AI Business Ideas

Cloudflare's Agent Internet: Three Businesses to Build Now

Direct Answer

Cloudflare is assembling credible infrastructure for an internet where software can discover, price, pay for, and retrieve digital resources inside an HTTP request. That is a meaningful platform shift. It is not yet a turnkey economy, and the title's promise of 1,000 new millionaires is a founder thesis rather than a forecast.

The practical opportunity is simpler. Agents need clean data, current business facts, reliable tools, and licensed expertise. A founder can sell those resources manually today, learn what buyers value, then expose the repeatable part through a search API, MCP server, or paid endpoint as agent purchasing becomes easier.

The useful sequence: sell the decision or workflow first, prove repeated demand second, structure the resource third, and automate payment last. A payment rail removes checkout friction. It does not create a reason to buy.

Watch Greg Isenberg's Episode

Episode and idea credit: Greg Isenberg. Watch the original video on YouTube or follow Greg on X. Greg states that he has no affiliation with Cloudflare; this article is also independent and is not endorsed by Cloudflare.

What Exists Today, and What Is Still Coming

The four Cloudflare products in the episode are related, but they are not at the same release stage. Treating all four as fully available would turn a good thesis into bad implementation advice.

LayerAugust 2026 statusWhat it doesWhat to do now
AI Crawl ControlGenerally availableShows AI crawler activity and lets site owners allow, block, or return a customizable 402 licensing message. Custom 402 messaging is for paid plans.Audit crawler traffic, define an access policy, and publish a licensing contact path.
Pay Per CrawlBetaLets participating publishers allow, charge, or block authenticated crawlers at a domain-wide price. Cloudflare records billing events and distributes earnings.Join only if crawler access itself is worth selling; do not forecast revenue before demand exists.
AI IndexPrivate betaPlans a site-owned index with an MCP server, search and bulk APIs, pub-sub updates, and generated llms.txt surfaces.Build clean source pages and structured data now; treat the managed index as an emerging option.
Monetization GatewayEarly-access waitlistPlans payment rules for pages, files, APIs, data, tokens, and MCP calls, with x402 verification at the edge.Validate the resource with normal billing, then join the waitlist if per-use purchasing fits.

Developers do not have to wait to test the pattern. Cloudflare already documents how to gate HTTP content and MCP tools with x402 in Workers. That is a build-it-yourself path, not the same as having the Monetization Gateway generally available.

How x402 Turns a Request Into a Transaction

HTTP 402 has existed as a reserved "Payment Required" status for years. The x402 protocol gives machines a concrete payment exchange around it:

  1. Request. A client asks for GET /resource.
  2. Price. The server returns 402 Payment Required with a PAYMENT-REQUIRED header containing the price, token, network, and merchant address.
  3. Authorize. The client creates a signed payment payload and retries with a PAYMENT-SIGNATURE header.
  4. Verify and settle. The server verifies directly or uses a facilitator, which checks and broadcasts the pre-signed transaction without holding funds.
  5. Deliver. The server returns the resource with a PAYMENT-RESPONSE settlement confirmation.

Cloudflare's documentation lists fixed-price exact payments across several networks and an EVM upto scheme for variable settlement. Stablecoins are a natural early rail, but x402 does not guarantee universal wallet support, tax simplicity, buyer budgets, or agent adoption.

Pay Per Crawl and x402 are not identical. Pay Per Crawl uses Cloudflare crawler authentication, pricing headers, aggregated billing, and Cloudflare as merchant of record. x402 is a broader open protocol for programmatic payment to many kinds of resources.

The Agent Internet Stack

The episode's strongest idea is not the payment technology. It is the stack underneath a resource an agent can trust enough to buy.

LayerOwner's jobFailure if missing
Rights and provenanceKnow who owns the source and preserve where each fact came from.The product cannot be licensed safely.
CollectionGather the pages, records, calls, files, or expert material.The resource is incomplete or biased.
NormalizationCreate consistent entities, fields, dates, and units.Agents receive plausible but incomparable data.
FreshnessDetect changes, timestamp records, and expire stale claims.A cheap answer causes an expensive wrong decision.
AccessExpose a report, API, feed, search endpoint, or bounded MCP tool.The resource cannot enter a workflow.
Identity and policyDecide who may discover, preview, buy, or reuse each resource.Access becomes impossible to govern.
Pricing and paymentPrice the useful unit and verify payment before delivery.Revenue does not track value consumed.
Quality and analyticsMeasure accuracy, successful calls, retention, and refresh cost.Volume hides poor trust or margins.

An llms.txt file can sit in the discovery layer, but it is an emerging proposal, not a universal instruction that makes models crawl, cite, or recommend a site. Visible source pages, consistent facts, conventional SEO access, structured data, third-party evidence, and direct integrations still matter.

Idea 1: Build a Niche Data Refinery

Pick one market where an expensive decision depends on scattered, frequently changing information. Collect the raw material, normalize it, verify it, and sell the decision-ready version.

QuestionFirst answer
WedgeOne niche, one city, and roughly 100 entities. Track only fields tied to a paid decision: services, prices, reviews, complaints, booking flow, ads, hiring, or availability.
First buyerAn agency, consultant, broker, software vendor, or operator already paid to understand that niche.
First versionA reviewed spreadsheet plus a monthly change report with source URLs, observed dates, confidence, and corrections.
First saleShow five findings that alter targeting, pricing, product, or sales decisions. Sell a pilot before a dashboard.
Product ladderReport, recurring dashboard, search API, MCP tool, then metered lookup after machine demand appears.

The defensibility is not scraping. It is a maintained entity model, rights-aware sourcing, historical changes, correction handling, and evidence buyers trust. Respect site terms, database rights, copyright, privacy law, and access controls. Do not collect personal contact details merely because a crawler can see them.

Idea 2: Sell Agent Readiness to Businesses

Run 20 to 50 buyer-intent questions across the AI systems that matter to a company. Compare the answers with official sources. Capture wrong prices, missing use cases, weak evidence, outdated comparisons, unsupported recommendations, and absent citations. The screenshot is the wedge; the corrected source system is the product.

  1. Establish ground truth. Approve current company facts, offers, locations, policies, evidence, and change owners.
  2. Measure the answer surface. Record prompt, model, date, answer, mentions, links, cited domains, errors, and uncertainty.
  3. Trace each failure. Separate inaccessible pages from contradictory facts, thin proof, weak consensus, and ordinary model variation.
  4. Repair the source layer. Improve pricing, use-case, comparison, FAQ, documentation, feed, schema, proof, and changelog pages.
  5. Publish machine interfaces when useful. Add feeds, APIs, MCP, or an experimental llms.txt map when a consumer exists.
  6. Rerun on a fixed cadence. Measure accuracy, qualified mentions, citations, corrections, referral quality, and sales attribution.

Sell the first engagement as an audit and remediation plan, not a promise to "rank in ChatGPT." Model answers are probabilistic and platform-specific. Greg's multi-thousand-dollar examples are pricing hypotheses; scope should follow products, markets, prompts, sources, and remediation hours.

Idea 3: Turn an Expert Archive Into One Useful Tool

Do not begin with "chat with this person's brain." Begin with one authorized expert, one recurring job, one audience, and one outcome. A positioning archive might become a critique workflow. A sales archive might become an objection-preparation tool. A technical archive might become a diagnosis assistant that links every recommendation to its source.

  1. Secure written rights. Define corpus access, outputs, commercial terms, attribution, name and likeness use, revocation, and updates.
  2. Build the corpus. Clean authorized videos, podcasts, newsletters, books, workshops, and examples.
  3. Tag for the job. Label topic, audience, framework, example, constraints, outcome, source, date, and confidence.
  4. Design a workflow. Intake, classify, retrieve evidence, apply the framework, state uncertainty, produce the artifact, and cite.
  5. Evaluate fidelity and utility separately. The expert checks misrepresentation; users check whether it helps complete the job.
  6. Sell through existing distribution. Start as a subscription, licensed team tool, cohort companion, or lead product.

A vector database is not the product. Rights, a narrow workflow, source fidelity, update discipline, and distribution are. Without those, the archive is an unauthorized imitation with retrieval attached.

Micropayment Economics Without the Fantasy

The useful equation is qualified requests x price x payment success = gross revenue. Subtract collection, licensing, normalization, refresh, inference, storage, payment, support, fraud, compliance, and sales costs to reach contribution margin.

Hypothetical monthly usagePriceGross revenueLesson
1,000,000 generic fetches$0.002$2,000Huge traffic can still produce little revenue when each request carries little value.
10,000 verified market lookups$0.25$2,500Decision-linked calls can support better economics.
1,000 specialist workflows$5.00$5,000Outcome pricing can beat raw access when the tool reliably completes valuable work.

These examples are arithmetic, not forecasts. Real prices, demand, payment success, costs, and legal obligations will vary.

Before adding x402, ask whether buyers want anonymity or accounts, whether refunds matter, who carries tax obligations, how budgets are controlled, and whether an API key plus monthly invoice is easier. Machine payment is strongest when buyer and seller have no prior relationship and the resource is valuable enough to buy instantly.

A Thirty-Day Validation Plan

  1. Days 1-3: choose one decision. What is expensive, messy, frequently changing, already funded, and improved by a trusted resource?
  2. Days 4-7: interview five buyers. Ask about the last decision, sources, delay, cost of error, and evidence required.
  3. Days 8-12: make the resource manually. Build one report, readiness audit, or expert workflow with citations.
  4. Days 13-17: sell three pilots. Charge normally. A compliment, waitlist, or free signup is not payment evidence.
  5. Days 18-22: measure acceptance. Did it change a decision, save review time, reduce errors, or generate revenue?
  6. Days 23-26: structure the repeatable core. Define schema, freshness, permissions, evals, and the billable unit.
  7. Days 27-30: expose one bounded endpoint. Create a read-only API or MCP tool. Test x402 only if it solves a real purchasing problem.
Go/no-go gate: continue when at least two buyers pay, use the output in a real decision, and ask for a repeat. Change the wedge when interest disappears at the invoice.

Video Chapters

TimeTopic
00:00Introduction
01:01The old paradigm of the internet
02:57The new paradigm of the internet
03:41What Cloudflare is actually doing
06:33An AI Index for customers
07:34The agent internet stack
09:09Why now is a good time to build
10:29Startup idea 1: niche data refinery
17:10Startup idea 2: agent readiness
23:43Startup idea 3: expert archives
30:36The filter for finding ideas
32:33Closing thoughts

Final Verdict

Cloudflare is not creating businesses by itself. It is making a missing transaction pattern more plausible: identify a machine buyer, state a price, verify payment near the edge, and release a digital resource without building a separate checkout relationship.

The durable opportunity sits one layer below the protocol. Build data that stays clean, facts that stay current, tools that finish a job, and expertise that is licensed and faithfully applied. Start as a service because that is where demand becomes visible. Productize the repeated work. Add agent-native discovery and payment when they improve an already valuable exchange.

The builders who win will not merely own a door. They will own something useful behind it.

Sources and Credits

Common questions

What is x402?
x402 is an open payment protocol built around HTTP 402 Payment Required. A service states its price and accepted payment details, a client submits signed payment proof, and the server or a facilitator verifies and settles the transaction before returning the resource.
Can any Cloudflare site charge AI agents today?
Not through one universally available Cloudflare switch. AI Crawl Control is generally available, while Pay Per Crawl remains in beta, AI Index is in private beta, and the Monetization Gateway has an early-access waitlist. Developers can already build x402-gated HTTP or MCP resources with Cloudflare Workers and the documented libraries.
Does llms.txt make AI systems crawl or recommend a website?
No. llms.txt is an emerging convention for presenting an LLM-friendly map of a site. It can be useful, but broad crawler support is not guaranteed and the file does not create authority, citations, demand, or recommendations by itself.
Which business idea is easiest to start?
Agent-readiness auditing usually has the shortest path because it can begin as a manual service with screenshots, source reconciliation, and a prioritized remediation plan. A data refinery can become more defensible but requires continuous collection and licensing discipline. An expert archive requires explicit rights before product work begins.
Do I need stablecoins to validate one of these ideas?
No. Sell the human-facing service first using normal invoicing or subscriptions. Add API, MCP, or x402 access only after customers repeatedly pay for the underlying data, audit, or workflow.
Will Cloudflare create 1,000 AI millionaires?
That is Greg Isenberg's opportunity thesis, not a measurable forecast. Outcomes still depend on useful resources, legal access, freshness, distribution, margins, and buyers willing to pay.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call