AT&T is the smartest AI company

Everyone is racing to build the next great AI model. AT&T just committed over $250 billion to making sure all of them have somewhere to run. On March 10, 2026, marking the 150th anniversary of the first telephone call, AT&T announced a five-year plan to expand fiber, 5G, and satellite infrastructure across the United States. No new foundation model. No autonomous agent platform. Just pipes, towers, and cables. It might be the smartest AI bet anyone has made.

The infrastructure thesis

Every flashy AI use case, from real-time video translation to autonomous supply chains, bottlenecks at the same point: connectivity. Edge AI needs 5G. Cloud AI needs fiber. Autonomous agents need always-on, low-latency networks. The model doesn't matter if the data can't move. AT&T's $250 billion isn't a moonshot. It's a calculated wager that no matter who wins the model race, they'll all need bandwidth. The spending covers fiber broadband acceleration, 5G home internet expansion, satellite connectivity through its partnership with AST SpaceMobile, and continued investment in FirstNet for first responders and AI-driven network security. The company already invested more than $145 billion between 2019 and 2023 on wireless and wireline networks. This next phase is about staying ahead of a demand curve that AI is bending upward.

Cheaper AI means more AI means more infrastructure

There's a useful economic concept here: the Jevons paradox. In the 1860s, economist William Stanley Jevons observed that as steam engines became more efficient, coal consumption didn't fall. It rose. Better efficiency made coal-powered machines viable for more applications, and total demand exploded. The same dynamic is playing out in AI. When DeepSeek released its R1 model in early 2025, claiming performance comparable to OpenAI's frontier models at a fraction of the cost, the market panicked. Nvidia lost $600 billion in a single day. The logic seemed obvious: cheaper AI means less demand for infrastructure. But that logic misses the second-order effect. Cheaper AI doesn't just replace existing workloads at lower cost. It makes entirely new workloads viable. Real-time 4K video translation, autonomous logistics simulation, always-on personal agents, these become commercially feasible only when inference costs drop low enough. And every one of them demands more bandwidth, more compute, and more connectivity. Global data centres consumed approximately 415 terawatt-hours of electricity in 2024, about 1.5% of worldwide demand. The International Energy Agency projects that figure will more than double to 945 TWh by 2030. Despite per-query efficiency gains, Google's overall carbon footprint has surged 48% since 2019, driven by AI infrastructure expansion. Efficiency isn't reducing consumption. It's accelerating it. For AT&T, this is the thesis in a single sentence: the better AI gets, the more bandwidth the world needs.

The historical parallel everyone forgets

This pattern isn't new. During the internet boom of the late 1990s, the most obvious winners seemed to be the dot-coms. But many of the biggest long-term winners were the companies that built the infrastructure underneath. Cisco's routers and switches were the backbone of the early internet. The company's stock peaked during the dot-com bubble, crashed brutally, and took 25 years to recover to those highs. But here's the thing: the infrastructure it built was never wasted. The fiber laid during the telecom boom, including 28 million miles of "dark fiber" mocked as wasteful in 1999, became the foundation for Netflix, YouTube, and the entire Web 2.0 era. As one analysis from KKR put it, the companies building too much fiber weren't wrong about the long-term need for the cables. What they got wrong was how fast it would materialize. Telecom services as a share of total consumption grew before, during, and after the boom, rising from 1.7% in 1995 to 2.4% in 2001. The construction of fiber catalyzed further innovation. Every new application led to demand for more bandwidth, and more bandwidth made more new applications possible. The same flywheel is spinning again, only now the applications are AI-powered.

The AI stack has more layers than you think

Most people think of the AI stack as models and chips. But the full stack is much deeper:

Each layer is a potential bottleneck, and each is a potential investment thesis. NVIDIA gets the headlines because its GPUs are the most visible constraint today. But as models get more efficient and inference moves to the edge, connectivity becomes the binding constraint. You can have the fastest model in the world, but if the network can't deliver its output in real time, it doesn't matter. AT&T is positioning itself at the connectivity layer, arguably the most durable layer in the stack. Chips get replaced every few years. Models become obsolete in months. But fiber in the ground and towers on hilltops last decades.

The skeptic's case

It would be irresponsible to present this without acknowledging the risks. AT&T has a mixed track record with big bets. The DirecTV acquisition in 2015, a $48.5 billion deal, was meant to create a bundled telecom and media powerhouse. Instead, it accelerated subscriber losses, ballooned AT&T's debt from $82 billion to $177 billion (after the subsequent Time Warner deal), and led to a messy partial divestiture. Forbes called it a deal that "hasn't really lived up to expectations," which is generous. The Time Warner acquisition followed a similar arc: regulatory battles, culture clashes, executive departures, and eventual spin-off. AT&T's stock trended downward through both deals. So is $250 billion over five years too much? Possibly. The dot-com parallel cuts both ways. Yes, infrastructure investment eventually pays off. But the companies that built it often didn't survive long enough to benefit. Many of the telecom companies that laid all that fiber went bankrupt. The infrastructure persisted, but the returns accrued to different players. The difference this time is that AT&T isn't building speculative infrastructure for a nascent market. It's expanding proven infrastructure, fiber, 5G, satellite, for a market that already exists and is growing. The $42.5 billion federal BEAD program for broadband deployment adds a government tailwind, and AT&T has already secured the largest share of that funding at about $1.06 billion.

The meta-lesson

When an entire industry zigs in the same direction, there's often more value in zagging. Right now, hundreds of billions of dollars are chasing AI models, AI chips, and AI applications. The infrastructure those models depend on gets far less attention. AT&T isn't trying to build the next GPT. It's building the roads that every GPT needs to reach users. In a gold rush, the people selling pickaxes and shovels don't need to know which hill has the gold. They just need to know that people will keep digging. That's not a guarantee of success. But it might be the closest thing to a sure bet in an industry defined by uncertainty.

References