The figures related to the AI data center construction have long ceased to be typical investment-cycle figures. Jensen Huang predicts that $3 to $4 trillion will be allocated to AI infrastructure by 2030. In fact, McKinsey's wider perspective puts the number near $6.7 trillion by 2030. If you look at just the four major US hyperscalers, they together invested $305 billion in capital expenditures in 2025, and their 2026 budgets are pointing notably higher.
This type of large spending isn't directed to one stock. It is going through several levels - from the chips that train the models, to the power systems supplying the racks, to the cooling loops removing the heat, to the neoclouds renting GPUs by the hour, to the utilities and real estate operators who are working to deliver gigawatt scale site. Different parts of that stack have very different risk-reward profiles and are differently exposed to the durability of the buildout itself.
Here is what investable layers of the AI infrastructure boom actually look like in 2026, and where the real winners are appearing in the public markets.
The Chip Layer Is Still Where the Biggest Dollars Land
Approximately 50% of hyperscaler data center spending is on silicon, and that's also the primary area where wealth creation has been most deeply concentrated. Nvidia still holds the main position. The company's Blackwell generation is producing in volume, the Rubin and Vera Rubin platforms are being introduced until 2026, and its partnership with OpenAI alone is the big commitment of at least 10 gigawatts of Nvidia-based systems with the chipmaker progressively investing up to $100 billion into OpenAI as each gigawatt goes live. The first gigawatt is scheduled for the second half of 2026.
Broadcom is the comparatively less advertised but structurally very important second pillar of the chip layer. This company provides the networking silicon which connects the tens of thousands of GPUs inside a training cluster, and it has become one of the most favored partners of hyperscalers designing custom AI accelerators. Its AI revenues have been on an upward surprise for several quarters, and analysts often mention it together with Nvidia as a chip-layer company with a solid moat.
The Neoclouds and Hyperscale Operators
The operators of capacity--that is, those companies who own or lease the data centers where the GPUs are actually being run--are the next layer of investment. CoreWeave, now widely recognized as the leading pure-play in this space, is one example. This firm leases Nvidia GPU power to AI research labs and enterprises. In fact, Nvidia put in another $2 billion in equity in January 2026, speeding CoreWeave's plan for 5 gigawatts of AI factories by 2030. Nebius is the European counterpart with its own facilities and a growing customer base.
Among the hyperscale players, the companies that stand to gain most obviously are Microsoft Amazon Alphabet, Meta, and Oracle. Each of them has revealed plans for multi-tens-of-billions in capex for 2026 and each is playing its own game. The Azure AI segment, helping to drive Microsoft's growth, is now one of the fastest-growing parts of the company. AWS is still building-out its GPU clusters with both Nvidia silicon and its own chips. Google Cloud is focusing on TPU as a differentiator. Meta is working on its own AI training infrastructure, including the 5-gigawatt Hyperion campus in Louisiana and the Prometheus site in Ohio. Oracle Cloud Infrastructure, which in the past has been regarded as a niche player, has transformed into a serious AI hosting contender. It recently secured contracts that, in our wildest dreams, would have seemed impossible three years ago.
How to Think About Positioning
The real issue for most investors in practical terms isn't whether the expansion will really happen -it's how to balance exposure to different layers so as not to miss the good times or be overly affected when one of the layers undergoes a correction. The chip layer has delivered the most returns but it also has the highest valuations and is the layer most at risk if there is a decrease in training capex. The power and cooling layer has smaller multiples, a clearer line-of-sight to revenue, and it benefits from the buildout no matter which chips become dominant. Neoclouds give concentrated upside with concentrated customer risk. Hyperscalers provide a more diversified exposure but at large-cap valuations. It is the utilities and real estate layer where mispricing is most likely, as historically fewer equity investors have contemplated those sectors in terms of AI.
Keeping current on which of these layers is actually delivering on its narrative matters more than it used to. A good AI industry resource covering hyperscaler capex, data center announcements, chip roadmaps, and the less-visible power and cooling supply chain is genuinely useful for an investor trying to separate durable trends from one-quarter noise. The pace of announcements in this space has been fast enough that quarterly analyst reports often lag by months, and primary coverage of the infrastructure news itself tends to produce better calls.
Closing Thought
Building AI data centers is the biggest infrastructure spend cycle in recent history by most reasonable measures. It's also so early that the list of companies who will profit from it keeps changing. The firms that appeared to have an edge in 2024 - mainly Nvidia and the top hyperscalers - are now seen alongside a much larger group of power cooling networking, real-estate, and utility companies that should be equally or even more involved in the buildout.
Expanding that list is likely good for the overall theme's long-term returns. Over-reliance on one stock or one layer is always though. Here the investors can not only take an AI infrastructure exposure at several points of the stack but also with different risk profiles. The trade that the market considers the safest is the picks-and-shovels, because it makes money even if the top of the stack winners turn out to be Nvidia and Microsoft or a very different set of names in five years.