Companies building the most capable AI models in 2026 are running out of electricity, cooling capacity, the chips and the cables that connect them. According to Bridgewater, giants like Alphabet, Amazon, Meta AND Microsoft are located in invest together $650 billion in AI-related capital spending in 2026 (up from $410 billion in 2025). These are commitments at a scale that only makes sense if you believe that the physical layer is where competitive advantage is actually built.
Most founders and investors are still optimizing for the layer above. They are building products that depend entirely on infrastructure that they don’t own, control, and can’t take for granted. This is a structural weakness.
So where does the value go from here? Who ends up controlling what the AI economy actually runs on? The answers are starting to come together.
Could infrastructure in 2026 be a bubble?
Whenever capital moves rapidly in a certain direction, one calls it a bubble. We’ve seen it before with dot-coms, with crypto, and, more recently, with a wave of AI startups that were little more than a thin layer of code sitting on top of ChatGPT.
Today, the word “infrastructure” covers the GPU clusters, fiber networks, and power grids that power modern data centers. This is a big step from what it once meant – roads, bridges and pipelines. When a word begins to encompass many different assets at the same time, the question of the bubble becomes inevitable.
A bubble, at its core, is demand that doesn’t yet exist. Capital is poured into something that may or may not materialize. With infrastructure, however, the reality is the opposite. You can’t train a modern AI model without large amounts of electricity, store sensitive data securely without a specific network, and perform more calculations without chips. These are hard physical limitations that define what is actually possible today. From Texas to Northern Virginia, data center expansion is already running afoul of local power constraints and grid capacity debates. And even the supply of advanced GPUs – dominated by companies like Nvidia— has become a strategic constraint.
Of course, big capital is responding to this. PitchBook data shows Increase in infrastructure investments by 44 percent year after year. What’s more, it’s a completely different type of investment compared to what fueled the AI boom. The infrastructure is constrained by strong physical constraints and demand cannot be produced arbitrarily.
The drivers behind the infrastructure boom
If capital is pouring into infrastructure on this scale, there must be powerful forces driving it. Geopolitics is one of the most important.
Governments around the world are retreating from centralized clouds they don’t control. If your most sensitive data resides on servers owned by a foreign corporation subject to foreign law, true sovereignty becomes questionable. This realization has helped accelerate what is now called “sovereign AI”—the idea that a country’s AI capabilities must run on infrastructure located within its borders. Building this infrastructure requires huge investments, so geopolitics has become a major driver of the current boom.
Then comes the cost of capital. When interest rates rise and uncertainty increases, investors become more selective and favor tangible assets. The infrastructure fits this profile. It is tied to physical capacity and contracted, active demand that exists regardless of market sentiment. Compared to supporting a startup that may pivot several times before finding its footing, the infrastructure can seem remarkably stable.
Energy completes the picture. Training a modern AI model can require approximately ten times more power than traditional computing workloads, and the appetite continues to grow. Network access has quietly become one of the most contested advantages in technology.
This dynamic is already visible from the large financial movements. BlackRockfor example, there is launched a $100 billion fund specifically dedicated to AI energy infrastructure. The logic is straightforward: whoever controls the electricity supply ultimately affects how quickly the AI economy can expand.
Why applications timed out
Infrastructure didn’t suddenly become attractive overnight. During the app era, startups built an app, acquired users, scaled quickly, and dealt with margins later. The infrastructure was already in place, too. Cloud computing was relatively cheap, smartphones were ubiquitous, and capital was patient enough to wait for the benefit. In other words, these conditions were almost perfect for application layer bets.
But markets eventually saturate. Suddenly, a new food delivery app couldn’t solve new problems, it could only fight for scraps in a crowded space where margins were shrinking and new users were expensive to attract.
At the same time, AI and Web3 introduced a level of computational and architectural complexity that older centralized architectures simply weren’t built to handle. After all, you can’t run a frontier model on infrastructure designed for a travel application. This was the real turning point, when infrastructure became the most attractive bet.
The next decade belongs to the builders below
In the age of apps, winners controlled distribution. Google owned search, Apple and Amazon owned phones dominated the online storefront. Everyone else built up and paid for access. Now the keepers are moving down a layer. Compute, power and connectivity are becoming the critical bottlenecks because every AI product ultimately depends on them.
This change changes the look of a winning company. In an infrastructure cycle, success is defined by the ability to deliver real capacity: more compute, lower costs, and better control over data. Buyers ultimately pay for reliability and operational control.
For investors, the picture favors capacity. Infrastructure and networks that are expensive to replicate, difficult to replace, and already stretched by real demand tend to stay.
The earliest opportunities are emerging in two areas. One is Decentralized Physical Infrastructure Networks (DePIN), where computation and connectivity are distributed beyond the larger cloud providers. The other involves hybrid operators, teams that own physical equipment while also running the software stack needed to keep that equipment productive.
These are not quick businesses to build, and that’s exactly the point. Institutional capital is often more comfortable investing in sectors with long time horizons, operational complexity and scarce physical assets. In the AI economy, value is likely to be concentrated further down the stack, where every application ultimately depends on compute, power and reliable connectivity.
