If you are responsible for capacity, security, or spend, the wrong mental model is to picture a new “AI internet” sitting beside the old one.

The right model is simpler and more interesting. The same products, the same internal systems, the same public sites, and the same media pipelines — now touched far more often, by both people and software, with more bytes per interaction when multimodal work shows up. Large language models are the amplifier. Your existing footprint is the speaker.

That matters on a 1-2 year horizon because the compounding starts now. A single human chat session is a thin slice next to video, but enterprise adoption is moving from “try an assistant” to “this is how we work.” Microsoft has reported more than 100 million monthly active users for Copilot-class tools, with Azure growth reflecting how much of that usage lands in hyperscale back ends. Those figures map to new steady-state paths through identity, logging, data residency, and egress policies you already operate. See Reuters coverage of Microsoft’s fiscal 2025 results and capex.

The more dramatic step is agentic workflows — not because science fiction arrived, but because “call the CRM API, search the wiki, file the ticket, attach the screenshot” is the same boring integration graph enterprises have built for decades. Agents simply traverse it more eagerly than humans do. Each step is HTTP or gRPC you already paid to expose; retries and tool-chaining multiply round trips; retrieval-augmented generation adds vector and document fetches on top of ordinary search. None of this requires new physics. It requires admitting that automation increases fan-out across the graph you already maintain.

Multimodal use is where byte volume stops hiding. Text is cheap on the wire compared with images, audio, and especially video. The International Energy Agency has been explicit that modalities differ in intensity. For example, noting that video generation is far more energy-intensive than text generation or AI-enabled search -- which is a useful proxy for how quickly payloads scale when assistants stop being chat-only. That same IEA line of analysis ties the AI build-out to concentrated infrastructure investment and local grid stress, not just to global averages. Start with the October 2024 IEA commentary on data centres and AI and the fuller April 2025 “Energy and AI” report if you want the policy and modelling detail behind the headlines.

Regionally, the story is “global demand, local crunch.” The IEA’s commentary highlights how US investment has surged while China and the European Union also expand — and how, in large economies, data centres may be single-digit percentages of national electricity use but still exceed ten percent in several US states and more than twenty percent in Ireland because capacity clusters where power, land, and fibre converge. That geography matters to CTOs and CISOs in a practical way- your users may be everywhere, but the places your providers train, batch, and serve models are not. Peering, egress paths, and incident blast radii inherit that map whether you chart it or not.

Measurement teams are already flagging the early innings. AppLogic Networks (publisher of the Global Internet Phenomena Report, formerly Sandvine) says AI assistants are not yet dominant by volume — but that their popularity should warn operators that future demand and its traffic footprint remain uncertain. Translate that for the enterprise- if carriers must watch this category, your WAN, cloud egress, and SaaS governance deserve the same discipline. See their March 2025 summary of the 2025 Global Internet Phenomena Report.

Why I believe the projections are reasonable comes from years watching teams who ship at the edge. The internet’s “boring” layers (TLS, caching semantics, rate limits, bot defenses, observability) were tuned for human pacing and predictable automation. Agents that parallelize and treat every document and image as fuel expose those assumptions fast. The optimistic part is that this is a known class of problem- capacity engineering, graceful degradation, API contract testing, tighter prompt data minimization, and honest egress accounting. The slope is what changed.

Treat AI as a multiplier on existing surfaces, not a greenfield. Inventory the APIs and data paths your assistants and agents will touch before you polish the prompt. Model fan-out and multimodal payloads the way you would for a major product launch — because that is what this is. The flood is quiet only until the metrics dashboards catch up.

Further reading and sources I used for this piece

• IEA — “What the data centre and AI boom could mean for the energy sector” (October 2024 commentary investment scale, regional concentration, modality intensity, grid implications).
• IEA — Energy and AI (April 2025 special report).
• Reuters — Microsoft AI spending, Azure scale, Copilot usage (July 2025).
• AppLogic Networks / Sandvine — 2025 Global Internet Phenomena Report overview (AI assistants trend line for operators).
• For the underlying economics of compute efficiency cited by the IEA, see Bick, Blandin, and Deming (2024), NBER Working Paper 32966 and Epoch AI’s hardware trends discussion — useful if you need scholarly footing alongside policy and market sources.