The Market Bought Traffic. Cloudflare Bills by the Month.
On September 9 the market paid five edge names $17.65 billion for the agent economy. Cloudflare's filings can say where a request lands today — on the cost side — and five tripwires say what would show the bet was right.
Prices are regular-session closes through 2026-09-10; filings are the FY2025 10-K and the Q2-2026 10-Q and 8-K; pricing pages and Cloudflare Radar figures as of 2026-09-11. Cloudflare reports Q3 on October 29 (date not yet company-confirmed).
The verdict, first
On Wednesday, September 9, Cloudflare rose 10.5%, Datadog 7.2%, Fastly 5.7%, Akamai 4.9%, Zscaler 2.6%. Together the five added $17.65 billion. The security companies next to them on the screen did not move: Palo Alto Networks fell 0.6%, CrowdStrike fell 1.1%, SentinelOne was flat. The Nasdaq 100 was down. The market bid the companies that carry internet traffic, and left the companies that inspect it alone.
It had a story to act on. In August, Cloudflare's chief executive told shareholders the company is "building the infrastructure, controls, developer tools, and payment rails for the Agentic Internet." The day before the move, Meta posted a personal agent that books, buys and negotiates on its own. Put the two together and the bet has a precise shape: as AI agents start doing tasks on people's behalf, the traffic they generate becomes revenue for the companies it passes through — the edge, the layer of networks that sits between websites and whoever, or whatever, is calling them. The bet is testable, because Cloudflare's filings describe exactly how the edge's largest company gets paid.
What the filings say: Cloudflare's revenue is "subscription fees to access its network and products, support services, and usage-based fees," and the usage-based part is "primarily related to fees charged for the Company's customer's use of excess bandwidth." Its pay-as-you-go tiers are $25 and $250 a month. Its Enterprise contracts run one to three years. The only product family that bills by the request is the developer platform: $5 a month minimum, ten million requests included, $0.30 for each additional million. What share of revenue is metered at all, the company does not disclose.
What the numbers already show: a request lands on the cost side of Cloudflare's accounts before it lands on the revenue side, if it lands there at all. Cost of revenue has grown faster than revenue for seven straight quarters. Gross margin fell in each of seven consecutive quarters, from 77.8% in mid-2024 to 71.2% in the first quarter of this year, and recovered 0.6 points in the second. The company's annual report expects both halves: that cost of revenue "will continue to increase due to higher network and bandwidth costs" as customers use the network more, and that it will "continue to benefit from economies of scale" as they do. The series shows the first. It does not yet show the second.
The overall point, stated once. Agent traffic may become a large business for the edge. Today, the edge's largest company sells subscriptions and pays for traffic. The products that would turn a request into revenue on their own terms — paid crawling, per-call payment rails, agent identity — exist, and are in closed beta, on a waitlist, or unpriced. The market has priced them as if they were shipping. The sections that follow take the bet apart: what was bought, how the edge gets paid, where a request lands in the accounts today, where it would show up if the bet is right, who else was bought that day, and what would settle it.
The session
Two candidate explanations circulated for Cloudflare's 10.5%. The first was a security partnership with OpenAI: a vulnerability-remediation service inside Cloudflare's Managed Defense, using OpenAI's Daybreak models. Cloudflare's press release and blog post for that service are dated September 3, six days earlier. On the day it launched Cloudflare rose 4.3%, against a 1.2% gain in the Nasdaq 100; the next day it fell 2.0%. The second explanation was Meta's Muse, a personal AI agent that "can open a browser, fill out forms, and negotiate on their behalf," posted on Meta's newsroom on September 8; Meta rose 6.6% on the 9th. Cloudflare's own newsroom carries no announcement for September 9; the day's only post is an engineering note on Workers.
We do not resolve which headline did it. The shape of the day is the evidence: the market bought companies that carry and observe traffic, and did not buy companies that sell security. Whatever the trigger, the thing being priced was traffic.
| Name | 9/9 move | Value added | Own announcement that day |
|---|---|---|---|
| Cloudflare | +10.5% | $10.6bn | none |
| Datadog | +7.2% | $5.4bn | none |
| Fastly | +5.7% | $0.2bn | none |
| Akamai | +4.9% | $0.7bn | none |
| Zscaler | +2.6% | $0.7bn | Agentic SOC launch |
| Palo Alto Networks | −0.6% | ||
| CrowdStrike | −1.1% | ||
| Nasdaq 100 | −0.3% |
Market value added on September 9 by Cloudflare, Datadog, Fastly, Akamai and Zscaler. Four of the five announced nothing that day. The security names on the same screen did not move. Whatever the trigger, the thing being priced was traffic.
What an agent does to a request
A request is one message from a program to a server: fetch this page, run this function, return this answer. A person browsing generates a few per action. An agent — software that completes a task rather than answering a question — generates many: it reads a page, follows links, calls an inference model, checks a price, tries again. Every one of those passes through some piece of edge infrastructure, and Cloudflare's network fronts a large share of the public web. Cloudflare's own Radar data already reports that more than half of internet traffic is non-human, that 52% of crawler requests in June 2026 were for AI training (up from 22% in spring 2025), and that requests where an AI bot acts on a live user's behalf were under 5% of AI-bot traffic in mid-2025. The agent share is small and rising. The question is not whether the requests come. It is what a request is worth to the company it passes through.
How the edge gets paid
Cloudflare's price list, set against how each part bills:
| Product family | How it bills | Price | Source |
|---|---|---|---|
| Pro / Business plans | Flat subscription, monthly or annual | $25 / $250 per month | Cloudflare blog; 10-K |
| Enterprise | Contract, 1–3 years; some usage terms for large customers | negotiated | 10-K |
| Excess bandwidth | Usage fee, per period | the main usage fee the filing names | 10-K revenue note |
| Workers (developer platform) | Metered by requests and CPU time | $5/mo min; 10M requests included; $0.30 per additional million | Workers pricing page; 10-K |
| Workers AI | Metered by "neurons" | $0.011 per 1,000 | Workers AI pricing page |
| AI Gateway | Core free; add-ons paid | free (analytics, caching, rate limits); 5% fee on inference credits bought through Unified Billing | AI Gateway pricing page |
| Pay per crawl | Publisher sets a per-request price; Cloudflare is merchant of record | Cloudflare's share not stated; closed beta | Cloudflare docs |
| Monetization Gateway | Per-call payment rules, settled in stablecoins | fee not stated; waitlist | Cloudflare blog, 2026-07-01 |
| Web Bot Auth | Cryptographic identity for bots and agents | plan not stated | Cloudflare docs |
Three things follow. First, the two products that would price a request on its own terms, pay per crawl and the Monetization Gateway, are in closed beta and on a waitlist with no stated Cloudflare take, and the identity primitive beneath them, Web Bot Auth, has no stated pricing at all; the one stated take that ships is the 5% fee on inference credits bought through AI Gateway. Second, the metered products that are shipping price requests at thirty cents a million and inference at $0.011 per thousand neurons. Third, the filing tells us the metered share of revenue is undisclosed, which means no outsider can compute what a marginal request earns. The toll booth the company describes has no posted toll.
The case that agent traffic is a revenue pool for the edge, in its strongest form. Four arguments, all of them serious.
The new products are the point. Agents that buy, book and negotiate need to prove who they are and pay for what they take. Identity and payment at the edge are not more of the old traffic at the old rate; they are new product lines, and Cloudflare has shipped the primitives — Web Bot Auth, pay per crawl, the Monetization Gateway, stablecoin settlement over the x402 protocol. Beta today is a launch schedule, not an absence.
Cloudflare bills both ends of the request. The website an agent visits is a Cloudflare customer. The developer who built the agent may run it on Workers at $0.30 per million and route its inference through AI Gateway. Two tolls, one request.
Subscription is the land; usage shows up at renewal. A customer whose traffic doubles does not pay more this month. It upgrades from Pro to Business, or from Business to Enterprise, or expands its Enterprise commitment when the contract comes up. That is what dollar-based net retention measures, and Cloudflare's is 120%, up from 114% a year ago. Customers paying over $100,000 a year rose from 3,712 to 4,698. Contracted revenue not yet recognized rose from $2.50 billion to $2.73 billion in six months. If agent traffic is arriving, this is where it would surface first, and it is surfacing.
The market was selective. It bid the names that carry traffic and left the names that sell seats. A blanket AI bid would have lifted CrowdStrike. It did not.
The test below is written against these four claims. The first and third are the ones we cannot settle from filings, and we say so.
The ledger
Now the numbers the company has filed, quarter by quarter. Revenue and cost of revenue from Cloudflare's XBRL filings; fourth quarters derived from full-year less nine-month figures.
| Quarter | Revenue | growth y/y | Cost of revenue | growth y/y | Gross margin |
|---|---|---|---|---|---|
| Q2 2024 | $401.0m | $89.0m | 77.8% | ||
| Q3 2024 | 430.1 | 96.0 | 77.7% | ||
| Q4 2024 | 459.9 | +26.9% | 108.7 | +30.5% | 76.4% |
| Q1 2025 | 479.1 | +26.5% | 115.6 | +35.9% | 75.9% |
| Q2 2025 | 512.3 | +27.8% | 128.7 | +44.6% | 74.9% |
| Q3 2025 | 562.0 | +30.7% | 146.3 | +52.5% | 74.0% |
| Q4 2025 | 614.5 | +33.6% | 162.0 | +49.0% | 73.6% |
| Q1 2026 | 639.8 | +33.5% | 184.2 | +59.3% | 71.2% |
| Q2 2026 | 696.1 | +35.9% | 196.5 | +52.7% | 71.8% |
Revenue growth has risen from 26.5% to 35.9% over six quarters, with one 0.1-point pause. That is the fact the price is built on, and it is real. Cost of revenue has grown faster in every one of the seven quarters shown, by 4 to 26 percentage points. Six points of gross margin left in two years.
The company names what the cost is. For the second quarter, the $67.9 million increase in cost of revenue was "primarily due to" three things: $30.1 million more in "third-party technology services," $20.9 million more in co-location, network and bandwidth costs "for operating our global network for our expanded customer base," and $10.1 million more in depreciation from "server acquisitions and deployments." For the full year 2025 the same three lines were $59.2 million, $54.1 million and $45.2 million.
Two of those three are traffic. Bandwidth and co-location are what carrying a request costs. Depreciation is the servers that carry it, and the 10-K says the deployment now includes "servers with graphics processing units (GPUs) to support our AI-related developer platform products." The largest line, third-party technology services, is not defined in the filing; the same phrase appears in the sales and administrative expense discussions, which marks it as a general purchased-services category. It may be inference capacity bought from others, software, or both. We cannot attribute it. The cost lines, like the expansion metrics in section 5, are consistent with customer traffic and consistent with Cloudflare building capacity for its own AI products; the filing separates neither.
Set the two halves together. The filing describes a business whose revenue is mostly subscription and whose named cost increases are carriage, server depreciation, and an undefined purchased-services line, and it reports seven quarters in which those costs grew faster than revenue. That is what agent traffic would look like on a subscription business: it arrives as cost first. The decline is not an accounting artifact: on the company's own non-GAAP basis, which strips out stock-based compensation and related payroll taxes and acquired-intangible amortization, gross margin went 79.0% → 76.3% → 73.1% across the same three second quarters, a 5.9-point fall against GAAP's 6.0. The 10-K expects costs to keep rising with usage and expects economies of scale to offset them. Seven quarters of the series show the first and not yet the second.
Where usage would surface, and when
The strongest argument on the other side is the lag argument: a subscription business monetizes usage at renewal. The evidence for it is the expansion metrics — net retention 120%, large customers up 27%, contracted backlog up 9.5% in six months — and they are moving the right way. What they cannot do is attribute. Net retention rose from 114% to 120% while the company was also adding products, raising list prices, and moving customers from monthly plans to Enterprise contracts; the filing does not separate agent-driven expansion from any of that, and offers no metric that would. So the honest statement is: the expansion metrics are consistent with the traffic thesis and consistent with everything else, and the company has disclosed nothing that distinguishes them.
| Expansion metric | Earlier | Later | Change |
|---|---|---|---|
| Dollar-based net retention | 114% | 120% | +6 pts |
| Customers paying >$100,000 a year | 3,712 | 4,698 | +27% |
| Remaining performance obligations | $2.50bn (Dec 2025) | $2.73bn (Jun 2026) | +9.5% in six months |
The tripwires below are built on that gap. The first thing that would change our view is a number, from Cloudflare, denominated in dollars, attributable to agent products.
The read-through
The 9/9 bid also lifted four other companies. The same question — does a request become revenue? — has a different answer at each, from their own filings.
Fastly is the clean case: "we derive our revenue primarily from usage-based fees," charged "based on their committed or actual use of our platform, as measured in gigabytes and requests," and "the majority of our revenue is usage based." A request is revenue at Fastly, by construction. It added $0.2 billion on 9/9, the smallest of the five.
Akamai says its revenue is driven by "the amount of traffic we serve on our network," and that bandwidth is "a significant portion of our cost of revenue" — the same two-sided exposure as Cloudflare, with more of the revenue side usage-linked. It is renegotiating multi-year contracts "to better align unit pricing" with its own hyperscaler cost increases.
Datadog bills committed usage with overage: "to the extent that our customers' usage exceeds the committed contracted amounts under their subscriptions... they are charged for their incremental usage." Agents generate logs and traces; Datadog meters both. It added $5.4 billion, the second-largest, and its exposure is the most direct after Fastly's.
Zscaler recognizes subscription revenue ratably over one- to three-year terms, and its own 10-Q lists "AI agents" as a driver of cost — "our cost of revenue will increase due to higher bandwidth and data center expenses" as customers add "applications, workloads, AI agents, and connected devices." Zscaler's 9/9 move also had its own news, a product launch that morning, so we exclude it from the no-announcement claim.
The read-through, then, was not indiscriminate: it went furthest for the name with the most subscription-shaped revenue and the largest market value, not the name with the most usage-shaped revenue. If the market had been pricing per-request exposure, Fastly, not Cloudflare, would have led the day. That is the per-request reading's prediction. The new-product reading — identity and payment rails, which Fastly does not sell — predicts what was seen, and the session cannot separate the two.
Verdict and tripwires
The question was whether agent traffic is a profit pool the edge can meter, or a volume it carries at a flat price. From the filings: today, for Cloudflare, it is the second. Revenue is subscription and bandwidth overage; the metered products price a million requests at thirty cents; the products that would charge an agent on its own terms are in beta with no stated take; and the two-year series shows traffic arriving as cost, with gross margin down six points while revenue growth accelerated. The market's $10.6 billion was paid for a product set the company still describes as being built. The market bought traffic. For now, Cloudflare bills by the month.
That is a description of the present, not a forecast. The company may convert. The tripwires are written to catch it if it does, on outcomes the company files rather than on traffic it reports.
Tripwires (armed):
- FOR THE POOLAT-1 the toll gets a number. Cloudflare discloses revenue, or a revenue-denominated metric, attributable to agent or AI products (Workers AI, AI Gateway, pay per crawl, Monetization Gateway, agent identity) at any print through Q4 FY2026.
- FOR THE POOLAT-2 traffic arrives as revenue. Gross margin at or above 74% with revenue growth at or above 30% year over year, for two consecutive quarters.
- AGAINSTAT-3 traffic keeps arriving as cost. Gross margin at or below 72% for two further consecutive quarters with no agent-attributable revenue metric disclosed.
- AGAINSTAT-4 the print passes in silence. Q3 (October 29, tentative) and Q4 pass with no AT-1 disclosure and revenue growth below 30% year over year.
- FOR THE POOLAT-5 a metered product ships with a stated take. Pay per crawl or the Monetization Gateway reaches general availability with a disclosed Cloudflare fee or revenue share.
Two prints cannot settle the renewal argument in section 5; AT-3 and AT-4 score what the 9/9 session priced, not the thesis. A metered-inference mix could grow revenue below 74% gross margin, so AT-2 tests carriage leverage while AT-1 and AT-5 carry the pool case.
Not a tripwire: Cloudflare Radar's AI-bot share of requests. Traffic rising while revenue stays flat is exactly this note's failure mode for the edge; a traffic figure cannot distinguish the two outcomes.
What would change our view. We abandon "agent traffic is a cost before it is a revenue" if AT-1 or AT-5 fires, or if AT-2 fires without a change in revenue mix that the filing attributes to something else. Any of those would show the company converting a request into a priced unit, and the ledger in section 4 would then be read as the build-out ahead of the revenue, which is how the company describes it. In the other direction, we would read AT-3 as build-out rather than carriage if the filing attributes the cost growth to GPU capacity for Cloudflare's own AI products. We would also revisit the note if Cloudflare re-segments its revenue to disclose usage-based revenue separately; the metered share is the number this note most wants and cannot compute.
Sources & method
Prices. Regular-session daily closes, Robinhood, pulled 2026-09-11, for NET, DDOG, AKAM, ZS, FSLY, PANW, CRWD, S, CIBR, QQQ, META, 8/31–9/10. Market value added = shares outstanding (Robinhood fundamentals, 2026-09-11) × close-to-close price change; share counts held constant. Cloudflare Q3 report date and verification status: Robinhood earnings data.
Announcement dates. Cloudflare press-release page and blog (checked 2026-09-11): OpenAI Daybreak service dated 2026-09-03; no release dated 2026-09-09. Meta newsroom: "Introducing Muse," 2026-09-08. Datadog, Akamai, Fastly, Zscaler: wire coverage (MT Newswires, Benzinga) via Robinhood; the Zscaler Agentic SOC launch is MT Newswires, 2026-09-09 03:51 ET.
Filings. Cloudflare 10-K FY2025 (0001477333-26-000016), 10-Q Q2-2026 (0001477333-26-000054), 8-K 2026-08-06 (0001477333-26-000053), 8-K 2025-07-31 (0001477333-25-000136) for the year-ago non-GAAP gross margin. The 10-K cost-of-revenue paragraph is quoted in full in the data archive. Quarterly revenue and cost of revenue from XBRL company facts via EDGAR; Q4 figures derived as full-year less nine-month. Datadog 10-Q (0001628280-26-054458), Akamai 10-Q (0001086222-26-000086), Fastly 10-Q (0001517413-26-000213), Zscaler 10-Q (0001713683-26-000096).
Pricing. cloudflare.com plan and developer-platform pricing pages; Cloudflare blog "Adjusting pricing, introducing annual plans"; developers.cloudflare.com pricing pages for Workers, Workers AI, AI Gateway, pay per crawl, Web Bot Auth; Cloudflare blog "Announcing the Monetization Gateway" (2026-07-01). All as of 2026-09-11.
Radar. Cloudflare blog "Content Independence Day, one year on" (2026-07-01) and "A deeper look at AI crawlers" (2025-08-28). Used for context; not a tripwire.
What this note does not do. It does not value Cloudflare. It does not attribute the 9/9 move to a cause. It does not forecast agent adoption. It does not use earnings-call transcripts; the "agentic Internet" quotation is from the 8-K exhibit.
Disclosures
Information only. TON618 Capital. This report is for information purposes only. Nothing here is an offer to sell or a solicitation of an offer to buy any security, fund interest, or digital asset, and nothing here is personalized investment advice or a recommendation regarding any instrument.
Publisher's exclusion. All research is published solely as general, impersonal information of regular circulation. It is not tailored to the objectives or circumstances of any individual and is not issued in connection with compensation from any client. The Fund has no clients and distributes all research free of charge. On that basis it publishes in reliance on the publisher's exclusion from the definition of "investment adviser" under the Investment Advisers Act of 1940 (§202(a)(11)(D); cf. Lowe v. SEC, 472 U.S. 181 (1985)).
Registration & conflicts. TON618 Capital is not registered as an investment adviser or broker-dealer in any capacity. The Fund is a Bitcoin fund and may hold or transact in the securities or digital assets it discusses; material conflicts are disclosed where they exist. Ownership: the Fund holds no position in any company named in this report as of the report date — Cloudflare, Datadog, Akamai, Fastly, Zscaler, Palo Alto Networks, CrowdStrike, SentinelOne and Meta Platforms among the public issuers, and OpenAI and Anthropic, which are private and named only as model providers; none is a Fund holding and none is a BTC-correlated instrument. Compensation: the Fund received no compensation from any party in connection with this report and charges nothing for it.
Use of AI. Artificial intelligence is used in the creation of this research. All methodology and data integrity are reviewed and approved before publication by TON618 Capital's Chief Investment Officer, Keyth Beck; errors may nonetheless occur, and readers should verify independently.
CFA. This report was prepared to align with CFA Institute analytical standards (methodology only). CFA® and Chartered Financial Analyst® are registered trademarks owned by CFA Institute. That reference describes the analytical framework applied; it does not imply the report was prepared, reviewed, or authored by a CFA charterholder, and the report is not issued, reviewed, endorsed, certified, or approved by — nor affiliated with — CFA Institute.
Risk & feedback. Past performance is not indicative of future results. Digital assets and equities are volatile and may result in total loss of capital. Corrections and feedback are welcome — please direct them to CIO Keyth Beck at keyth@ton618capital.com. Version 1.0.