No PriorsNo Priors Ep. 126 | With Cloudfare CEO Matthew Prince
CHAPTERS
- 0:00 – 0:47
Matthew Prince and what Cloudflare actually is (not just a CDN)
Sarah introduces Matthew Prince and frames Cloudflare as critical internet infrastructure. Matthew reframes Cloudflare’s origin story as “a firewall in the cloud” that accidentally made the internet faster—expanding into what the network should have been from the start.
- •Cloudflare’s positioning: security-first, not “just a CDN”
- •Original thesis: move networking/security appliances into the cloud
- •Performance parity goal turned into performance gains
- •Mission: faster, more reliable, secure, private, efficient internet
- 0:47 – 2:08
Cloudflare’s scale, footprint, and the 15-year build
The hosts establish Cloudflare’s scale and product breadth, then Matthew anchors the timeline: launched in 2010 and approaching year 15. The segment sets up why Cloudflare has a privileged vantage point on internet traffic and threats.
- •Cloudflare scale and broad product surface area
- •Launched September 2010; approaching 15 years
- •Infrastructure vantage point across the web
- •Security and reliability as core differentiators
- 2:08 – 3:24
How Cloudflare became dominant: “customer zero” and solving emergent problems
Matthew explains Cloudflare’s growth as a sequence of solving problems created by its own adoption. The company repeatedly became its own first customer and built new capabilities—policy, security, even a registrar—when the ecosystem forced it.
- •Customer-zero approach: build what Cloudflare itself needs
- •Free service enabled data collection and massive scale
- •Scaling brought attacks, abuse, and edge-case users
- •Expansion into policy, security hardening, and registrar services
- 3:24 – 6:34
The web’s interface shift: from search to AI (and collapsing clicks)
Matthew argues the last 30 years of web value creation were driven by search, but usage is moving to AI assistants. As AI delivers “derivatives” instead of sending users to original sources, referral traffic declines sharply—threatening content economics.
- •Search as the dominant value-creation model of the web
- •AI assistants replacing search as the primary interface
- •AI overviews/answer boxes reduce click-through
- •Cloudflare data: far harder to get clicks now vs 10 years ago
- 6:34 – 9:43
Agents change everything: commerce is easier than content (and law may not save publishers)
Elad introduces agents that take actions on users’ behalf, further reducing time spent on the open web. Matthew separates commerce—likely to benefit—from content, where copyright and fair-use dynamics may favor derivative AI outputs, worsening attribution and traffic.
- •Agents reduce direct browsing and consolidate interfaces
- •Commerce agents can be mutually beneficial for sellers
- •Content faces a structural monetization problem
- •Derivative-use/fair-use legal precedent may not favor creators
- 9:43 – 11:44
Building a market requires scarcity: Cloudflare’s default blocking and standards push
Matthew lays out a market-design argument: no market without scarcity. Cloudflare’s “Content Independence Day” initiative blocks AI training crawlers by default, and the company is pushing for clearer crawler disclosure standards to enable fine-grained permissions.
- •Markets need scarcity; “free by default” breaks pricing
- •July 1 announcement: block training crawlers by default (even for free users)
- •Work with IETF/standards bodies for crawler declarations
- •Goal: humans free, bots pay—enforceable controls beyond robots.txt
- 11:44 – 13:05
Robots.txt isn’t enforcement: detecting evasive crawlers and naming bad actors
The discussion extends robots.txt into a world where some bots misrepresent their behavior. Matthew claims some prominent AI companies use tactics resembling hostile actors to evade blocks—and Cloudflare can detect, stop, and publicly call them out.
- •robots.txt is blunt and often ignored
- •Bots can spoof identity and evade publisher restrictions
- •Cloudflare’s strength: detection, blocking, and attribution
- •Promise to publicly expose “misbehaving” AI crawlers
- 13:05 – 16:01
De-emphasizing traffic as value: filling the “Swiss cheese” holes in knowledge
Matthew argues Google taught the web to worship traffic, incentivizing clickbait and rage-driven content. AI models look like a “Swiss cheese” approximation of knowledge; the most valuable new content fills the gaps, suggesting a better incentive system than clicks.
- •Traffic as a proxy for value led to clickbait optimization
- •A/B testing headlines to maximize emotional response
- •LLMs prune redundant content; gaps become most valuable
- •Pay creators to fill knowledge holes, not to drive rage clicks
- 16:01 – 18:03
Spotify as an existence proof: unmet-demand signals and pie-expanding payouts
Matthew uses Spotify/Daniel Ek as an analogy for compensating creators at scale and expanding the market. Spotify surfaces unmet demand via failed queries; creators can target those gaps—an idea Matthew thinks can translate to AI-era content incentives.
- •Spotify now pays >$10B/year to music industry (pie expansion)
- •Surfacing unmet demand via queries with poor matches
- •Creators can profit by producing missing content
- •AI could similarly identify and reward content that fills gaps
- 18:03 – 22:32
Will we run out of human content? The Medici-style risk and how deals should scale
Elad asks whether human-generated expert content will hit a limit as models improve. Matthew worries less about running out of discoveries and more about losing independent creators—leading to a future where a few AI labs employ the journalists/scientists, creating siloed knowledge. He also critiques flat-fee licensing deals that don’t scale with AI company growth.
- •Human discovery continues; the bigger risk is incentive collapse
- •“Medici” scenario: AI labs directly employ creators, creating silos
- •Compensation should scale with model business growth (subscription/ad share)
- •Flat-fee ‘all content for $X’ deals are structurally naive
- 22:32 – 26:03
Cloudflare’s AI infrastructure bet: inference at the edge and network-as-platform
Elad pivots to Cloudflare’s product evolution from caching pages to running inference. Matthew describes early GPU-at-the-edge work (initially ignored, later booming), and argues Cloudflare’s network position makes it a natural layer for agent protocols, security, and payments rails.
- •2020 edge GPU partnership (with NVIDIA) was early; demand arrived later
- •Belief: many models run on-device; bigger ones run at the edge
- •Cloudflare’s position “in front of” much of the internet enables governance
- •Protocols (e.g., MCP) need security rails and payments infrastructure
- 26:03 – 27:21
Open vs. closed models: pragmatic support, skeptical of ‘open source ends the world’
Sarah asks about model openness; Matthew says Cloudflare runs both but leans pro-open based on the company’s open-source culture. He rejects extreme doom arguments, suggesting regulation should focus on downstream real-world capabilities (e.g., pathogen printing), not just model access.
- •Cloudflare supports both closed and open model customers
- •Cultural bias toward open source and openness
- •Skepticism of “open models = catastrophe” framing
- •Regulate dangerous outputs at execution points, not only model weights
- 27:21 – 30:58
How the content marketplace actually emerges: forcing Google into the same rules
Matthew describes the tactical prerequisite: Google must stop being treated as a special case with privileged access to content. Once Google is in the same bucket as other AI companies, scarcity and paywalls for bots can trigger scalable marketplaces—via pooled negotiations, base rates, or micropayments.
- •Google’s historic ‘content for traffic’ bargain is breaking down
- •AI transformations (answer boxes/overviews/Gemini) require a new deal
- •Flatten the playing field so startups aren’t disadvantaged
- •Cloudflare could enable pooled licensing, standard rates, or micropayments
- 30:58 – 45:16
Advice to creators + the timeline to local models and agentic infrastructure
Matthew advises creators to regain control by creating scarcity and negotiating from that position, noting early signs like Google pilots paying news providers. The conversation then expands to local/on-device inference constraints (power efficiency), model compression, GPU utilization, and a coming wave of agent standards, identity/permissioning, and payments (potentially leveraging cryptography/ZK and blockchain-like primitives).
- •Creator playbook: control access, create scarcity, then negotiate
- •Early signal: Google experimenting with paying news providers
- •Local inference depends on power efficiency, utilization, and compression (DeepSeek as a lesson)
- •Agent infrastructure needs standards, identity/permissioning, and scalable payments
