Analysis by veteran technology journalist Om Malik outlines how artificial intelligence is reshaping the internet’s hardware and network topology.

Asymmetry of traffic

For three decades the internet was optimized for downstream content delivery to homes and devices rather than for heavy upstream usage from endpoints.

Outbound traffic now grows at 21.7% annually, roughly twice the pace of incoming traffic, driven primarily by AI agents sending data to cloud services.

According to Nokia at MWC 2026, AI generates 77 exabytes of traffic each month, representing 20% of global network traffic and expanding at 23% per year.

Inside data centers

Traditional north–south user-to-server patterns are giving way to dominant east–west flows between GPUs inside AI clusters and across racks.

AI deployments demand about five times the bandwidth of conventional data centers, and legacy architectures struggle to accommodate these increases.

Google announced the Virgo Network to separate AI traffic and provide a flat, large-scale topology, while Meta and AWS introduced internal fabrics and new routing topologies respectively.

Undersea cables and ownership

Ownership of transoceanic capacity has shifted markedly since 2012, when carriers controlled about 80% of undersea resources.

Today Google operates 33 cables, owning 17 outright; Meta has 16, Microsoft 6, and Amazon 4, together accounting for roughly 71% of global undersea capacity.

These networks carry about 80% of transpacific traffic and approximately 90% of transatlantic flows, concentrating the physical layer among hyperscalers.

Investment and sector implications

Hyperscale operators plan significant capital spending, with projected expenditures of $600–$700 billion in 2026 to expand networks and compute capacity.

Network builders such as Zayo are laying 15,000 new route miles of fiber, while Microsoft pilots hollow fiber technologies claiming roughly 47% speed improvements over conventional options.

AI platform providers like Anthropic and OpenAI rely on hyperscalers for infrastructure, underscoring that product leadership in AI does not equate to ownership of the underlying network layer.

The changing topology affects related ecosystems, including blockchain networks, decentralized protocols and agent layers that increasingly interact with large-scale infrastructure.