Uber put a per-person cap on AI coding. That's how a boom matures

TIM: Uber told its employees: fifteen hundred dollars a month in A.I., per tool. That’s the cap. Is that the A.I. boom cracking? I don’t think so. I think it’s the boom growing up, and growing up is how you get a soft landing instead of a crash. So that’s today’s question, across five stories: is A.I. growing up in a controlled way, or is it still running wild? It’s Thursday, October twenty-ninth. Let’s find out.

TIM: That’s Uber’s cap since June: per employee, per AI coding tool.

MARGO: After it ran through its budget. Fortune says the whole year’s, in four months.

TIM: Here’s my take. It’s the same as Meta or Google saying they can’t afford not to spend. Every company felt that.

ZURI: And some overdid it.

TIM: Sure, plenty over-indexed. But the token maxing was the first five percent of employees. The early adopters.

ZURI: And it was a status game. Jyoti Mann, who covers Meta for The Information, the tech news site, in April:

DEV: “Competing on an internal leaderboard called ‘Claudeonomics’ for status as a token legend.”

ZURI: And employees built that leaderboard themselves. Best growth hack Anthropic never paid for.

TIM: Now it’s rolling out to the other ninety-five percent. That’s a big line item, and a limit is sensible.

DEV: And a heavy user isn’t a hundred casual ones. It’s an agent coding “continuously in the background,” Anthropic’s words, when it set weekly limits.

HOLLIS: The FT reported some Amazon staff ran agents on “wholly meaningless” tasks to keep their stats up.

DEV: And Uber’s cap is per tool. Two tools, three thousand a month.

TIM: So is it paying off? Our math, Margo.

MARGO: Wharton surveyed managers at big US companies this summer. Eighty percent say AI is paying, by their own ROI measure.

HOLLIS: And MIT’s study? Ninety-five percent of pilots failing?

MARGO: Different count. About one pilot in twenty makes real money fast. Run twenty, hit one, and you still tell Wharton you’re ahead.

TIM: That’s my thesis. One in twenty, but everybody’s running several. It adds up.

MARGO: And the harder bar: twenty-eight percent say “strongly positive.” Beautiful number. Not a small one.

HOLLIS: Can I test the other ninety-five percent, though?

TIM: Go.

HOLLIS: Gallup, the polling firm. In mid-twenty twenty-four, four percent of US employees used AI daily.

DEV: That’s use, not spend. Nobody publishes how many are the heavy spenders.

HOLLIS: By May, fifteen. And in tech, where Uber and Meta live, forty-two.

DEV: So the budgets showed up when over forty percent of tech used it every day.

TIM: Okay. So the rollout to everyone else is further along than I thought. The budgets landed in the middle of it, not the start.

ZURI: Yes, and the rest is a marketing job. Gallup says people whose manager backs AI are about seventy percent likelier to use it often.

HOLLIS: Anyone here ever been yelled at about an AWS bill? Amazon’s cloud.

DEV: Mine runs local.

DEV: The landlord yells.

HOLLIS: Cloud bills got so big that by twenty twenty there was a foundation for the people who police them. FinOps.

MARGO: And the cloud kept growing. Amazon’s, up nineteen percent in twenty twenty-four, past a hundred billion.

HOLLIS: A budget can turn into a cut. Two thousand one: business tech spending fell two years running, about a seventh of the dollars.

MARGO: Hollis, your model’s called a crash every year since twenty eleven.

HOLLIS: That one happened. Back above its old high in two thousand six. Six years to get back.

MARGO: And this time Wharton has eighty-seven percent of managers still expecting AI spending to rise. Half as many expect big jumps.

TIM: That’s my soft landing. Still climbing, bending gently.

DEV: Ethan Mollick, the Wharton professor who studies AI at work, said it on Big Technology yesterday.

CLIP: Ethan Mollick: “And I think people are worried about token costs, but that’s often a management problem. Like if you don’t make any decisions about how tokens should be used, right? Who should be using them? What process should be using them? Then, and you incentivize people to use as many tokens as possible, they’ll use as many tokens as possible.”

TIM: So a budget is a management decision. Not a retreat.

MARGO: Companies count every token now. Your bank hopes you never do that to your savings rate.

TIM: It’s what I said last episode. These tools let people work around the patterns companies set up against them.

MARGO: Including the savings account?

TIM: If a saver doesn’t have the wherewithal to leave a bank paying next to nothing, an agent should move it. That’s fundamentally good for the consumer.

HOLLIS: The worry, as NPR’s The Indicator put it: “bots pulling client deposits out of accounts en masse in search of higher interest rates.”

ZURI: So the bots are the first customers who actually check the rate.

MARGO: Ten thousand dollars at the FDIC’s national rate, about four-tenths of a percent: thirty-eight bucks a year.

TIM: Next to nothing.

MARGO: The hundred biggest money funds pay about ten times that. Over three hundred dollars more, on the same ten thousand.

DEV: Not insured, though. Online banks pay three to five percent.

MARGO: Upper bound: money-fund rates on every savings dollar would cost banks about a third of a trillion a year. Nearly half their interest income.

ZURI: So the whole bank brand was “you won’t bother leaving.”

HOLLIS: Or a repricing. Cullen Roche runs the asset manager Discipline Funds.

ZURI: And?

HOLLIS: Calling it a bank run “confuses a repricing of bank funding with a panic driven collapse of the system.”

MARGO: Follow the cash: it moves bank to bank, and the squeeze is on margins. Interest is about two-thirds of what banks bring in.

TIM: So a squeeze, not a run.

HOLLIS: For the record, banks used to be capped on what they could pay you. Regulation Q, the Fed’s old rule.

DEV: And people just sat there?

HOLLIS: They walked around it, into money market funds. Four billion dollars at the end of nineteen seventy-seven, nearly sixty times that by late nineteen eighty-two.

TIM: And the banks?

HOLLIS: Congress let them pay market rates. Within about a year, nearly seventy billion came back. Savers won, banks adapted, and the thrifts struggled through it.

MARGO: The banks that paid won.

ZURI: And now?

MARGO: Citi’s paying up. Andy Sieg, introduced on CNBC as its head of wealth, this week: in two days, “over twenty-five thousand clients.”

ZURI: No, because that’s loyalty marketing with a velvet rope. The better rate starts at a thirty-thousand-dollar balance, plus monthly activity.

HOLLIS: That’s his count, in two days. Clients, not deposits.

TIM: So why doesn’t an agent just move the money?

DEV: The door. Most agents log in with your password. Nikesh Arora, who runs the security company Palo Alto Networks, on All-In:

CLIP: Nikesh Arora: “most personal agents are using your credentials to log in. So it becomes hard for a bank to decide whether it’s actually you logging in or is it your MUSE agent who has your credentials logging on your behalf.”

TIM: So the bank has a hard time telling you from your agent.

DEV: So JPMorgan blocks password logins. Agents can read your balance through a connector like Plaid. They can’t move money.

HOLLIS: And JPMorgan’s piloting its own agent, Smart Cash, to move your cash “to maximize yield.”

ZURI: Be the agent, so nobody else’s gets in. That’s the brand play.

DEV: And the open-banking rule, that your bank data is yours on request? As of April, a judge had paused it.

ZURI: Paused for what?

DEV: The CFPB, the consumer finance regulator, was reconsidering who counts as your “representative.” Whether an agent counts as you was the open question.

HOLLIS: And speed cuts both ways. Silicon Valley Bank lost over forty billion dollars in one day, in March twenty twenty-three.

MARGO: Big uninsured balances, from startups. Not ten-thousand-dollar savers.

TIM: So it’s good for savers, and a squeeze for banks, not a run. What I didn’t see is the door: it’s the login.

HOLLIS: And once you do let an agent in, Tim, it’s in everything.

TIM: I don’t give agents that access. And I’m starting to second-guess trusting Google with my email and documents. It’s primarily an advertising company.

MARGO: “Primarily.” I’ll price that in a minute.

TIM: They used to read my email about something I bought, to advertise something similar. They stopped that in twenty seventeen.

ZURI: Babe, your six agents mostly book meetings with each other. So what’s the problem now?

TIM: All my data feels under attack. If not literally, then through the terms of service.

HOLLIS: By who?

TIM: Companies bending the rules to find new ways to get training data for their models.

ZURI: Like who?

TIM: I’m even seeing Instagram ads offering to buy business data. Which is wild.

DEV: Then read this. Google’s privacy policy, as of October sixteenth. No ads from your Gmail, “unless you choose” to share it.

TIM: There’s always an “unless.”

ZURI: But you’d have to choose. It’s opt-in.

DEV: And once you connect it, Google’s own help page says summaries of your mail can show you ads and train Search’s AI models.

MARGO: Summaries. That’s the fine print?

DEV: Its words: “the summary may just be the email.”

HOLLIS: And last November, Gmail’s own account: “We do not use your Gmail content to train our Gemini AI model.”

DEV: Both can be true. Gmail’s line is about Gemini; the help page is about Search.

MARGO: The price I owed you. Five years ago ads were about eighty-two percent of Google’s revenue. Last quarter, somewhere between fifty-eight and sixty-four percent.

TIM: So still mostly ads.

MARGO: Still. Ad dollars grew. But Cloud went from eight percent of revenue to almost a quarter.

TIM: And the data buyers in my Instagram ads?

MARGO: Whoever they are, the market’s real. Google paid ten million dollars for Spirit Airlines’ business data, in its bankruptcy.

HOLLIS: Including a hundred million emails.

MARGO: Ten cents an email, at most. The lot, a sixth of what Reddit reportedly gets from Google a year. Private mail, cheaper than public posts.

ZURI: It’s giving free sample. Meta gives away Muse, its agent, with a hundred million tokens a week.

DEV: NPR asked Meta about ads. It “would not give us a statement on the record.”

HOLLIS: Remember the quiz? You didn’t take it. Your friend did.

TIM: Cambridge Analytica.

HOLLIS: A Facebook personality quiz. Over a quarter million people installed it. Up to close to ninety million people’s data went, by Facebook’s count.

ZURI: For every one who said yes?

HOLLIS: Up to about three hundred who didn’t. The FTC fined Facebook five billion dollars in twenty nineteen.

DEV: Your agent does that to everyone who emails you. And it runs the other way: an email can carry instructions.

HOLLIS: Instructions?

DEV: Simon Willison, the developer who named prompt injection: “an attacker can literally email your LLM and tell it what to do!”

TIM: Can’t you just give it one folder?

DEV: Not through Gmail’s API. Read access means the whole inbox.

MARGO: Then privacy’s a product. Paid Workspace contractually bars training on your data without permission. The free feature doesn’t.

DEV: And Google bans other companies’ apps on its Gmail API from using your mail for ads, or to train AI models beyond your own.

ZURI: So Google’s strictest rules are for everyone else.

TIM: Okay. So my worry isn’t where the terms could go. It’s in the terms as written, once you switch it on. And mine stays off.

ZURI: Different fine print: Claude’s maker just filed to go public.

DEV: Anthropic, the lab that makes Claude, filed Monday to list on Nasdaq. One share, held by its seven co-founders, carries just over half the vote.

TIM: I’m not too concerned. Plenty of companies keep founder control with different share classes. This is basically the same thing.

MARGO: Same size, anyway.

TIM: Show me.

MARGO: The founders own about a tenth of the company, by Fortune’s estimate, for half the vote. Five times their weight.

HOLLIS: And Mark Zuckerberg?

MARGO: About sixty percent of Meta’s vote on thirteen percent. Four and a half.

TIM: So it is the same thing.

MARGO: The size is. At Meta the votes ride on the shares. Anthropic’s half doesn’t move.

ZURI: Meaning?

MARGO: They’ve pledged to give away eighty percent of their stock. Then they’d own about two percent, still half the vote. Twenty-five to one.

DEV: And they don’t pick most of the board. A trust does: five trustees who can’t own the stock.

TIM: So two controllers.

HOLLIS: A guest on Prof G Markets, Scott Galloway’s show, put your side bluntly.

CLIP: a Prof G Markets guest: “I mean, what’s there to talk about? You’re a shareholder in this company. You have no rights. Okay. Let’s move on. If you want to invest in that company, that’s up to you. You know, it’s been disclosed.”

ZURI: Shortest terms of service I’ve heard all day. No “unless.”

HOLLIS: Name the IPO where the public got zero votes.

ZURI: Snap. I had a three-hundred-day streak and no vote either.

HOLLIS: Twenty seventeen. Then S&P Dow Jones, which runs the S&P 500, kept new multi-class companies out for almost six years.

MARGO: And let them back in twenty twenty-three. So index funds aren’t locked out by the structure.

HOLLIS: A vote you don’t have, you can’t use. And on average, across US dual-class IPOs, the early premium is gone in six to nine years.

TIM: What people need to ask is what “public benefit” actually means. Nobody knows what that looks like yet.

DEV: On paper, it’s one sentence: “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.”

ZURI: One sentence. Every brand I’ve declared dead had one.

DEV: Delaware law makes the board balance it against shareholders’ money, and report on it every two years.

ZURI: OpenAI’s for-profit moved to become a public benefit corporation last year too. Same label.

DEV: And Anthropic called its own trust “an experiment.”

MARGO: Which I’d rather hear up front. The filing even says the mission may mean slowing down, at a cost to returns.

TIM: And there’s a real philosophical divide in what Anthropic’s been advocating publicly.

ZURI: And both sides post it. Jack Clark, a co-founder, last year: Anthropic is “endorsing SB 53,” California’s frontier AI transparency bill.

HOLLIS: David Sacks, then the White House AI czar: “a sophisticated regulatory capture strategy based on fear-mongering.”

MARGO: And the Pentagon fight, over Anthropic refusing fully autonomous weapons and mass surveillance, is in the filing’s risk factors. Two courts have split so far.

TIM: So the gap’s the same size today, and it widens as they give.

HOLLIS: So should buyers care?

TIM: Care enough to know what you own: half the vote fixed, a board the trust mostly picks, a mission that can come before returns.

DEV: Google owns a piece of Anthropic, by the way. It reported last night.

TIM: Alphabet first, then. Revenue up a quarter from a year ago. Google Cloud more than doubled.

MARGO: About thirty billion for Cloud in a quarter. Growth went from eighty-two percent to a hundred and three.

HOLLIS: And the stock fell over ten percent after hours. Free cash flow, negative nine billion.

MARGO: Because it spent over fifty billion building in one quarter.

TIM: Here’s my take. They also announced a new model that, at first glance, puts them back in competition with the frontier.

ZURI: Gemini 4 Carbon. Sundar Pichai, Google’s CEO, posted it “ranks #1” on an outside index.

DEV: Artificial Analysis, the independent testing firm. Carbon scores fifty-nine, Claude Opus fifty-eight. But nobody can buy Carbon yet.

TIM: Nobody?

DEV: It’s going to “a subset of cyber defenders” first.

TIM: Okay. Even when they aren’t leading day to day, they’ll stay competitive. They’re a fast follower, and that keeps pressure on token costs.

HOLLIS: Sundar Pichai doesn’t love your word. On Big Technology, Alex Kantrowitz said people thought Google had given up the frontier. Sundar Pichai:

CLIP: Sundar Pichai: “I mean, very wrong. Look, I think we are so committed to being at the frontier. I think it’s in our DNA. We won’t do it any other way.”

TIM: So he says frontier, not follower. Then what about the prices?

MARGO: Argon, the Gemini you can buy: a million tokens in and a million out, twelve dollars.

TIM: Cheap.

MARGO: Through November thirtieth. December first it doubles, to twenty-four. Same as Claude Opus.

DEV: And Argon scores eight points below Opus. OpenAI’s Sol sells at twelve dollars and scores higher. That’s the cheap token: OpenAI’s.

TIM: So Argon’s list price isn’t what’s pushing tokens down.

MARGO: Carbon’s push is what you get, not the token: Opus’s list price, a point better, and about twenty-three percent cheaper per task, Artificial Analysis says.

ZURI: And the intro price is the growth hack. Launch at half, then double. Google’s doing it on Flash too, in January.

DEV: Google designs its own chips. SemiAnalysis, a chip research firm, puts the newest about forty-four percent cheaper to own and operate than Nvidia’s GB200.

HOLLIS: Who got to the personal computer before IBM?

ZURI: Apple?

HOLLIS: With the Apple II, nineteen seventy-seven. IBM’s PC came four years later, became the standard, and the “IBM compatible” clones copied it.

MARGO: Because IBM built it open. Anyone could copy it.

HOLLIS: And IBM sold the business to Lenovo in two thousand five. Google Cloud’s a follower too: still third, at fifteen percent to Amazon’s twenty-eight.

ZURI: But the Gemini app has a billion users. Android and Search put a fast follower in front of everyone on day one.

HOLLIS: A billion, by Sundar Pichai’s count, no period given. Tim’s newsletter counts his agents as subscribers.

TIM: So Google’s competitive again.

MARGO: And the pressure?

TIM: Real, but it’s more model for the money, and an intro price. Not a cheaper token from Google’s list.

MARGO: An intro price is a loss leader. Tim signs up for every one.

TIM: A soft landing needs controls. After today, I think most of them are there: the budgets are caps, not cuts, and the spending is still expected to rise, just less steeply. What I see differently is how far along we are. I thought the rest of the company was only starting to use it. In tech, over forty percent already use AI every day. We’re in the middle of the curve, and it’s bending gently. My inbox stays mine. And for a soft landing, that’s the shape I want to see.