Should We Start Preparing for an Agentic Internet?

A New Participant Online

A normal search journey leaves the work with the person. We open results, compare claims, move information between tabs, fill a form, and decide which next step belongs to us. An agent can carry more of that sequence. It may gather approved product data, compare options against a stated need, ask for confirmation, and use a supported purchasing or booking route.

This does not make the open web irrelevant. Agents still need trustworthy sources, current terms, identifiers, prices, availability, policies, and routes into the business. In fact, weak information becomes more expensive when software acts on it. A vague service page may confuse one visitor today; tomorrow it may cause several assistants to omit the business or represent it badly.

We therefore see the agentic internet as an additional access layer. Websites remain useful for explanation, trust, branding, discovery, and direct control. Machine-readable information and bounded tools let an authorised agent work with the same business. The winning architecture is likely to support both rather than force a choice between them.

Compute Is Physical

Agentic tasks ask more of infrastructure than a short text response. An agent may reason through several steps, retrieve information, call tools, inspect results, retry a failure, and keep enough context to remain coherent. Each useful action can involve more inference, memory, networking, and storage. Multiply that by millions of people and businesses and the demand stops looking like ordinary web traffic.

Memory supply is already under pressure. In its fiscal third-quarter 2026 remarks, Micron said it did not yet have line of sight to when industry memory supply would catch rising demand, even as it expected gradual improvement in 2028. High-bandwidth memory for AI systems also competes for manufacturing capacity with other memory products. Calling this a simple RAM shortage misses the shape of the problem, but businesses should not assume unlimited hardware supply either.

Data-centre capacity has similar bottlenecks. New sites need land, grid connections, generation, cooling, transformers, networking, skilled construction, and long-term finance. OpenAI's own Stargate material describes an effort to bring capacity online faster because demand is outrunning the available compute base. The shortage is not one empty warehouse. It is a chain of constrained systems.

Training frontier models remains expensive, and widespread agent use makes inference a large continuing cost rather than a one-off bill. Prices can still fall while total spending rises because people run more tasks, longer tasks, and more capable models. We expect the unit economics to improve, but a business plan should work with today's costs and leave room to change models or providers later.

The Next Hardware Wave

Software advances arrive in waves because bottlenecks move. The current wave has concentrated on finding useful AI applications. At the same time, chip designers, model labs, data-centre operators, and energy companies are working on the systems underneath those applications. Better kernels, smaller models, routing, caching, specialised accelerators, and improved memory movement can all reduce the resources needed for a result.

OpenAI's Jalapeño chip is a useful example. OpenAI and Broadcom designed the accelerator specifically for language-model inference, and OpenAI published early performance-per-watt results in August 2026. Initial deployment is planned by the end of 2026 as part of a broader platform. We treat that as evidence of where the industry is investing, not a promise that every agent call will suddenly become cheap.

Lower cost per task often creates more tasks. Faster models make real-time agents practical; cheaper inference encourages longer workflows; better hardware expands the number of people who can use it. The capacity problem can shrink at the unit level while growing in total. That is why we believe prices will trend down without pretending infrastructure demand will disappear.

Efficiency Needs a Destination

The environmental question cannot be reduced to whether one text prompt uses a small amount of electricity. The International Energy Agency says energy use per AI task has been falling rapidly, but reasoning, video, and agentic workloads can require hundreds or thousands of times more energy than a simple text generation. Usage is also expanding. Both facts matter at once.

The IEA projects electricity generation for data centres rising from 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030 in its base case. Renewables are expected to meet nearly half of the additional demand, but natural gas and coal still supply a substantial share of near-term growth. An efficient chip plugged into a carbon-heavy grid is an incomplete answer.

We want AI to remove repetitive work and make useful capability more accessible. That benefit has to justify the infrastructure behind it. Providers need better disclosure, efficient models, cleaner power, careful siting, water responsibility, hardware reuse, and workloads that accomplish something worth the resources. Businesses also have a role: choose the smallest model and shortest workflow that reliably complete the job instead of treating compute as invisible.

Our Practical Preparation

Our preparation starts with information. We publish structured service, product, course, policy, and business details, maintain clear page metadata, and provide an `llms.txt` route that points machines towards useful public material. This does not guarantee an answer or a citation. It gives an agent a cleaner source to retrieve than a collection of decorative pages with inconsistent facts.

The next layer is action. We are building MCP surfaces that can expose narrow business capabilities to compatible agents, with permissions and deterministic results around the operation. Commerce is one obvious direction. OpenAI is already developing richer product discovery and an Agentic Commerce Protocol. A business should be able to let an authorised customer agent find an appropriate offer or complete a supported purchase without opening arbitrary access to its backend.

Inside the business, we use agents to reduce busy work: research, drafting, code changes, checks, structured data work, and repeatable operations. Products such as our Web Conversation Engine and Fonte UI come from that experience. They are not ideas pasted onto an AI trend. They are our attempts to turn what we learn while helping ourselves into systems another business can actually operate.