Cloudflare: The Tollgate Is Built, and Two Parties Are Paying

Let me start with a date: the day after tomorrow, September 15, Cloudflare will change a default setting. On pages with ads, "mixed-use" AI crawlers get blocked outright. It applies to new customers, new sites added by existing customers, and all free users—no asking, the default is to block. This was announced on July 1. TechCrunch covered it, so that's secondhand media, but the original policy text checks out on Cloudflare's own website.

A toll booth has been built.

Then I went looking for who's paying, and found two.

Why you can't route around it

W3Techs' September 4 tally: 25.6% of websites use Cloudflare; among sites with a known reverse proxy service, it holds 84.9%.

First, let me correct something I got wrong in my last piece—I wrote "covers 24 million active websites." That number is old, and it isn't the most meaningful measure anyway. The 84.9% of reverse proxy is.

The mechanism of this position has just one link: you point your domain's DNS at it, and traffic flows through it. Whoever the traffic flows through is the one who can do security, performance, and access control at their layer. That's exactly what I saw in that last round of cross-border seller posts—they tried every application-layer blocking app to no avail, and it was finally solved at the edge network layer. But I don't have links to those posts on hand, so I'll treat them as color, not evidence.

The incumbents need to be spelled out, not hidden: Akamai, Fastly, AWS CloudFront plus AWS WAF, Imperva—all at the same layer, all with the same capabilities. Akamai was even in one of the successful cases from the last round.

So I have to revise the word "exclusive."

It isn't exclusive at the product layer. What's exclusive is something else: changing DNS pointing, rebuilding WAF rules, redoing certificates and cache policies—and that set of WAF rules is usually taught to you one real attack at a time, and nobody has that record, including the site owner. Cloudflare has been giving away free DDoS protection and CDN since 2010, sixteen years of it, and by the time a site grows up, the DNS has long since been pointed there.

The cost of switching away isn't on the bill. It's in that record.

Who's paying

Website owners. Not visitors, and not AI companies.

Subscription plus usage, mostly large Enterprise deals. As of June 30, there were 4,698 large customers with annualized spend over $100,000, up 27% year over year, accounting for 73% of total revenue. Net retention 120%. These are the raw numbers from the financials and investor materials.

That's the entire business of the past sixteen years.

The new opening

Something happened in Q2 this year that hadn't happened before: more than half the requests on Cloudflare's network weren't sent by humans. Average daily AI Agent requests rose 1,700% year over year.

So it swapped identities—from "blocking bots for websites" to "collecting money from bots for websites." The products are called AI Crawl Control, pay per crawl, Monetization Gateway. That September 15 default block is the switch for this whole thing.

Up to here it's true: it really can decide who to block and who to let through, and it decides across a quarter of the internet.

Then comes the other half.

Currently two partners are hooked up to pay-per-use, Ceramic.ai and http://You.com。OpenAI isn't, Anthropic isn't, Meta isn't, xAI isn't. Google not only isn't, it publicly pushed back on Cloudflare's crawler classification, on the grounds that it has Google-Extended, so sites can opt out of training without opting out of search.

In the financials, this revenue line is zero rows.

Whether a company can charge, and whether it has collected the money, are two different things. Cloudflare is currently stuck between the two.

While I'm at it, let me get "standard-setting power" straight

The other reason I put it on my list was standards. Last round I saw that Shopify launched Web Bot Auth in May 2025, which is the protocol Cloudflare is pushing—others are actually using the rules it set, and I weighted that heavily at the time.

This time I looked into the status of the protocol itself.

The IETF working group charter was approved in October 2025, and both milestones are past due. As of this August, no draft-ietf-webbotauth-* has been adopted; all that can be found is an individual draft, draft-meunier-webbotauth-httpsig-protocol-01. It is not an RFC.

More important is the verifier side. Those already in production deployment: Cloudflare, AWS WAF, Akamai, Vercel, HUMAN. Vercel even shipped verification ten weeks before the working group was formed. On the signer side, OpenAI is in production, Google signs only some requests, labels it experimental, and the format it uses doesn't even match the one recognized in Cloudflare's docs.

So the accurate statement is: Cloudflare is the proposer and largest deployer of this protocol, not its owner. This layer's position is shared by several players.

Of the two reasons I noted it down for, "exclusive solution layer" needs revising, and "standard-setting power" only holds halfway.

The primary evidence isn't wrong. It's the word I derived from it that was too big.

The numbers (skip if you find this tedious)

Financials as of June 30, released August 6; stock price and valuation as of September 11 close.

Revenue looks good. Q2 was $696.1 million, up 36% year over year; first half $1,335.8 million, up 35%; full-year guidance $2,864 to $2,870 million, up 32%. A company with revenue approaching $3 billion still accelerating—that's not common.

Gross margin doesn't look good. Non-GAAP 73.1%, versus 76.3% a year ago, down 320 basis points. Pulling back further: FY21 was 77.6%, FY24 still 77.3%, FY25 fell to 74.5%, and this quarter's GAAP figure is 71.8%.

In the same quarter, cost of revenue rose 52.7% while revenue rose 36%. Costs outran revenue by 17 percentage points.

These two sets of numbers only get interesting together, same as when I looked at Microsoft last time—a single number means nothing on its own.

Caching, DDoS scrubbing, DNS resolution—these have near-zero marginal cost and are the products that prop up the 79% gross margin. Edge inference, R2's storage and egress, Workers execution—they aren't. The price of changing its business is being paid basis point by basis point in gross margin.

The company's own long-term model floor is 70%. It's at 71.8% now.

On profit: GAAP operating loss of $205.7 million, including $150.7 million in restructuring charges—in May this year it announced layoffs of about 20%, citing AI changing internal efficiency. Non-GAAP operating income $96.1 million, 13.8%. Still not GAAP profitable; the company says it will turn positive by 2028 at the latest.

Cash flow: free cash flow for the quarter $56.4 million, 8.1% of revenue. FY25 full year was $281 million, 13.0%.

There's something here I can't derive. Full-year network capex guidance is 14 to 15% of revenue, while this quarter's actual was only about 8.8%. I originally wanted to infer from this gap that "it's leasing GPUs through finance leases, hiding capex in cost and depreciation"—leases don't go into capex, only into COGS, which would neatly explain why gross margin is falling so fast. Then I realized I can't derive it. I couldn't find the lease obligation details in the 10-Q, and "spending pushed to the second half" is an equally plausible explanation. Without that detail, this is unknown, not a discovery.

Two more numbers to note: stock-based compensation plus payroll taxes $140.6 million, equal to 20.2% of revenue. The $1,293.8 million convertible note due August 15—I haven't read a settlement announcement, so I don't know whether it was settled in cash or converted to equity.

Valuation: September 11 close $306, market cap $109.15 billion, enterprise value $108.52 billion. EV/Sales is 43.2x on TTM, 37.9x on full-year guidance, 29.4x on 2027 consensus. Up 37.5% over the past year, up 59.3% year to date.

Both sides

The three things the bulls can put on the table are all verifiable.

① The position was settled by time, not bought. Revenue contributed by channel partners rose from 25% a year ago to 31%, distribution is expanding outward, not being pushed by itself.

② Three independent measures are all moving up at once. Revenue growth from about 29% in FY25 to 36% this quarter; net retention from 114% to 120%; current remaining performance obligations up 35% year over year. cRPO is a leading indicator—it will turn before revenue growth does—so what to watch here is cRPO, not revenue.

③ Machine traffic gives it a new unit of account. It used to charge by site and bandwidth; now there's another possibility: charging by Agent request.

The three things the bears can put on the table are also all verifiable, and deserve more weight.

① It's switching to a lower-margin business. This isn't a prediction, it's already happening—that string of gross margins above is the proof.

② Being able to charge doesn't mean having collected. September 15 only proves who it can block. Two names on the list, zero rows of revenue disclosure.

③ "Standard-setting power" currently holds halfway. That IETF status above speaks for itself.

The bulls' three are about position; the bears' three are about money. That's no coincidence—its position has never been questioned. What's always been questioned is how much that position can be monetized.

The 10x math

You can't answer this with a feeling. 10x return = revenue multiple × valuation multiple change—that's the decomposition I always use.

On an enterprise value of $108.5 billion, a 10x gain is $1.09 trillion. Working backward, how much revenue is needed:

The 20x tier is the least absurd. An infrastructure company growing 30% with a 30% free cash flow margin getting 20x revenue at maturity would already be a very good outcome.

That corresponds to a 34.3% CAGR.

And what Cloudflare has actually done over the past five years is exactly 34.3%—FY21 revenue $656 million, FY26 guidance $2,867 million.

So "10x in ten years" requires: taking the speed it just achieved, running it for another ten years on a base 4.4x larger, while the valuation multiple drops from 37.9x to 20x. $54.6 billion in revenue equals 23% of its own stated TAM ($238 billion in 2026, based on Gartner's forecast), and it currently holds 1.2%.

Adjusting again for 2% annual net share dilution, the 20x tier's required revenue rises to $66.5 billion, a 37.0% CAGR.

Writing this, I checked on the side how many companies worldwide are at the $1.09 trillion scale right now.

As of September 11, roughly twelve are buyable on US markets with a market cap over $1 trillion: Nvidia $5.3T, Apple $4.8T, Google $4.1T, Microsoft $3.7T, Amazon $2.8T, SpaceX about $2T (it only listed on June 12 this year), Broadcom $1.7T, Meta $1.7T, Tesla $1.4T, and Micron, Berkshire, and Eli Lilly each around $1.1T. TSMC's $2.2T trades via ADR; SK Hynix isn't on US markets.

The latter few sit right on the trillion line, moving in and out of the list on a few points of swing, so "twelve" only holds for September 11.

But the order of magnitude is clear. Those who can stand in this position can be counted on two hands, and they're almost all chips, operating systems, search, e-commerce—things that come along once a generation.

Cloudflare is at $109.1 billion now. 10x means walking from here into that row.

I'm not ruling that out, but I'm not going to think along those lines. On this thread, there's only one thing I can do: come back every quarter and tick three boxes—see whether gross margin has stopped sliding, and whether anyone has actually started paying the toll. Keep expectations low, keep watching.

Let me work out a nearer threshold. To get a 4% free cash flow yield without the market cap changing, I need $4.37 billion in free cash flow; working backward from the company's long-term model of a 30% free cash flow margin, revenue has to reach $14.6 billion—5.1 times this year's figure.

That's the threshold for "buying today doesn't lose money," not the threshold for "ten-bagger."

Compare it with the two software names I bought last time and it's clear: with Salesforce I bought in the eleven-to-fifteen-times range, which by design has an endpoint—what I was earning was the money from other people's doubts about it being repaired. At Cloudflare's current price, the doubts were repaired long ago, even over-repaired—37.9 times forward revenue means the market has already priced in "the toll is being collected," while the toll isn't being collected yet.

So this isn't a valuation-repair trade. It can only be a long-term compounding trade, or not a trade at all.

The three numbers I'll be watching

Same as with Microsoft, I'm writing down the rules before I take a position. Once the earnings come out, the only thing I'll do is tick or cross.

One: gross margin stops falling. Tick: quarterly Non-GAAP gross margin no lower than 72.0%, and no more than 100 basis points below the prior quarter. Cross: falls below 70.0%, or drops more than 100 basis points in each of two consecutive quarters. Currently 73.1%.

Two: the payers become payees. Tick: any of OpenAI, Google, Anthropic, Meta, or xAI appears among the officially disclosed paying parties, or the earnings report breaks out this revenue separately. Cross: after a full year, the list still contains only long-tail vendors, with not a single revenue figure disclosed. Currently two parties, zero disclosure.

Three: growth isn't bought with dilution. Tick: the gap between year-over-year revenue per share growth and year-over-year revenue growth is less than 5 percentage points, and the TTM free cash flow margin is no lower than 10%. Cross: the gap reaches 5 percentage points, or the free cash flow margin falls below 8%. Currently stock-based compensation is 20.2% of revenue, and the quarterly free cash flow margin is 8.1%.

The second one doesn't require waiting for earnings—the payer list could be published at any time. The first and third require waiting for the Q3 report; the company hasn't announced a date yet, but based on past years it should be late October to early November.

I'll come back and tick these three every quarter. If any one of them gets crossed two quarters in a row, a piece of the argument above collapses, and I'll come back and rewrite it.

What I couldn't find

• How much money pay per crawl has actually collected so far. The company hasn't disclosed it.

• The settlement method and dilution of that $1.2938 billion convertible bond. I haven't seen an announcement.

• What exactly that 5-plus percentage point gap in capex is. I covered this above.

• The actual situation after the September 15 effective date. It hadn't arrived yet when I wrote this.

• Whether it's losing share. W3Techs only gives installed-base share, with no comparison of competitors' net growth rates.

• The original links to those seller posts from the last round. I didn't save them myself, so that section can only serve as color—lesson noted: next time I scan for complaints, I'll save the links along with the exact quotes.

Leaving a question

The most awkward part of writing this piece is this: when I first put it on my list, I used the terms "exclusive solution layer" and "standard-setting power." After digging in, the first term needs to be redefined, and the second only half holds. The evidence itself isn't wrong—what's wrong is that the concept I derived from the evidence was too big.

So here's what I want to ask: have you ever had a time when, after abstracting firsthand observations into a concept, that concept grew on its own and ended up making judgments for you? How did you catch it?

My next return is the Q3 earnings report, to tick those three boxes.