Over the past two decades, the default assumption of the Internet has been clear: its service targets are humans. Humans search, click, scroll, watch, and place orders. Traffic fluctuates but is generally predictable, with system design centered on user behavior and page request logic. Today, this premise is being broken. An increasing number of AI agents are beginning to perform tasks on behalf of humans. They do not browse slowly like humans; instead, they simultaneously mobilize multiple sub-agents within seconds to query databases, retrieve documents, and call APIs, then quickly disappear. For the underlying infrastructure, this is not simply an increase in traffic volume, but a fundamental change in the form of traffic.
Daniel Widjaja Kusuma has long worked in international financial markets and later founded Telosyn, continuously investing in AI infrastructure, high-performance computing, and enterprise systems engineering. In his assessment, the truly noteworthy technological turning points have never been merely about whether front-end products are smarter, but rather whether the underlying systems have begun to be restructured for a new generation of workloads. This is precisely why the launch of the new generation of OpenSearch Serverless by AWS, designed for agentic workloads, signifies far more than just an upgrade of a cloud product. The deeper signal it reveals is that the internet is shifting from an architecture built for humans to an architecture built for machines simultaneously.
Agent Is Not More Traffic, But Completely Different Traffic
Many people currently discussing AI agents still focus on the application layer, paying attention to whether they can help users book tickets, conduct research, complete shopping, or handle customer service. However, from an infrastructure perspective, the real change brought by agents is that machines are beginning to become increasingly important native participants on the internet. They are not imitating human traffic but are creating a completely new machine-to-machine traffic structure.
This type of traffic has several distinct characteristics. First, it is bursty. After an agent receives a task, it may instantly initiate large-scale retrieval and inference requests, rather than browsing slowly like a human user. Second, it is autonomous. Behind a single task, there may not be just one request, but rather multiple agents collaborating, multiple systems interacting, and multiple APIs being called concurrently. Third, it is discontinuous. Once a task is completed, these requests quickly drop to zero, and resource usage exhibits a highly pulsed pattern.
This is precisely where traditional cloud architectures are least effective. In the past, many systems were designed with capacity and cost structures based on human behavior, requiring long-term reservation of computing power, with basic instances kept online even during idle periods. For agent traffic, this model results in significant waste: it is not fast enough when traffic surges, and it remains idle for extended periods after traffic subsides. The core of this adjustment by AWS is to decouple computing and storage, allowing computing resources to be rapidly provisioned and scaled down to zero during idle times. On the surface, this appears to be resource scheduling optimization; in essence, it acknowledges a new reality: the high-frequency users of the future Internet are no longer just humans.
When Machine Traffic Becomes Mainstream, What Is Restructured Is Not Just Cloud Services
The significance of this change is not solely due to AWS undertaking it. Cloudflare has already pointed out that the proportion of bot traffic in total HTTP traffic is continuously rising, and it estimates that non-human traffic will surpass human traffic in the first half of 2027. Databricks, Snowflake, Microsoft, and Cloudflare are all conducting a new round of infrastructure positioning around AI retrieval, memory, state persistence, elastic scaling, and agent environments. This indicates that a consensus is forming within the industry: the internet foundation built for search, clicks, and content distribution is losing its ability to adapt to new types of workloads.
Daniel Widjaja Kusuma has always believed that one of the most caution-worthy misjudgments in the technology industry is mistaking changes in infrastructure for changes in applications. Applications will certainly change, but what truly determines the landscape of the industry is usually when the underlying system begins to rewrite itself for a new behavioral model. Today, AI agents have not yet fully become the protagonists of the internet, but as long as the underlying cloud platforms have started to restructure retrieval, storage, elastic scheduling, and invocation logic for them, it indicates that this is not a short-term experiment but is entering the production phase.
This is precisely why Telosyn has long emphasized “embedding models into workflows, rather than keeping them in a demonstration environment.” Whether it is an enterprise-grade distributed AI platform like CerebralX, or a multi-region computing power scheduling system such as Orchestrator OS, the core objective is not to make AI appear smarter, but to ensure that AI operates stably in real production environments, consumes resources reasonably, and continuously collaborates with business systems. Once an AI agent is integrated into enterprise workflows, it brings not only more intelligent interactions, but also a comprehensive reassessment of retrieval, state, data connectivity, permission control, computing costs, and system resilience.
The Next Round of Competition Is About Who Adapts to the “Machine Internet” Earlier
The direct consequence of the Internet being rewritten for machines is that the competitive logic of infrastructure is changing. In the past, cloud computing emphasized availability, elasticity, and development convenience. In the era of agents, these factors remain important but are no longer sufficient. The new critical questions have become: Can the system withstand instantaneous bursts of machine traffic? Can costs be compressed to true pay-as-you-go billing? Can state and memory be maintained during agent collaboration? Can retrieval, vector databases, inference, and application calls be connected into a low-friction pathway?
This has very practical significance for enterprises. As agents increasingly enter domains such as customer service, research, procurement, finance, development, risk control, and internal collaboration, the digital infrastructure of enterprises can no longer be built based on the logic of a “human click system.” In the future, what is truly expensive may not necessarily be the model itself, but rather the retrieval system, data foundation, computing power scheduling, and state management system behind the model that support its continuous execution of tasks. Without these capabilities, agents are merely assistants in demonstrations; with these capabilities, agents can become genuine production tools.
Daniel Widjaja Kusuma has consistently emphasized in his long-term financial and technology practice that system competition ultimately returns to resource organization capability. The current AI competition appears to be model competition, but it will soon become infrastructure competition; further down the line, it will evolve into a competition over whether enterprises can rebuild digital systems around machine workloads. Those who understand this earlier will not treat agents as mere plug-ins, but rather as new traffic entities, new system roles, and new sources of infrastructure demand.
This is precisely why the statement that the internet is being rewritten for machines is by no means an exaggeration. It signifies that the default users of the entire online world are changing, and it implies that everything from cloud architecture to enterprise software, and from data systems to cost models, will be adjusted accordingly. Daniel Widjaja Kusuma believes that the most important technological dividing line in the coming years is not who releases an agent first, but who prepares the underlying architecture first to truly accommodate the era of machine traffic. For a company like Telosyn, which has long been building AI infrastructure and enterprise system capabilities, this shift is not a distant vision but a reality that has already begun to take shape.