→ Join 500+ data center pros for the Data Center Dealmakers summit! In Atlanta on October 7th. Hyperscalers, power providers, developers, site selection teams, and capital. Request an invitation here!

Now, let’s get to it.

The Internet of AI Is Being Built Around Data Centers, Fiber, and Power

The AI boom is often described as a data center story.

That is only partly right.

In a recent essay, longtime technology analyst Om Malik argues that AI is creating something bigger: a new “Internet of AI,” a private, hyperscaler-led network of data centers, fiber routes, optical systems, subsea cables, and power infrastructure built specifically for AI workloads.

For data center owners, developers, investors, utilities, and infrastructure vendors, the key takeaway is simple:

AI is not just changing demand for compute. It is changing the physical architecture of the internet.

From Consumer Internet to AI Infrastructure

For most of the last 30 years, internet infrastructure was designed around downstream consumer demand: streaming video, websites, social media, gaming, and cloud applications.

AI changes the pattern.

Malik notes that upstream bandwidth consumption is now growing faster than downstream traffic, with average upload usage growing 21.7% year over year, more than twice the rate of downstream growth. But he argues the bigger shift is not happening inside consumer broadband networks. It is happening inside and between data centers.

AI workloads create new traffic patterns:

• GPUs communicating with GPUs inside the same facility

• AI clusters synchronizing across large campuses

• Inference workloads serving many users and agents

• Data center campuses linked by private fiber

• Hyperscaler-owned subsea routes connecting continents

The result is a new network stack built around AI.

The Four Layers of the Internet of AI

Malik describes four layers of the emerging AI network:

1. Inside the AI data center

2. Data center interconnect

3. Private AI internets

4. Planetary AI networks

For the data center industry, each layer has major implications.

1. Inside the AI Data Center: East-West Traffic Takes Over

Traditional data center networks were designed for traffic moving between users and servers. AI training workloads are different.

Inside AI data centers, the dominant traffic flow is increasingly “east-west,” meaning GPUs communicating with each other inside the same facility. Malik notes that AI clusters can require up to five times more connectivity than traditional hyperscale topologies.

That creates pressure on legacy network designs.

The traditional “fat tree” or Clos-style data center network was built for general-purpose cloud workloads. AI training creates massive, sustained bursts of traffic between GPUs that need to stay synchronized. Malik writes that major hyperscalers are now moving away from old network assumptions and designing proprietary fabrics for AI.

Examples cited in the piece include:

• Google’s Virgo Network, designed as a specialized flat, low-latency network architecture for AI

• Meta’s separate front-end and back-end network approach for AI training clusters

• AWS’s random graph topology work intended to reduce switch counts and improve AI workload performance

The implication: AI data centers are becoming purpose-built machines, not generic warehouses full of servers.

2. Data Center Interconnect: Fiber Demand Is Exploding

The next layer is the fiber connecting data centers to each other.

This is where AI demand becomes highly visible in the market for dark fiber, wavelengths, and long-haul connectivity.

Malik cites Zayo’s bandwidth data showing that bandwidth purchased for data center connectivity grew 330% from 2020 to 2024, with total bandwidth purchases more than doubling to 42.4 terabits. Hyperscalers accounted for 57% of metro dark fiber installations between 2020 and 2024.

The scale of orders is also changing. According to Malik’s summary of Zayo’s data, pre-AI long-haul orders were often 8 to 12 fibers. Hyperscalers are now routinely ordering 12 to 48 fiber pairs per route, with some orders reaching 144 to 432 fibers.

This is not just a Silicon Valley or Northern Virginia story.

Malik points to Memphis, where long-haul and metro wavelength demand reportedly grew from 0.3 terabits in 2023 to 13.2 terabits in 2024. Salt Lake City also saw major growth. These are markets where hyperscalers found power, land, and routes to pull fiber.

For developers and investors, this reinforces a critical point:

AI site selection is becoming a three-variable equation: power, land, and fiber.

GET IN THE ROOM:

Power. Land. Capital. Delivery.

Join 500+ data center developers, investors, brokers, utilities, and construction leaders in Atlanta for a day built around what actually moves deals: power certainty, site strategy, capital stacks, and AI-driven demand.

→ Click here to request an invitation

3. Private AI Internets: Hyperscalers Are Building Their Own Backbones

The third layer is the private backbone network.

The largest hyperscalers are no longer simply buying network capacity from telecom providers. They are building and controlling more of the infrastructure themselves.

Malik cites Microsoft as an example, noting that it operates one of the world’s largest backbone networks, spanning more than 500,000 miles of fiber, and added more than 120,000 new fiber miles in the U.S. in one year to extend its dedicated AI wide-area network.

This creates a two-tier system.

Hyperscalers can lock in fiber supply, customize networks, and vertically integrate around their own AI workloads. Enterprises and smaller operators are left with less control, longer lead times, and less negotiating power.

For the broader data center market, the message is clear: the most powerful AI players are not just leasing capacity. They are shaping the infrastructure map.

4. Planetary AI Networks: Subsea Cables Become Strategic AI Infrastructure

The final layer is subsea fiber.

Malik notes that as of early 2026, there are more than 600 active and planned subsea cables totaling more than 1.5 million kilometers of fiber. But the more important shift is ownership.

Before 2012, hyperscalers accounted for less than 10% of total subsea cable usage. By 2024, Google, Meta, Microsoft, and Amazon collectively had stakes in 59 cable systems, with Google leading the group. Malik reports that hyperscalers now control roughly 71% of global subsea fiber capacity, around 80% of trans-Pacific bandwidth, and approximately 90% of transatlantic capacity.

This is a major structural shift.

The internet’s global backbone is increasingly being shaped by a small number of AI-scale companies.

The Binding Constraint: Power

Despite all the focus on chips, fiber, and network architecture, Malik argues that the ultimate constraint is power.

His framing is worth repeating:

Bandwidth follows power. Power does not follow bandwidth.

AI infrastructure is being built where cheap, reliable power is available, and the network is being redesigned around those locations. Malik cites forecasts that hyperscaler capital expenditure in 2026 could exceed $600 billion, with some estimates closer to $700 billion. He also references McKinsey projections that global data center capacity demand could nearly triple by 2030, with roughly 70% of that demand coming from AI workloads.

This is why data center development is increasingly tied to:

• Utility capacity

• On-site generation

• Transmission access

• Grid interconnection timelines

• Fiber routes

• Water and cooling availability

• Permitting and political support

The winning data center markets may not be the largest population centers. They may be the places where power, land, and fiber converge.

What This Means for Data Center Investors and Developers

Malik’s essay is technical, but the real estate and infrastructure implications are straightforward.

1. AI data centers are becoming specialized assets

Generic data center capacity is not the same as AI-ready capacity. Power density, cooling, network architecture, and fiber access matter more than ever.

2. Fiber is becoming strategic, not secondary

Fiber availability and route diversity can materially affect site value. Data center interconnect is becoming a core underwriting item.

3. Power markets are driving geography

The AI buildout is pushing development toward markets with available power, scalable land, and reasonable fiber connectivity.

4. Hyperscalers are vertically integrating

The biggest players are increasingly controlling chips, data centers, fiber, subsea cables, and software layers. That creates opportunities for vendors, but also raises the bar for independent operators.

5. Edge AI remains uncertain

Malik is skeptical that edge inference will become as large as some vendors suggest, although he notes that robotics, autonomous vehicles, drones, and industrial AI could create real edge use cases if physical AI scales.

The Bottom Line

The AI boom is not just a leasing story.

It is not just a GPU story.

It is not even just a hyperscale data center story.

It is a full-stack infrastructure story involving data centers, fiber, subsea cables, power generation, network architecture, and global capital allocation.

For the data center industry, the “Internet of AI” thesis points to a new competitive map.

The winners will not simply be the companies with the most servers.

They will be the companies and markets with the best combination of power, land, fiber, capital, and execution.

Source: Om Malik, “Say Hello to the Internet of AI,” On my Om, May 4, 2026.

→Try our resources for data center pros:

Advertise Listings | Share a Closing Announcement | Search Properties | Data Center Real Estate News | Comps | Professional Directory