The Talented Mr. Ripley

THE NUMBER: 95%. That’s the discount Meta (NASDAQ: META) will give you on its new model if you let it train on what you type into it. Read it the other way and it’s a price tag. The gap between the private tier and the data tier works out, by Tomasz Tunguz’s math, to $1.24 per million tokens, roughly $454,000 a year for a business running a billion tokens a day. That’s not a discount. That’s Meta telling you, to the penny, what your data is worth to it. Hold that number next to a simple question every business owner should be asking this fall: if the AI is nearly free, what is it actually charging me?

You didn’t hire Brock Purdy

Every lab wants you to picture its AI as Brock Purdy. Mr. Irrelevant, the 262nd and last pick in the whole draft, the kid who carries the clipboard until the starter goes down, then steps in and turns out to be humble, coachable, and better than anyone had any right to expect. That’s the story they’re selling. Your AI is the eager backup who was only ever waiting for a chance to help. Give it the reps and it’ll make you look good.

It’s a lovely story. It’s the wrong movie.

You didn’t hire Brock Purdy. You hired Tom Ripley. In Patricia Highsmith’s 1955 novel, and in the Minghella film everybody remembers, a shipping magnate hires an agreeable young man to sail to Italy and bring his son Dickie home. The young man is charming, useful, quick. He studies how Dickie dresses, how he spends, how he signs his name. And then, on a small boat off San Remo, he stops being the help and becomes Dickie, forging the signature well enough to cash the checks. “I’d rather be a fake somebody than a real nobody,” Ripley says, and it’s the truest line in the picture, because the whole grift is that the assistant was never trying to help. He was learning the part.

That’s the frame for the entire AI business right now, and it’s the thing nobody’s naming while they argue about benchmarks. The eager, brilliant, unbelievably useful assistant you’ve invited into your business is studying how you sign your name.

Not all data is equal

Start with the thing that changed this week, because it’s the tell. Meta shipped Muse Spark 1.3, a model good enough that Zuckerberg called it “frontier performance almost too cheap to meter” and Alexandr Wang put it at parity with Claude Fable and ahead of GPT at coding. Set aside whether that’s true (almost nobody actually at the frontier is going to switch to it). Look at how they priced it. Two tiers. The Standard tier is $1.25 per million tokens in, $4.25 out, and Meta never trains on you. The Contributor tier is ten cents in, twenty out, a 92 to 95 percent cut, and in exchange Meta keeps the right to train its next model on everything you send. Nobody else has made the barter this explicit. Tunguz’s read is the correct one: the spread between the two prices, $1.24 a million tokens, is Meta’s revealed valuation of your data. For a company pushing a billion tokens a day, choosing privacy is a $454,000 annual surcharge. That is the number, printed on the menu.

Here’s what makes it matter, and it’s the second-derivative point I want you to hold onto: not all of your data is the same kind of data, and the difference is the whole game.

When you type a query into Google (NASDAQ: GOOGL), you’re handing over intent. You want a running shoe; Google learns you want a running shoe and shows you one. That’s a fair trade, honestly one of the best trades in the history of commerce. You got the answer, they got the intent, everybody’s square. When you scroll Meta or TikTok, you hand over attention and a little of your psychology, and you get dopamine back. A rougher trade, but still a trade. You knew the terms.

Now point a frontier model at your actual business. Let it run your collections process, or draft your pricing, or triage your customer complaints, or build the model your analysts use. Ten thousand times. What did you trade? Nothing. There’s no consideration coming back the other way, because the thing it’s learning isn’t a query you can walk away from. It’s how you work. Your margins. Who your customers are and what they’ll tolerate. The specific, hard-won, un-Googleable knowledge of how your shop actually runs. You didn’t sell it any of that. You taught it, for free, and you paid for the privilege in tokens.

Consumer data is barter. Business data is unilateral disarmament. And the men buying it know exactly which one they’re getting.

The ad company always wins

Why is Meta the one printing this price? Because it can afford to. An ad business that already turns human behavior into $130 billion a year can lose the entire inference margin and come out ahead, because the data it collects on the way is worth more to its real business than the tokens ever were. The 95 percent “discount” isn’t generosity and it isn’t a loss leader. It’s Meta paying you for your data with free intelligence, and because it monetizes data better than almost anyone alive, it can outbid every pure-play model company in the world.

That’s the part OpenAI and Anthropic can’t match. They don’t have a $130 billion ad engine underwriting the giveaway. So they can’t win a data-price war against Meta or Google on the merits, which is going to force them somewhere else, and I’ll come back to where.

Google is running the same play with the lights off. We learned this week, from a study of two million product listings across a hundred thousand searches, that Google’s AI Mode shows shoppers the same products at 21.6 percent higher prices than plain search, and when the two disagree, AI Mode is the pricier one 68 percent of the time. It surfaces four options where normal search showed twenty-eight. Only 1.28 percent of the products in classic search even reappear in the AI answer. Same product, same query, same day, and it routes you there through Chrome without mentioning it.

I don’t think Google is individually sniffing out your personal willingness to pay, the way Delta uses AI to find the ceiling on a fare. I think it’s doing something simpler and more durable. It deleted the comparison. For twenty-five years Google’s whole business was showing you everything and letting the cheapest option win, and that comparison is exactly what protected the shopper. AI Mode replaces the aisle of twenty-eight with a hand-picked four, and the cheapest doesn’t make the cut. The consumer surplus that used to come from comparison shopping just became somebody’s margin. And the receipt is in the earnings: Alphabet’s Search ad revenue grew 17 percent last quarter to $63.3 billion, even as clicks per search fell. Fewer clicks, more money. That only happens if each search got worth more. The AI didn’t get you a better deal. It quietly removed the thing that used to.

Two ad giants, one asset. Meta prices your data explicitly, on a menu. Google prices it implicitly, in the twenty-four products it stopped showing you. Both of them have spent two decades learning to turn you into money, and both of them are better at it than any lab will ever be.

Ripley goes to work

Which brings me back to the labs, and to where the pressure sends them. If intelligence is trending to zero (we made that case Wednesday, and I won’t relitigate it here) and they can’t win the data-price war against the ad companies, then selling tokens is a dead business on a long enough timeline. So they do what Ripley did. They stop being the help.

The move has a name now, and one of your own crowd said it out loud this week. A sharp markets account posted, plainly, “Obviously the labs are going to roll out outcome-based pricing and revenue share by vertical.” Read that twice, because it sounds like a pricing footnote and it’s actually the whole thesis. The moment a vendor stops charging you for tokens and starts charging for your outcome, two things happen. First, you stop caring about the token bill, which is the point. And second, and this is the Ripley part, they have to measure your outcome to charge for it, which means they’re now studying whether they can produce your outcome without you. Getting paid on the result is how you learn the business from the inside. And once you’ve learned it, the tokens were never the product. You were the market.

We’ve watched this exact movie before, twice, from the two companies best at it. Amazon (NASDAQ: AMZN) built a marketplace, watched what sold, and launched Amazon Basics on top of its own sellers, using their demand data to pick the winners to undercut. AWS did the software version, taking open-source projects like Elasticsearch and standing up managed copies that ate the companies that built them. Apple (NASDAQ: AAPL) does it so routinely it has a nickname: “sherlocking,” from the day Apple watched a popular Mac app called Watson succeed on its platform and simply built the feature into the OS. The platform watches what works on top of it, then becomes the thing that works.

The labs are already on the conveyor belt, and coding is where it’s furthest along, because coding is where the outcome is cleanest to measure (the code runs or it doesn’t). Every lab now ships its own coding harness, Codex from OpenAI, Claude Code from Anthropic, that competes directly with the third-party tools built on the same models. This week Anthropic open-sourced an entire commerce-agent blueprint, the scaffolding a retailer would use to build a shopping agent, given away free. Ask yourself why a company gives away the blueprint for the exact vertical it’s studying. It’s the same reason Amazon gave sellers a great marketplace. You learn the business by hosting it.

Anonymity protects records, not process

So what do you do about it, if you’re the business owner and not the lab? The instinct is right: pay for anonymity, or opt out and run open weights, which are cheaper anyway. But there’s a catch in it that decides who’s actually safe, and it’s worth being precise about.

Anonymity protects records. It does not protect process.

If your edge is a number, you’re fine. Jane Street, Two Sigma, Renaissance Technologies, the great quant shops, can hand a model a stream of trades with every identifier stripped out, because the asset is the number and the number anonymizes cleanly. They pay top dollar for frontier AI and sleep at night, protected by moats a lab can’t cross: their capital, their execution, their decades of proprietary data, the sheer institutional apparatus. Their real risk isn’t that Anthropic copies them. It’s the macro one, because those three reflexive workloads, AI research, coding, and trading, are roughly half of frontier inference revenue, and the quants are helping fund the very loop that could unwind on them. Different risk. Keep it separate in your head.

But if your edge is how you work, you cannot anonymize the pattern, because the pattern is the identity. A mid-market distributor can strip its company name off every document it feeds a model, and the model still learns exactly how a distributor like that runs, prices, and keeps its customers, because it’s executing the process ten thousand times. The name was never the valuable part. The process was, and the process is precisely what can’t be hidden from a system that’s doing it. So “just anonymize it” is a real shield for the number and a leaky one for the operator, and most of the economy is the operator.

That leaves running your own open-weight models, which is genuinely cheaper on inference and genuinely private. It’s the right move for a lot of shops. Just know it isn’t free either: you inherit the harness, the ops, and the awkward fact that the best open weights right now are increasingly Chinese, which trades a data-exposure risk for a supply-chain-and-trust one. Worth it for many. Frictionless for none.

The company of free data isn’t the one you think

There’s an obvious counter, which is that some frontier company will just become the DuckDuckGo of AI, the one that promises never to touch your data, and win the trust the others are burning. I’d watch for the opposite, and I’d watch Grok.

xAI can afford to give the model away and take all the data, not because it’s private, but because it isn’t. It has an empire behind it, X for the social graph, Tesla for the fleet, SpaceX for the compute, all of it hungry for exactly the data flowing through the model. That’s not the DuckDuckGo of AI. DuckDuckGo’s whole promise was we don’t take your data. Grok’s move is we take all of it, and we don’t need to charge you, because the empire already pays. That’s the Gmail of AI, free and excellent and total. And it proves the earlier point at the largest possible scale: a model company can’t outbid an ad company for data, and it definitely can’t outbid a conglomerate that has a dozen other businesses starving for the same inputs.

Which means the real privacy play, the actual DuckDuckGo of AI, was never going to be a frontier lab at all. It was going to be the thing that keeps your data out of the model in the first place.

The two seats Ripley can’t take

Here’s where it lands, and long-time readers will see the old destination coming, so let me be honest that it’s an old destination reached by a new road. There are exactly two seats in this whole business that never ask for your secrets, and they’re the two we keep coming back to.

Apple sells the first one. The bet Apple is making, and getting mocked for being late on the frontier, is that for the 99 percent of the world doing bounded work, a good-enough model that runs on the device and keeps your data on the device beats a brilliant one in someone else’s cloud. Privacy as a hardware feature, not a policy promise you have to trust. There’s nothing to harvest if the data never leaves the phone. Ripley can’t study a signature he never gets to see.

SpaceX sells the second one, through Colossus and the compute it rents to the labs themselves. The landlord gets paid whether OpenAI wins or Anthropic wins or Meta wins, and it never needs to know what’s running through the racks. Neutral compute takes its rent regardless of whose business gets eaten upstairs. It’s the one position in the trade that’s indifferent to the outcome, which is the same thing as being immune to it.

Neither one needs your data, so neither one can go downstream and become you. That’s not a coincidence, and it’s not nostalgia for a call we’ve made before. It’s the logical end of the whole argument. In a world where the free thing is the reconnaissance and the smart money is buying your business one outcome at a time, the only safe places to stand are the ones that were never in the data business to begin with. The phone in your pocket, and the shovel under the whole thing.

The rest of it, the free tiers and the outcome-based pricing and the brilliant assistant that does your work better than your last three hires, is exactly as generous as it looks. Which is to say, it’s Ripley on the boat off San Remo, being enormously helpful, right up until the moment he isn’t.

You didn’t buy a tool. You let Ripley in. And Ripley never hands the job back.