Why One Internet Pioneer Thinks the Original Model Broke
Summary
Barrett Lyon argues that the Internet’s original open model broke as communication consolidated around corporate servers and policy migrated into the network layer. Carrier-grade NAT weakened peer-to-peer applications, while identity and age controls wrongly treat “an IP address” like “a driver’s license.” His central question for WhatsApp, Signal, iMessage, Telegram, and VPNs is deceptively simple: “Whose server is that?”
Docs.net is Lyon’s attempt to restore direct, private communication through a programmable parallel network. Its internal ecosystem includes 196 domains that do not exist on the Internet, API-driven provisioning, mesh and public addressing, and peer-to-peer calls, video, chat, and file transfer. Lyon says phones can send files directly, while the system can transfer even 20GB files between branches over a private network with low latency and “up to 3 to 4 levels of encryption.”
The investable distinction is infrastructure ownership, because a privacy promise is only as credible as the hardware and counterparties beneath it. Lyon’s conclusion from examining VPN providers was categorical: “You cannot set policies if the equipment is not your own.” Docs.net therefore owns the stack “from software to wiring,” and Lyon describes separate US and European operations, with a Swiss company responsible for Europe. The company also makes its ownership and governance visible.
AI has allowed a roughly 12-person company to build 26 sites around the world in six months without network or system administrators. Agents handle network operations on company-owned hardware, with contacts and inferences processed internally. The security thesis is organizational as much as technical: fewer employees with access to the network means a smaller insider-threat surface.
Lyon sees the combination of ubiquitous telemetry and scalable, “coin-operated intelligent systems” as the genuinely dangerous AI transition. Individual tracking events may appear harmless, but aggregated data can expose income, routines, purchases, childbirth, divorce, and other life changes—sometimes “before your spouse does.” Once millions of intelligent units can interrogate that dataset, surveillance and targeted persuasion become dramatically cheaper.
The episode’s contrarian infrastructure call is to defend data centers while attacking the tracking systems built around them. Lyon and Joel de la Garza challenge portrayals of data centers as continuously “drinking” water, explaining that sealed systems require replenishment when equipment breaks down or water evaporates. They argue that these facilities create skilled jobs and improve nearby infrastructure. Their sharper conclusion: “What people should be targeting isn’t data centers, but tracking systems.”
Lyon’s latest project extends a career built around turning personally encountered Internet failures into shared infrastructure. A high-school DDoS problem led to Prolexic; DefenseNet became F5 Silverline; BitGravity addressed video-delivery waste; and a $50 bet produced the Opte Internet map, with two pieces in MoMA’s permanent collection. The recurring pattern is turning a personally encountered failure into shared infrastructure.
Deep dive
1. An engine failure made abstract risk brutally concrete
De la Garza opens with the story behind Lyon’s unusual commute: when a Bay Area drive expanded from 15 minutes to almost 45, Lyon began flying between work and his family near Tahoe. He accumulated about 1,300 hours on a Sirus aircraft before an engine failure after departing San Carlos forced him to decide where to crash so he would not harm people.
Lyon considered deploying the aircraft’s ballistic parachute—crash reports suggested pilots died because they declined to use it—but saw a large former military airport within gliding distance and landed there instead. Only when a firefighter took his hand did “a sudden wave of fear” arrive; until then, the problem had been procedural.
The failure was not a single clean event. Lyon says an intake-control component broke, metal entered the manifold and apparently affected cylinders 3 and 6, the newly installed electronic ignition failed, and the backup magneto also stopped working under higher RPM. He eventually sold the plane and concluded that any future aircraft should have “fewer moving parts.” De la Garza also suggested reserve pilots and systems, which Lyon endorsed.
2. The Internet’s policy layer is being asked to do the wrong job
Docs.net reconnects Lyon with his teenage hacker identity and first business: he sold shell accounts for $5 a month to as many as 1,000 users before cloud services existed, then left the domain in place after closing the company. The new venture grows from the same fascination with protocols, but against an Internet Lyon now sees as stagnant: “There isn’t much new, and everything is built on top of what already exists.”
His technical complaint is carrier-grade NAT, which “is destroying the ability of applications to perform cool functions” by obstructing direct connectivity. With powerful edge Wi-Fi, mesh protocols, and peer-to-peer networking now practical, Lyon wants to recover the earlier Internet’s promise as an unrestricted “marketplace of tremendous ideas.”
Lyon’s policy objection is equally architectural. Policy movements increasingly ask networks to enforce age, identity, and content restrictions even though “the network does not perform identity-verification functions” and “an IP address is not a driver’s license.” If access must be restricted, he argues, the decision belongs in another layer rather than being embedded indiscriminately in connectivity.
Misclassification makes those controls more dangerous: a local information provider can mistakenly identify a user’s location and apply the wrong jurisdiction’s policy, while automated filters may brand legitimate services as adult content. De la Garza cites a partnership-oriented chatbot business that half the classifiers treated as adult; the point is not merely inconvenience, but how brittle labels become infrastructure-level exclusion.
3. Docs.net replaces server-mediated messaging with private edge-to-edge exchange
Lyon describes conventional VPNs as encrypted connections to rented exit servers: useful for separating identity from an IP address, but fundamentally a workaround. Docs.net instead describes a network adjustable through APIs, allowing domains, public or mesh addresses, and policies to be created in milliseconds across an internal ecosystem containing 196 domains that do not exist on the Internet.
Account creation requires solving a simple puzzle to receive a token. Users can change the token or delete the account and create a new one. Inviting a friend quietly exchanges certificates and establishes trust without presenting users with cryptographic machinery—the design lesson being that “if you tell people PKI, it does not work properly.” Lyon jokingly calls the result “Private Key Infrastructure,” because the public-key layer is abstracted behind an ordinary chat interface.
Lyon’s critique of WhatsApp, Signal, iMessage, and Telegram is that they appear to send data to another person while actually uploading it to a server for retrieval. “Whose server is that?” is therefore the load-bearing question. On Docs.net, phones communicate over the mesh and act as servers themselves: “No intermediary or server is needed at all.”
He says this architecture supports direct calls, video, chat, and file transfers without exposing phone numbers, email addresses, or names. The system can transfer files as large as 20GB between branches over a private network with low latency, with up to 3 to 4 levels of encryption. Privacy, however, is not presented as immunity from law: Lyon explicitly says the network must still obey applicable laws “as much as possible.”
4. Privacy promises become credible only when the operator owns the stack
De la Garza’s trust challenge is sharper than a generic preference for encryption: a company claiming privacy is still unconvincing if it does not disclose who owns it, why it exists, or who is involved. The episode’s cautionary example is the secure-messaging service the FBI created and sold exclusively to organized-crime groups; exclusivity and security branding did not reveal the actual operator.
Free VPNs create the same unresolved principal-agent problem. “If you do not pay, you become the product,” Lyon says, but even a paid provider’s “no logging” policy leaves questions: are the servers owned or leased, who is the lessor, what happens beyond the cage, and can the provider truly enforce its stated rules?
Lyon’s eventual conclusion was that “you cannot set policies if the equipment is not your own.” Docs.net consequently controls software, equipment, wiring, routing, and the AI components running on its own hardware. Lyon also describes a regional split: the US operation is based under US law, while a Swiss company handles European operations. The team emphasizes transparent ownership and governance rather than asking users to trust a faceless privacy label.
5. AI compresses operations while telemetry magnifies the threat
Docs.net built 26 sites around the world with redundant routing in roughly six months using a team of about 12. Lyon says it employs neither a network administrator nor a system administrator: “Everything is handled by AI.” The agents operate inside a secure environment on Docs.net hardware, keeping contacts, inferences, and other operational data away from large external AI providers.
De la Garza says the most serious cyberattacks are tied to insider threats, and Lyon agrees. A small staff means fewer people can access the network; physical equipment is monitored, and the camera company’s agent raises alarms on movement or suspicious activity. The team even considered a Mission: Impossible-style laser system, but the data-center lights could not be turned off to make the beams visible.
The larger danger is AI joined with advertising telemetry. Lyon argues advertisers may already be “far ahead” of intelligence agencies: trackers across apps and websites can aggregate coffee habits, wake times, income, shopping patterns, and major life events. The objection is not whether anyone cares about one Target underwear purchase; it is that the combined record reconstructs a person’s life.
Scalable, “coin-operated intelligent systems” change the economics of exploiting that record. Millions of such units can be deployed alongside vast telemetry, with intelligence enabling effective targeting. Lyon’s prescription is to dismantle tracking at the application and network levels and restore an Internet where people can connect and build without continuous advertiser or government observation.
6. The data-center backlash targets the machinery society still needs
Lyon and de la Garza regard much data-center commentary as technically uninformed or manipulated. They challenge portrayals of facilities continuously “drinking” water: Lyon says water needs to be replenished when equipment breaks down or evaporation occurs, while de la Garza describes a sealed system that must be refilled if a chiller failure requires draining it.
Their economic case is equally direct. Large facilities bring well-paid technical jobs and attract high-tech talent; de la Garza argues that infrastructure and connectivity also improve around them. Despite public controversy, Lyon says visits to more than 20 sites revealed no crowds of protesters—sometimes little more than “a single rabbit in the field.”
The pair suspect public opinion can be manipulated for commercial gain when anti-data-center narratives appear systematically across platforms, although they do not identify a responsible party. Their stated conclusion is narrower and categorical: weakening data-center infrastructure at this technological turning point is “the wrong decision,” while eliminating ad tracking would make life substantially easier.
7. Lyon repeatedly turns personal failures into shared infrastructure
Lyon traces Prolexic to friends attacking his high-school server. His survival techniques became a SANS presentation, people began contacting him, and the problems escalated from attacks on two watch companies during Christmas sales to gambling sites disabled before the Super Bowl. Because few firms could afford the necessary equipment, Prolexic offered it as a rental service—one of the first cloud services. The business survives as Akamai Prolexic.
DefenseNet later became F5 Silverline, while BitGravity applied the same systems instinct to video delivery. Lyon recalls how progressive downloads wasted expensive bandwidth when viewers skipped into a video’s middle; the company built technology around large files, large objects, and video-specific distribution.
The Opte project began as a $50 lunchtime bet that Lyon could adapt internal-network mapping tools to the entire Internet. After a bicycle accident left him unable to walk, he had time to code it; the resulting project, including its maps and animation, toured museums, and two pieces entered MoMA’s permanent collection. His reaction the next morning was not triumph but impostor syndrome: “Will I ever feel like an artist again?”
Lyon also describes a book that he thinks covers the beginning of Prolexic and tracking Russian cybercriminals, though he says the events only looked ordinary at the time. That mixture of obsession and uncertainty closes the episode. Whether mapping the Internet, building Burning Man projects, or reconstructing private networking, Lyon’s recurring belief is that passion creates “an aura” others can feel. Docs.net, he insists, remains “just the beginning.”