charliebrizz.xlsx SavedCharlie Brizz

Own the Loop

Filed
Reading
9 minutes
Length
2,073 words
By
Charlie Brizz

Anthropic filed its S-1 on October 26. Revenue for the first half of 2026 was $16.3B, 14x the same half last year. That’s the headline. The number I care about is the margin.

Anthropic doesn’t report a gross margin. But the filing does disclose what it cost to serve customers: $2.4B in the half, against $16.3B of revenue. That works out to a gross margin of roughly 85%, the same figure the Motley Fool landed on. NVIDIA, the company everyone calls the big winner of this cycle, runs in the mid-70s. That 85% leaves out training and R&D. Counting every compute dollar, Anthropic spent $8.9B on compute and infrastructure in the half, and its operating loss was still only $693M. But on each token an app rents, the lab keeps most of the dollar.

The company selling the intelligence keeps more of each serving dollar than the company selling the chips. Every AI app, every portfolio company running agents, and every PE firm with an enterprise AI license sits on the other side of that trade. We are all renting intelligence. Tractorbeam included: we build on these same labs’ models.

In January I wrote that the scarce resource in AI is focus. The question I keep getting from PE teams is fair: focus on what?

Own the loop.

A loop is simple. Real work goes in, feedback comes out, and the feedback improves the product in a way the lab can’t copy. It’s the only part of an AI product you actually get to keep. It’s also how you take margin back. Once you can prove what good looks like for a task, you can move that task to a cheaper model, or train your own, without quality dropping. Without a loop, you pay whatever the lab charges.

Three Questions for Every AI Deal

If you take one thing from this post, take these. Every AI vendor and every portfolio company with an “AI strategy” should answer them.

  1. Whose model does it run on, and what happens to margin as usage grows?
  2. What feedback does it collect that the lab can’t see?
  3. What happens the day the lab launches a competing product?

If the answers are “OpenAI,” “none” and “we’re faster,” you’re looking at a reseller. Price it like one. The rest of this post is why.

The Labs Are Keeping the Money, for Now

The cleanest way to see who keeps the dollar is to follow one customer.

In January 2025, a Cursor co-founder wrote on the company’s forum that Cursor was “without a doubt” Anthropic’s largest customer. Business Insider later reported, citing one employee familiar with the numbers, that at one point in its early days Cursor alone was roughly 40% to 50% of Anthropic’s revenue. Neither company confirmed that number. But the hottest AI app in the world was mostly a pipe for someone else’s model.

Then Anthropic launched Claude Code, which passed $2.5B of run-rate revenue about a year after launch. Cursor’s biggest supplier was now its biggest competitor, and Cursor paid it for every token its users burned. An app that resells a lab’s model is in a race with its own supplier, and the supplier controls the price.

In June, SpaceX, which had acquired xAI in February, agreed to buy Cursor for $60B in stock, and the deal closed in August. Two weeks later, OpenAI said it would wind down its contract supplying models to Cursor, with a proposed shutoff of November 12. OpenAI pointed to concerns about SpaceX, not Cursor. Either way, the supplier decides.

Meanwhile, the S-1 says no single customer topped 10% of Anthropic’s revenue in the nine months to September 30. The lab diversified away from its biggest app. The app got sold.

Intelligence Is Getting Cheaper. Your Bill Isn’t.

The obvious pushback: model prices are falling off a cliff, so the labs can’t keep those margins forever.

That’s half right. Stanford’s AI Index found that the price of GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024, more than 280x cheaper. Ramp, which tracks what businesses spend on AI, shows the price companies actually pay per token down 41% from its March peak, and frontier models’ share of tokens down from a 53% peak in August to 45%.

But nobody’s bill is going down. Tunguz puts token usage up 50% since July. People want the newest model, and the newest one is never the cheap one. Agents burn tokens with no human slowing them down. Once usage stops tracking the hours someone spends in an app, no flat subscription price holds up.

So the buyers who win stop asking which model is best and start asking what each task costs. Rogo, which builds AI for banks and funds, found the same task cost about $1.26 on a frontier model and $0.02 on a lighter one. It says 70% of enterprise queries belong on a cheaper model. Hg’s Head of AI, Amr Ellabban, makes the same point: you don’t win by owning the model, you win by owning everything around it. That’s the loop.

The Apps That Win Own the Loop

There are three kinds of loops, and each one turns into margin.

The usage loop. Cursor’s Tab model handles more than 400 million requests a day. Every accept or reject is a training signal, and Cursor can roll out a new checkpoint and collect the next round of feedback in 1.5 to 2 hours. Per Cursor’s own post, the result was 21% fewer suggestions and a 28% higher accept rate. Then it trained its own coding model, Composer, on real tasks inside its own tools.

So what did the loop do for Cursor, if it still got squeezed? It made Cursor worth $60B. What Cursor didn’t have was a model of its own good enough to run that loop on, so the margin kept flowing to its supplier until a buyer with a model showed up. The loop is the asset. Owning it early enough to swap the model underneath is the margin.

The expert loop. Harvey started on OpenAI models, then added Anthropic and Google so it could pick the best model for each legal task. In August it previewed Tenet, its own legal model, post-trained on an open-weight base model across roughly 1,750 legal tasks, each scored against an expert rubric of what the finished work must contain. Harvey says it used no customer data. The asset was experts writing down what good legal work looks like.

Bridgewater ran the same play in finance. With Thinking Machines, it fine-tuned an open-weight model on six financial filtering tasks and hit 84.7% accuracy, against 78.2% for the best frontier model it tested, at roughly 1/14th the cost per task. The detail I love: the outsourced labels were often wrong. So Bridgewater had its own investment experts settle the cases where the model and the labels disagreed.

The outcome loop. Intercom’s Fin answers customer support questions and charges $0.99 only when it actually resolves one. Every conversation tells Intercom whether it worked. On a podcast last October, its team said resolution rates went from about 35% at launch to about 65%, and that model upgrades explained only a couple of points of that. The rest came from testing on real conversations. CEO Eoghan McCabe said they lost money on each resolution at first, at an estimated cost of $1.21, before reaching software-level margins. That’s the whole thesis in one product. The loop made the product better, and the better product paid for itself.

Jaya Gupta and Ashu Garg at Foundation Capital name the raw material: “Salesforce knows what happened. It doesn’t know why.” The exceptions, overrides and reasons a deal got approved live in Slack threads and people’s heads. Whoever captures them at the moment of decision builds something the system of record never sees.

The test I now steal constantly comes from Ellabban: “How do you evaluate whether AI is getting better at the specific thing your customers pay you for?” A specific answer means there’s a loop. A vague one means they’re renting.

One Note on the Circle

Last October I wrote that the house always wins, and Sam Altman owns the house: chipmakers invest in labs, labs commit to clouds, clouds buy chips. A year later the circle is bigger, and the strain shows in the middle, where CoreWeave lost $626M in the second quarter, mostly to interest expense. If you own infrastructure, ask who eats the loss when usage disappoints. If you’re an adopter, an overbuild just means cheaper compute for you.

The Loop Inside Your Firm

The same rule applies to your own firm, and the loop you own is your playbook. AI is very good at work that runs on one: consistent inputs, clear decision points, defined outputs, and feedback on whether the output was right. Marking up an NDA against the positions your firm already wrote down. A first pass on a data room, tested against what you’ve seen in prior deals. Building the purchase price model and running the sensitivities once the structure is set.

What it doesn’t do is sit across from a CEO and decide they’re not the one. Or figure out whether the CIM survives the customer calls. Or decide what to pay and own the call. Harvard and BCG’s “jagged frontier” study of consultants using GPT-4 found exactly this split. Inside the model’s frontier, people got faster and better. On a task just outside it, they did worse, because they trusted a persuasive wrong answer.

George Sivulka at Hebbia explains why most firms haven’t seen the gains yet. When textile mills swapped steam engines for electric motors in the 1890s, output barely moved for about thirty years. The gains came when someone redesigned the factory. In his words, “We’ve swapped the motor; we have not yet redesigned the factory.”

So don’t buy ten tools. Pick one workflow. Write down how your best people actually do it. Build a way to grade the output. Log every decision and how it turned out. Write your own view before you ask the model, so you’re grading it and not the other way around.

Do that for a year and you’ll have something no lab can sell your competitors: a record of how your firm decides, and whether it was right.

The model is rented. The loop is yours.

Sources

Enjoying the insights? Let’s apply them to your PE firm. Say hello