indigo: The Agentic Internet is rising

On this trip to the Bay Area, I focused on attending two events, one by Vercel and one by Cloudflare. My biggest takeaway: the Agentic Internet is rising. On the streets of San Francisco and Silicon Valley, aside from the ubiquitous Muse and Grok Bot ads, what you see is competition among various Agent Infra players. Five years from now, Agent network traffic will be 1,000 times that of humans, and human behavior on the network will become a "rounding error" 🤔

Actions happen on the network

Models are the foundation of AI, but today's world is dominated by Agents. Models make decisions—that's a GPU problem; Agents take actions—that's a CPU and network problem; GPU clouds are built for decisions, while actions happen on the network.

This is exactly the opportunity for Agent Infra, and also where Cloudflare is positioned. Next, we will see a wave of developer platforms built around Agents, such as Cloudflare Workers, Vercel, and Railway, among others.

New business models, new battlegrounds for attack and defense

Agent Payment will become the new business model of the Agentic Internet. Agents' frequency and intelligence far exceed those of humans, so cyber offense and defense will become the norm. This is the most important issue of the Agentic Internet, and also its biggest opportunity. The service interface of the internet will also become fully Agentified. All enterprises and services that cannot adapt to the Agentic era will be eliminated in this round of paradigm shift.

Robotic beyond Agentic

Spatial intelligence has another path to realization: large language models can do it too. GPT-6 Astra has already integrated the Robotic modality, which will affect the technology roadmaps of many spatial intelligence and robotics model startups. Agents will turn the model itself into a multi-layered architecture, and capabilities from vision to action will ultimately be integrated into this architecture.

The self-cycle of silicon

An Agent that can write code and call tools can do almost anything, including improving itself and designing the carrier that runs it: chips. OpenAI is jointly developing GPT-Synopsys, a vertical large model dedicated to chip design, with Synopsys. If wafer fabs can also achieve fully automated control, then the self-cycle of silicon will officially close the loop.

A new method for AI to do scientific research

On the last day, in a conversation with Teacher Wan Weigang, we talked about new methods for AI Agents to solve scientific research problems. They are not yet capable of making entirely new scientific discoveries, but they can merge clues across tens of thousands of human papers and find those subtle differences, thereby helping us discover new solutions.

Just three days ago, OpenAI used 1,946 GPT-Rosalind agents to link million-scale mouse whole-brain gene perturbation data with 7,586 interdisciplinary literature citations in 32 minutes, finding new mechanistic clues for neurodegenerative disease.

In closing

While everyone is still worried about slowing growth in token consumption, Agent usage is rising. When Agents' use of the internet is a thousand times that of humans, we will welcome entirely new network structures, network services, security offense and defense, and entirely new model architectures—to solve scientific research problems, and also to solve real-life and service problems.