The infrastructure of AI

TL;DR:

• Many people talk about AI models, but there are many infrastructure layers below it
• If one wants to build scalable AI systems, mastering its infrastructure is a prerequisite
• The infrastructure is what makes AI so interesting for the whole economy, not just software companies
• Chances and risks of AI span from AI labs to energy companies providing gas turbines for AI data centers
• We might be in a financial bubble, but the long term benefits of AI will be great for society

Building out AI is more than building great models

There is a reason why AI has been able to push major capital markets indices to all-time highs over the last couple of years. Its economic impact can be divided into two areas: the buildout and the payoff. Anticipating major opportunities for selling AI solutions in the future, companies have been investing unimaginable amounts of capital into the buildout. While frontier models mostly come from AI labs, those labs are mostly financed by MAG7 companies. After using their cashflows in the beginning, companies began financing this buildout by taking on debt. Their investments cause ripple effects that matter to the whole economy.

AI

Data

Platform

Silicon

Facilities

Energy

AI

At the AI level, you have frontier labs training the latest AI models. Those models are trained using data on the scale of the entire internet. The most famous AI labs are OpenAI and Anthropic. To build their AI models, they need three things: data, compute and the best AI researchers available. The AI labs are mostly financed by companies from the data, platform and silicon layer. This enables them to purchase products and services from exactly those companies. As the data, platform and silicon companies typically trade at a very high multiple of their revenue, they use this as a means to push their own capital markets valuations. This gives them better access to capital, decreasing their capital costs. Also, AI models have to be deployed somewhere. As MAG7 companies are the only ones that can serve the global demands for AI, they know that AI models will eventually be deployed in their clouds, whose services they can sell at high margins.

Data

Being relatively generic after training, companies need to provide AI models with context to create actual business value when using them, making data the first layer below the AI models. Extensive knowledge databases need to be created, giving companies the opportunity to retrieve context relevant for a request. Especially for older companies this is challenging, as they oftentimes have fractured systems and most of the company’s knowledge is in the heads of their employees. In enterprise systems, companies like Databricks and Snowflake provide platforms for working with vast amounts of data.

Platform

AI models deployed at scale need to be managed by an underlying software infrastructure that routes user requests to different models and monitors for errors. These platforms ensure AI does not crash under heavy loads while meeting the highest security standards. They can be built using orchestration software like Kubernetes, which is able to balance loads efficiently and scale model deployments depending on demand. Winners in this space are the cloud hyperscalers (Amazon, Microsoft, and Google), which provide these infrastructure platforms on a highly scalable pay-per-use basis. They dynamically allocate cloud resources where they are needed most by using a layer of abstraction over classical servers while ensuring customer privacy. Many companies choose them because this way, they do not have to build data centers themselves. Additionally, the pay-per-use model is highly attractive as it gives companies the option to start and stop cloud services as they like in no time. If a service is no longer needed, hyperscalers use the underlying infrastructure for other cloud resources paid for by other customers.

Silicon

AI models basically work on matrix multiplication. They rely on highly specialized microchips that enable parallel processing. Access to these highly demanded and very expensive chips determines who can train and run the most advanced AI models. Equipped with huge cash flows, hyperscalers buy chips from companies like Nvidia or AMD. As Nvidia only designs chips, they are usually built by TSMC, a Taiwanese company that heavily relies on lithographic machines from ASML, a company from the Netherlands. At the heart of ASML’s lithography machines lie laser systems manufactured by Trumpf and highly specialized mirrors manufactured by Zeiss. Both are private engineering companies from Germany!

Facilities

The AI supercomputers used by the big cloud providers need to be housed somewhere. They require huge data centers equipped with advanced cooling systems. Therefore, massive, multi-year construction investments are needed. In the end, data centers are just massive buildings that host hardware and make sure it runs safely. Those buildings need to be built. Equinix and Digital Realty are the major data center developers. They buy the land, secure the power permits and design the data center architecture. Additionally, investment management firms like Blue Owl Capital and Pimco provide huge amounts of capital for building data centers.

Energy

Meta, part of MAG7, currently plans to build Hyperion, a data center for which they bought an area of roughly 14.8 km². The area will host not only data center buildings, but also cooling systems, power infrastructure, security buffer zones and more. Hyperion will draw up to 5 GW, the equivalent of roughly 3 million homes in that region. The energy required for data centers of this scale is so enormous that companies like Meta invest in the energy sector to ensure long-term power availability. In January 2026, Meta signed a deal to secure over 6.6 GW of nuclear power for their future data centers. The Hyperion facility will host three natural gas turbines and Meta has committed to adding 1.5 GW of new solar and wind energy in the region. Winners in this space are Siemens Energy and GE Vernova who provide gas turbines and grid infrastructure.

The MAG7: The Magnificent Seven

The Magnificent 7 (MAG7) are the dominant technology companies driving AI investment and infrastructure buildout.

These seven companies control the majority of AI development, compute resources, and infrastructure investment globally.

Let’s see where this takes us

The huge investments in the AI space represent both opportunity and risk for the economy. As they usually lead to reliable profits, data centers have been major success drivers for MAG7 companies. Whether this continues at such a scale remains to be seen. At each step are companies seeking to make profits, which inevitably increases prices for companies later in the value chain. As a result, more and more companies are beginning to vertically integrate their supply chains to reduce long-term costs. This means that they try to build parts of the supply chain on their own. For example, Google already designs specialized AI chips themselves to be more independent. Google and Microsoft also build their own AI models to be less dependent on OpenAI and Anthropic.

It might very well be the case that we currently see an overshooting in the build-out and demand will not follow right away. This could lead to a scenario where some companies may not be able to pay for their investments. At the scale we are currently at, this could lead to a ripple effect that disrupts the whole economy. In this case, companies with lots of liquidity would be the kings as they could then buy in at lowered costs. Possibly, companies would be bailed out by governments or die. History shows that while financial crises occur relatively predictably every few years, their exact occurrence can never be forecasted. They depend on information that does not yet exist. An economic downturn of this type would typically be triggered by a single event, oftentimes called a Black Swan. From a future perspective, a crisis because of AI companies currently overplaying their cards could look sensible in hindsight. Nevertheless, we never know for sure when the next crisis will be and what will be its cause. I can only assure you that there will be crises in the future.

While we might currently be in a financial bubble, we are definitely not in a technological bubble. AI is here to stay, and an economic downturn will only strengthen the capital markets in the long term by weeding out companies unable to weather the storm. Even after a financial crisis, AI will still be the major driver for productivity, efficiency and prosperity by enabling novel research methods, personalized education, advanced healthcare and much more.