The conventional read on Vinton Cerf’s remarks at the Open Frontiers conference this week is optimistic: the father of the internet sees a bright future for AI. The less comfortable read — and the one Cerf actually delivered — is a structural warning that the AI industry is courting the same closed-system traps that nearly strangled the early web.
Cerf, 83, co-invented the TCP/IP networking protocols that form the backbone of the modern internet, speaking on a panel alongside computer scientists and Databricks co-founder Matei Zaharia at the Open Frontiers conference. His message, according to the panel discussion, was not a victory lap but a diagnostic: three principles that determined the internet’s success, and which AI has yet to satisfy.
The Assumed Story and the Overlooked Tension
The headline narrative is that AI is on a trajectory resembling the early internet — a chaotic, fast-moving technology that will eventually settle into something transformative and universal. Cerf himself invited the comparison: “AI reminds me of the early days of the internet,” he said, according to the panel’s account.
But the parallel cuts both ways. The early internet was not inevitable. It succeeded specifically because its founders made a deliberate and politically contested choice: to keep the underlying protocols open, ungoverned by any single company, and accessible to anyone willing to follow shared technical rules. “It only worked because it was going to be distributed to begin with,” Cerf said. “We just said if you can find somebody to connect to and you follow the rules of the protocols, it should work.”
That choice was not the industry default in the 1980s and early 1990s. Proprietary network architectures from IBM, Digital Equipment Corporation, and others competed directly with the open TCP/IP stack. The internet won not because it was technically superior in every dimension, but because open access compounded faster than any closed ecosystem could. The question Cerf is implicitly raising — and which the AI industry has not squarely answered — is whether today’s AI builders are making the same foundational bet on openness, or whether they are defaulting to the losing side of that historical trade-off.
Cerf’s three principles — open standards, agent-to-agent communication protocols, and platform-layer thinking — are individually familiar talking points inside AI research circles. Taken together, however, they describe a single systemic risk: that AI’s most powerful capabilities may be locked inside incompatible proprietary silos at precisely the moment that multi-agent architectures require the opposite. The internet’s lesson is that interoperability problems compound quietly and then break catastrophically at scale. The current race to ship closed agentic products may be seeding exactly that failure mode, even as individual products appear to be flourishing. This concern is consistent with broader anxieties documented among economists and institutional observers about the structural turning point AI is approaching in 2025.
Three Principles, One Unresolved Problem
Open standards over closed systems. Cerf’s first principle is the most direct. The internet’s universality derived from a single deliberate design choice: no entity owned the protocol. A university research network in California, a Department of Defense lab, and a commercial internet service provider could all exchange data because they spoke the same technical language — not because any one of them mandated it, but because the rules were public and joinable. Cerf told the panel that AI is “approaching an inflection point” at which the growing number of AI agents will demand equivalent “interoperability and standardization.” The industry’s current posture — where major frontier labs each maintain proprietary agent frameworks, APIs, and memory architectures — is structurally inconsistent with that requirement.
Agents need a better language than English. Cerf’s second principle is the one most likely to surprise practitioners. Natural language, he argued, is insufficient for reliable agent-to-agent coordination. “I don’t think English is going to be the best choice,” he said. “There is ambiguity, and I think precision for inter-agent interaction is going to be very, very important.” Human languages are optimized for human cognition: they are rich in context, inference, and polysemy — the same word carrying multiple meanings depending on situation. Those properties, which make language expressive for people, become liabilities when two automated systems need to confirm a shared commitment unambiguously. “An agent really needs to be sure the other agent understands what it is that they just agreed to do together,” Cerf added. This framing aligns with ongoing research into formal agent communication languages and structured task protocols, though no consensus standard has yet emerged from the research community or standards bodies. The challenge also connects to a subtler concern: if AI systems already reshape how humans use language, poorly specified inter-agent protocols could propagate ambiguity at machine speed.
Platform thinking over product thinking. Cerf’s third principle is perhaps his most broadly applicable observation. “A lot of the successes come from enabling technologies,” he said, “whether it’s a platform or some other fundamental element that others can build on.” Google, Amazon, and Netflix did not build the internet; they built on top of it — and that layered architecture is what allowed the web’s value to compound across millions of independent developers. Cerf applied the same logic directly to AI: “If you really wanted to look for impact, think about things that enable other people to do things that they want to do.” The implication is that AI systems designed as vertically integrated products — doing everything in-house, keeping APIs restricted, controlling the full stack — are structurally less likely to achieve internet-scale impact than systems designed explicitly to be built upon.
The Strongest Counterargument
The most credible objection to Cerf’s framework comes from the history of standards bodies themselves: open standards are notoriously slow, politically fraught, and frequently captured by incumbents. The ISO OSI networking model, which competed with TCP/IP in the 1980s, was in many respects technically superior — and it was the product of a formal international standards process. It lost, not because it was open, but because TCP/IP was already deployed and good enough. Critics of the “open standards will save AI” thesis — a position associated with researchers skeptical of top-down governance of fast-moving technologies — argue that premature standardization risks locking in inferior architectures before the field understands what it actually needs. Under this view, the current diversity of proprietary agent frameworks is not a bug but an evolutionary phase: the Cambrian explosion that precedes natural selection of the most useful approaches.
This is a fair objection, and Cerf does not fully address it in the reported remarks. However, it does not obviously weaken his core conclusion. The TCP/IP case suggests that the critical variable is not who writes the standard but whether the winning protocol is openly joinable. A proprietary protocol that becomes the de facto standard through network effects — as has happened in messaging, social networking, and payments — typically extracts rents and creates fragmentation over time. Cerf’s warning is less about the mechanics of standardization than about the structural incentive to keep winning protocols closed. That incentive is, if anything, stronger in AI than it was in networking, given the capital concentration now visible across frontier model development. The concerns raised by leading economists about AI’s market concentration risks reinforce why this structural question is not merely technical.
Where This Ends Up
The most likely near-term outcome is a fragmented agent ecosystem that eventually forces standardization from the bottom up — driven not by foresight but by the practical pain of enterprise customers who cannot afford to rebuild integrations every time a vendor changes a proprietary protocol. That is precisely how TCP/IP won: not through a mandate, but through the accumulated cost of the alternative. If Cerf’s analogy holds, the AI industry is probably two to four product cycles away from that inflection point, and the protocols that achieve early adoption momentum — whatever their provenance — will be very difficult to displace afterward.
The less likely but consequential alternative is that a major regulatory intervention, particularly from the European Union under emerging AI interoperability provisions, forces open-standard compliance before market dynamics do. That path has historical precedent in telecom and finance, and it would accelerate the timeline Cerf is describing — though it would also introduce the standards-capture risks that TCP/IP’s grassroots adoption managed to avoid. The condition that tips the balance toward this second outcome is whether a high-profile multi-agent coordination failure — a significant security incident or a large-scale commercial breakdown — arrives before voluntary industry coordination does. The stakes are high enough that, as the labor market implications of agentic AI continue to sharpen, getting the foundational architecture right is no longer just an engineering question.
