numbers so that nobody has to trust a secret. It keeps a record that cannot be back-dated, forecasts sealed before the outcome and graded after it, misses included, because time is the one input no competitor and no model can buy.
It keeps success found where it actually happens. The filings say what worked; the launches, campaigns, deals, retail and social moves, events and hires behind the brands that win say how it worked, and the house reads both every day and merges them, so the record and the story check each other. It keeps identity and signatures as well: the Legal Entity Identifier becoming the identifier of record, verifiable credentials on filed facts and signed ledgers, so that who said what, and when, can be proved.
The leaders' own outlooks this season say the same thing in their own words. The brands are cutting their own layers, not only their agencies': Procter & Gamble is removing up to 7,000 overhead roles and Nike expects about $2.5 billion of savings, mostly in people. A client with fewer people to manage vendors needs work that arrives finished, measured and easy to check. Every outlook now carries a tariff line, Tapestry builds a tariff rate into its guidance and PepsiCo books refunds, so a program has to be planned on the landed cost, not last year's price. And the shopper is stretched: McDonald's credits an "intervention on value" for its recovery and Berkshire's businesses report "sluggish customer demand", so what a program earns has to be proved at the price the shopper actually pays, not in how many people it reached.
It keeps trust that can be shown rather than promised. Visa now calls trust "the next growth vector" and is paying $2.4 billion for fraud intelligence. A successful house shows its trust: a named chain of custody, a dated record, and an answer when something goes wrong.
It keeps the client's data where it belongs, with the client. Microsoft says it helps organizations "build and own their own intelligence", and WPP's clean room lets partners work together without moving the data. The rising house works inside the client's data and leaves it there; it never copies it out, keeps it or claims it.
It keeps the agents that are built on real data. Every company is now building AI agents to read, decide and buy for it, and an agent is only as good as what it reads. An agent fed on summaries acts on guesses, at machine speed, in the company's name. The agents that last read the primary record itself, the filing, the price, the order, the signed mandate, cite the line they acted on, know exactly who the counterparty is from its legal identifier, and are measured against what actually happened afterwards. Anthropic's own platform returns the exact passage behind every claim, the Model Context Protocol lets an agent call a source directly instead of scraping a copy, and the payment networks now require a signed mandate before an agent can spend. Amazon now ships rules to "prevent autonomous agents from causing unintended harm", and Microsoft has made agents a reporting segment of their own. Building agents that work with real data, and that can show their working, is now the job, and the limits come first: what the agent may read, buy and sign, and who answers when it errs. It is the job any successful house has to do for its clients.
Above all it keeps accountable execution. An AI can tell you who is winning; it cannot run your event, make your pieces or answer for the result. The future belongs to the house that turns what works into the brand's own program, makes it, and measures it again when the next filings land, and that can be hired by a person or by an agent: a feed an agent can read and a quote it can accept under a signed mandate. The payment networks are building the rails for it: Mastercard paid $1.4 billion for stablecoin infrastructure and Visa runs stablecoin settlement of its own. A successful house takes the agent's order, settles on whatever rail the signed mandate allows, and keeps the record.
The short version: the future keeps what can be checked and what can be done, and it discards what can only be summarised.
Why the old ways are now dangerous, and how fast they sink a brand
Everything above says the old ways are going. The harder truth, and the subject of the house's next piece, is that while they last they are dangerous. The old networks were built to pass things along: a buyer's list to a broker, a brief to a subcontractor, a file to a factory, a customer record to a vendor's vendor. Nobody checked who held it next, and nobody was asked to. That pass-through was the business model, and it made every brand in the chain complicit in whatever the weakest link did.
The record already shows where that ends. In the SEC filings the house reads every day, companies now report the damage themselves. On 8 September 2026 Veradigm told the SEC that "one of its third-party vendors experienced a cybersecurity incident that impacted certain data associated with a small number of the Company's customers"; six days later CenterPoint Energy Resources reported that an unauthorized third party had obtained personal information of a portion of its customers. The house's own Breach Clock and cyber-disclosure boards hold every one of these, filing by filing, each linked to its source. The regulators have gone after the data trade directly: the Federal Trade Commission has finalized orders barring X-Mode and its successor Outlogic, Gravy Analytics and Venntel, and Mobilewalla from selling sensitive location data, and proposed a ban on InMarket selling precise location data. And the House Edge papers have traced what it does to a brand: the small supplier who sinks a billion-dollar mark, the vendor who becomes the client's headline, the publicist and producer whose conduct a search engine cannot tell from the brand's own.
What has changed is the speed. A brand used to have days to answer for a bad link: a news cycle, a statement, a quiet change of supplier. There is no room any more for the questionable old media monger, angry and posting vitriol, while accessing your company's records at the same time; they are now filed below the bottom feeders. Now the first sentence anyone reads about a brand is written by a machine, from whatever stood closest to its name, the moment someone asks. A breach in a vendor's vendor, a supplier's scandal, a partner's post: each is read, joined to the brand and repeated in every answer at once. Brands no longer sink over days. They sink in microns of a second, and the old networks are where the hole is.
That is why the house works the other way: a checked network small on purpose, a chain of custody it can name, blind shipping so suppliers never hold the buyer, no personal data in any score, and a record that shows what it said before the outcome. The next piece is about exactly that: why the older ways and approaches are dangerous and no good. The creepy pass-throughs that handed a buyer's data on and never asked who held it next, the complicity of every network that looked away, the lack of action when it went wrong, and how all of it now sinks a brand in microns of a second, shown with first-hand proof and the real record.
---
The full edition, with the leaders' own filings, the complete data and every source: What goes away, and what the future keeps: the full edition.