Interstellar

THE NUMBER: 80% / 1%. Eighty percent of OpenAI’s and Anthropic’s business revenue comes from one percent of their customers. That’s Ramp’s card and bill-pay data across 70,000 American companies, published yesterday by its lead economist, Ara Kharazian, who called it “a level of concentration risk unseen in any other software category we track.” Then he said the part that matters: the one percent “skew heavily toward the tech sector and AI products and services.” The labs’ customers are the labs. Hold that next to the two IPOs everyone’s waiting for, and next to the $80 billion of compute Anthropic signed for in the last two months. The demand side of the AI trade is the supply side wearing a different badge.

The risk factor

Every S-1 has the paragraph. It sits under Risk Factors, and it reads, in the same words every time: “We derive a substantial portion of our revenue from a limited number of customers, and the loss of one or more of them could materially harm our business.” I read those for a living once. You learn to skim it. Twenty percent from the top ten is normal for enterprise software. Forty percent from three is a yellow flag. You stop reading and pick up the phone when it’s eighty from one.

Ramp’s number isn’t eighty from one customer. It’s eighty from one percent of customers, which is worse, because it tells you the shape of the whole book. Kharazian’s own words, in order: unseen in any other software category; not getting better; the top one percent is tech and AI; “what happens in a market correction? All these companies are highly correlated, and an increasing share of our economy is invested in them. Especially as we approach blockbuster IPOs for OpenAI and Anthropic.”

The spend-per-employee table underneath it is the picture. In July, the median American business on Ramp spent $11.95 per employee per month on AI. The top 10 percent spent $650. The top 1 percent spent $7,400. That’s not a distribution. That’s two different economies using the same word. One of them buys a seat of something for the accountant. The other one buys the model by the megawatt, and it’s the one paying for the frontier.

Ed Zitron compressed it into one sentence yesterday, and I’ll give him the credit because it’s a good sentence: “the AI bubble is based on the whims of maybe a few hundred companies spending money on two companies to justify five companies spending money with one company.” The few hundred are the one percent. The two are the labs. The five are the hyperscalers. The one is Nvidia.

Who the one percent is

Giovanni Cattani published the taxonomy a day before Ramp published the number, in a long X article that pulled 670,000 views, and it’s the most useful thing I read this week because it answers the question Ramp’s data raises. Who are these people, and why do they buy so much?

His frame is a two-by-two. Down one side: bounded versus unbounded tasks. Doing your taxes is bounded; there’s a ceiling on how complicated a tax return can be, and once the model clears it, more intelligence buys you nothing. AI research is unbounded; you can always do more of it. Across the other side: short horizon versus long horizon, borrowed from METR’s chart of how long a task a model can do on its own. o3 could run for about thirty minutes. Mythos runs for about three hours.

Then the sentence that does the work: “Demand for AI is ultimately demand for labor.” Bounded tasks, which is most of what companies do, produce cost-saving ROI and “converge on the cheapest possible option.” There’s only so much demand for accounting. Unbounded, long-horizon tasks are the only ones where a better model pays for itself, because the upside is open-ended. And there are, by his count, three of them: AI R&D, software engineering, and trading. Together, he estimates, they’re roughly half of all frontier inference revenue.

All three are reflexive. A lab spends tokens on research, gets a better model, raises money on the better model, spends more tokens. A startup burns tokens writing software, ships, raises, burns more. A quant shop tests a strategy against the market in real time and reinvests the profits in more tokens. “Larger spend on tokens yields more revenue and a stronger balance sheet, which is spent on more tokens.” Dwarkesh Patel asked on his podcast why the labs aren’t spending even more, given the $100 billion per gigawatt some people expect frontier intelligence to earn; at that rate, Cattani notes, Anthropic could be at $500 billion of run rate by year-end if it monetized every watt it’s contracted.

That’s the loop from the inside. Cattani is honest about the other direction: reflexivity “is great on the way up, awful on the way down.” A pullback in any one of the three could take 15 to 20 percent of frontier revenue directly, and because they’re correlated, up to half. The key demand drivers, he writes, are “correlated and super procyclical.”

So the S-1 paragraph, filled in: we derive a substantial portion of our revenue from a limited number of customers, and those customers are funded by the same trade we are.

Harder to mine

We wrote in Canary in a Coal Mine that intelligence and Bitcoin come out of the same mine, data centers and scarce power and purpose-built silicon, on opposite blueprints. Bitcoin is engineered scarce, so it appreciates. Intelligence is engineered abundant, so it depreciates. Add the piece I left out that day: each block is harder to mine.

Every frontier generation costs more to train than the last. That’s not controversial; it’s the whole scaling argument. What’s new is who it’s being trained for. If the one percent is the customer for the frontier, then each generation is a bigger bet sold into the same ring of buyers, and the ring doesn’t grow as fast as the bill.

Look at what Anthropic (private) actually did on Tuesday. It shipped Fable 5.1 and held list price, $10 per million tokens in and $50 out, still twice Opus 5. The one thing it cut was cache reads, 75 percent, from a dollar to a quarter. That’s the price of re-reading context, which is what a long-running agent does all day, which is what the one percent runs. Artificial Analysis measured Fable 5.1 at max effort and found it costs 20 percent more per task than Fable 5, because it generates 1.7 times the output. Cognition measured its coding agent, where 95 percent of tokens are re-reads, and found the cost per task fell from $5.84 to $2.68. Same model, same day, opposite bills. The Implicator had it right: “the price question now depends on your own traffic.” If your traffic is a loop, you got a discount. If it’s a one-shot question from the accountant, you got a raise.

And the loop is small. Ramp’s August index had Fable 5, Anthropic’s most expensive model, at 6 percent of Anthropic’s tokens and 11.4 percent of its dollars in its first month. Kharazian’s line then was “more performance is not worth the price tag.” AI Secret’s line yesterday was sharper: the best model in the world, “priced 6× the alternative, and the whole launch was spent defending a premium the market already walked away from.”

Google (NASDAQ: GOOGL) said the same thing without saying it. Gemini 3.8 Flash shipped yesterday, the third Flash in six weeks, at the same $0.75 in, $3.75 out as 3.7, 73.7 percent on DeepSWE, “approaching the performance of higher-cost frontier models.” Google hasn’t shipped a Pro since early in the year, and Ars Technica’s read is that we may never see the promised 3.5 Pro. That’s not a delay. Google looked at the same data Ramp has and decided the workhorse is the product and the frontier is a press release.

Meanwhile the token itself keeps falling. Silicon Data’s index hit 97 cents on Monday, less than half the late-May peak. Juniper says US labs’ share of work on OpenRouter went from about 70 percent to about 30 in a year, because Chinese models run up to 90 percent cheaper. A MotherDuck engineer ran Qwen’s 27-billion-parameter model on a laptop against OpenAI’s GPT-5.6 Luna on an agentic SQL benchmark; the laptop won, and a thousand questions cost less than fifty cents of electricity. A professor at Nanjing University told his software engineering students that if they couldn’t afford paid tokens they should drop the course, then recommended fifteen dollars of DeepSeek. Fifteen dollars.

Each block harder to mine, sold to a ring that doesn’t grow, into a spot price that falls. That’s the star.

The star

A star is a business with a cost of goods that rises every quarter. It fuses hydrogen for billions of years. When the hydrogen runs low the core contracts and heats and it starts fusing helium, which lasts millions. Then carbon, which lasts centuries. Then neon, oxygen, silicon, each stage hotter, shorter, and yielding less energy per reaction than the one before, until the core is iron. Iron is where it stops, because fusing iron costs more energy than it releases. The juice isn’t worth the squeeze. The core collapses in about a second, and depending on the mass, what’s left is a neutron star or a black hole, which is where I started with this metaphor before I realized the black hole is the ending, not the story.

I said Interstellar because of what the star throws off on the way out. Every element heavier than helium, everything your desk and your car and your blood are made of, was made in the core of a star that burned itself to iron making it. The iron doesn’t go into the black hole. It goes into the next planet.

Run the labs through it. Hydrogen was GPT-3: cheap, long-lived, changed everything. Helium was GPT-4 and Claude 3. We’re in the carbon stage now, Fable 5.1 and whatever Astra turns out to be, where each generation is hotter and shorter and the yield per dollar of training is visibly falling, and the only customers who still need the extra heat are the ones building the next stage. Iron is when the software’s been rewritten, the trading alpha has been wrung out, and the frontier lab’s only remaining move is to cut the price to reach the other 99 percent. At that point it’s not a frontier. It’s Flash.

And there’s a clock problem, which is the other Interstellar beat. Cooper lands on Miller’s planet, where every hour is seven years back on Earth. He’s gone for what feels like a few hours and comes back to twenty-three years of messages from his kids. The frontier lives on Miller’s planet. It ships a generation every few months, prices it for people who live on the same clock, and reads its own reflexive demand as the world’s. Back home, the world is doing bounded work on a seven-year clock, and it’s fine.

Where most of the world lives

Tomasz Tunguz did something this week that more people should do: he measured himself. Five years of his own posts, scored by an AI panel on a rubric, plus the edit logs from his current workflow, in which an agent drafts and he never touches the raw text, only issues instructions and reviews diffs. The result surprised him. AI did not make him faster. His line-level edits held steady at a median of 136 per post. What changed was the floor. Draft churn went from 47 to 3. The 10th-percentile post improved from 2.59 to 3.81 on his scale, nearly twice the gain of the 90th. His sentence: “Efficiency in knowledge work does not increase our hours. It concentrates our craft.”

That’s the ROI for most of the world, and it’s real. AI deletes your bad drafts. It doesn’t write your best one. That’s bounded work; there’s a ceiling on it, which is exactly why it goes to the cheapest model that clears the bar.

Jason Lemkin’s team at SaaStr built a renewal agent in half a day on top of their AI revenue system. Before agents, they built five custom renewal decks a year, for the Diamond sponsors. The agent built twenty to thirty, including for the Silver sponsors, the smallest checks, the ones who historically got a templated email and renewed at the lowest rate. The Silvers replied at a higher rate than the Diamonds. The labor that made it possible to treat every customer like a king never scaled before. It scales now. And the customers who appreciate it most are the ones who used to get the form letter. (The agent still makes up numbers. Amelia wrote DON’T MAKE UP NUMBERS into its instructions four times, in caps. Check every figure. That’s bounded work, too.)

Yesterday’s letter was about Korn Ferry deleting its junior associates. That’s the same story from the other side. The report the junior wrote was bounded work; the model does it better; the firm’s clients are happier. What got lost was the unbounded part, the apprentice who becomes the partner, and nobody’s paying frontier prices for that either, because you can’t buy it. It was never for sale.

Peter Diamandis put out a piece this week titled “There Is No Job Apocalypse,” and I don’t fully buy it, but his mechanism is right: AI displaces tasks, not occupations, and firms that adopt it heavily grow headcount about 10 percent over the following two years, per the Ramp and Revelio paper Kharazian co-wrote in June. Uber (NYSE: UBER) cut 3,300 people this week, a fifth of its managers, and got sued in Amsterdam by drivers who say a “soulless” pay algorithm sets their fares. Both are true. The bounded world is changing, and it’s changing on the cheap model.

The two exits

If the frontier’s customers are the frontier, the labs have two ways out. Find an industry that needs the frontier and doesn’t yet know it. Or become one.

Watch the second, because it started this week. OpenAI (private) wired ChatGPT into Epic, which holds the records of most American hospital patients; clinicians can ask what changed since the last visit and the model answers from the chart. Shelly Palmer’s read: “OpenAI is working through the economy one industry at a time. ChatGPT Finance moved in on the tax preparers, consumer ChatGPT Health is going right after health apps, and this release is placing an AI layer over the clinicians’ core information workflow.” Marc Andreessen said software was eating the world. Palmer: “Now, AI is eating the software.” ChatGPT’s ad business crossed a $1 billion run rate in under 200 days. The Pentagon gave three million people ChatGPT Mil. Anthropic shipped blueprints for retail agents.

And OpenAI told Cursor, its largest coding customer, that OpenAI models stop working inside Cursor on November 12, because SpaceX bought Cursor. That’s the sentence to sit with. The lab’s biggest software customer just found out what “customer” means to a lab. Nate’s newsletter got the lesson right: “the model is replaceable; six months of accumulated working context is not. Custody is a more consequential form of lock-in than a monthly subscription.” Keep your files, your decisions, and your instructions outside any provider. Concentration with an exit.

The first exit, a new industry, is the one everyone’s waiting on and nobody can name. Drug discovery, maybe. Materials. Something unbounded and long-horizon that isn’t already inside the loop. It might come. But it’s a hope on the S-1, not a customer.

Lucent

The last time I saw a supplier’s customers this dependent on the supplier, it was 1999, and the supplier was Lucent.

Here’s how it worked. The telecom act of 1996 created hundreds of competitive local carriers, CLECs, with no revenue and enormous build plans. Lucent, Nortel and Cisco sold them the switches and routers, and when the CLECs couldn’t pay, the vendors lent them the money. Vendor financing. By 2000 Lucent had roughly $8 billion of it committed on its books. The CLECs’ purchases showed up as the vendors’ revenue, the vendors’ revenue justified their stock prices, and the stock prices funded more financing. When Winstar filed in April 2001 it owed Lucent more than $700 million. Lucent never recovered. The equipment did; most of the fiber those carriers laid was eventually lit, twenty years later, by someone else. The build was real. The revenue was the vendor’s own capital coming back around.

Now read the Anthropic-Lambda deal with that in your head. Anthropic signed for $35 billion of compute over six years with Lambda, which is backed by Nvidia (NASDAQ: NVDA). The compute sits at a Hut 8 (NASDAQ: HUT) campus in Texas where Nvidia holds fifteen-year leases. Nvidia is the chip supplier, the lessee, and an investor in both parties. Its supply and capacity commitments just hit $279 billion, up 506 percent in a year. This week it bought $3.5 billion of MediaTek (TWSE: 2454) convertibles to get NVLink inside the chips customers are designing to replace Nvidia’s own. SB Energy filed to go public on Tuesday with zero operational data-center revenue, a $3.2 billion first-half loss, OpenAI named 306 times in the prospectus, Nvidia in for $3 billion, and OpenAI holding warrants worth about $5.5 billion for agreeing to be the tenant. PitchBook estimates Anthropic’s gross margin at 44 percent and says it needs about 70 to carry its commitments.

That’s vendor financing, one floor up. And the Ramp data is what makes it Lucent instead of just circular: the customers are funded by the trade. The eighty percent is the CLECs.

Paying the light bill

One more thing about stars: they’re hot, and you feel it before you get any use out of them.

George Noble passed along a post yesterday from Andy O that says it cleanly. Treasury Secretary Bessent says the AI capex boom will be “extremely disinflationary” and we’ll see the benefits within six months. But the productivity is still theoretical and the inflation is already in the data. Electricity is up 4.2 percent year over year. Information technology goods are up 3.6 percent. Computers and peripherals are up 5.1. The Dallas Fed estimates data-center power demand alone adds five basis points to PCE inflation this year; Allianz puts it at 19 basis points of CPI in the exposed regions. Hundreds of billions of dollars are competing for the same power, gas, transformers, copper, construction labor, HBM and DRAM at the same time, increasingly on borrowed money. “That is a massive demand shock before it is a productivity shock.”

Tim Cook told the Wall Street Journal that Apple (NASDAQ: AAPL) has been absorbing “huge increases” in memory prices and can’t anymore, so Mac and iPad prices go up and iPhones will follow. The Yahoo Finance writer who connected that to the Lambda deal had the right image: a stadium tour buying out every hotel room in the city, and everyone else quietly paying surge pricing. “The frontier model you’re chatting with has a tab. You’re quietly picking it up.”

SightBringer wrote a thread this week about the pandemic checks that applies here with the names changed. The public got temporary cash; asset owners got a permanent repricing of the world around them. “The check arrived once. The dilution kept compounding.” Swap in the AI build: the one percent gets the tokens and the repricing; the rest of us get the light bill. Bessent’s six months are on Miller’s planet, too.

Where I stand

Three moves, and then the allocation view, because you’ll ask.

Don’t sign a multi-year frontier deal this month. The price is coming down to meet you, not because the labs are generous but because the one percent can’t carry the next generation alone. Anthropic cut cache reads 75 percent on launch day. Google shipped three Flashes and no Pro. The token index is 97 cents. If there’s a Fable or GPT renewal on your desk, push it to month-to-month and let them chase. You’re the customer they’ll be courting in 2027.

Point the agent at the work you do badly, not the work you do well. That’s where Tunguz’s floor rose and where SaaStr’s Silvers replied. Today: list the twenty accounts, reports or customers that get the form letter. That’s the queue. Cheap model, real ROI, and nobody’s frontier premium required.

Run the risk factor on your own book. If a large share of your revenue comes from customers who are themselves funded by the AI trade, you own the loop whether you meant to or not. Sort your top ten by who pays them.

On allocation: I don’t want the star. I want what benefits from bounded work and the general growth of AI (the picks and the ash: the memory, the power, the harness, the open weights, the boring software that got a floor raised under it), and I want businesses that don’t get 80 percent of their revenue from one percent of their customers. The frontier lab is the one asset in this trade whose cost of goods rises every generation and whose customer base is itself. Everything else in the stack gets cheaper inputs as the star burns. A star is a magnificent thing to watch and a terrible thing to own.

Reflexivity is great on the way up, awful on the way down.”
— Giovanni Cattani, September 1, 2026

The frontier runs on Miller’s planet. You don’t. Don’t buy the star. Build with the iron.

---

The email edition (THE NUMBER, the three moves, and the day’s five stories) is in your inbox. If the question on your desk is which workloads are bounded and what’s the cheapest model that clears the bar, that’s the conversation we run at Outsider Labs.

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.