Artificial intelligence
Compute, models and applications shift value creation into software and infrastructure. The chain runs from AI accelerators and memory chips through networking, data center construction, power and cooling to cloud platforms, applied software and AI at the edge — winners are therefore not only the well-known chip and platform giants but every company along this chain. How much a given company benefits depends on how close it sits to the actual bottleneck: providers with real pricing power over scarce resources like compute, memory or power capacity are positioned differently than pure suppliers that only benefit indirectly from rising demand.
AI accelerators & fabrication
Every AI model is trained and run on specialized chips, making compute the scarcest resource in the entire boom. Manufacturing itself is extremely concentrated: only a handful of foundries and equipment makers master the most advanced process nodes, giving them years of near-guaranteed demand and real pricing power. Investing here means buying into the actual bottleneck of the AI economy, not just an interchangeable supplier relationship.
GPU-/ASIC-Design · Foundry · Advanced Packaging · Lithografie und Halbleiter-Equipment · EDA-Software
Memory & storage
An AI chip is useless without enough fast memory right next to it, which is why HBM and DRAM have become the second big bottleneck after the accelerator itself. Because only a few manufacturers can supply the most demanding memory types, demand keeps outrunning capacity, supporting prices and margins for years. On top of that, training today's models on today's datasets requires massive storage systems just to feed them.
HBM/DRAM · NAND · Enterprise-Storage · Speicher-Controller
Networking & optics
No single chip trains a modern AI model, thousands of GPUs have to work together like one giant computer. That only works with extremely fast interconnects: switch silicon, optical modules and fiber determine how efficiently an entire compute cluster actually gets used. Because every new chip generation needs more bandwidth, this market grows structurally with the AI buildout, regardless of which individual chipmaker ultimately wins.
Rechenzentrums-Switches · Optische Transceiver · Glasfaser und Kabel · Konnektoren
Data center construction & equipment
Every compute cluster needs a physical home: buildings, power distribution, backup power and thousands of server racks before a single chip even runs. This build phase is capital-intensive and takes years, giving the contractors, electricians and equipment makers involved a long, plannable order cycle. Whoever owns the land and buildings rather than just building them also earns rent for decades, much like any other long-leased infrastructure.
Server- und Rack-Hersteller · Elektro-Generalunternehmer · USV und Schaltanlagen · Rechenzentrums-REITs · Baustoffe für Industriebau
Power for data centers
A large data center now uses as much electricity as a small city, and hundreds of them are being built at the same time. This pushes power grids and plants to their limits and forces utilities to add capacity faster than ever, often through long-term supply deals directly with data center operators. Turbine makers, nuclear operators and grid equipment suppliers benefit from a demand wave that has nothing to do with AI chips directly, but would never have happened without AI.
Versorger mit Rechenzentrums-Pipeline · Gasturbinen-Hersteller · Kernkraft-Betreiber · Netzausrüster · Unabhängige Stromerzeuger
Cooling & water
The denser chips get packed, the more waste heat is produced in a small space, and at some point normal air cooling simply stops being enough. Liquid cooling, special refrigerants and large amounts of water go from a side detail to a hard requirement for every modern data center. Suppliers of this cooling technology are not benefiting from a nice-to-have, but from a physical constraint no data center operator can avoid.
Flüssigkühlung · HVAC und Kältetechnik · Kältemittel und Spezialchemie · Wasseraufbereitung
Models, cloud & hyperscalers
Whoever trains the large AI models and sells them through the cloud sits at the source: developers and companies build their own products on top of these platforms and rarely switch provider afterward. That creates high switching costs and recurring revenue, much like classic enterprise software, only growing far faster. That is exactly why these are the best-known names in the whole theme, even though a large share of the actual value creation happens further down the chain.
Cloud-Hyperscaler · Modellanbieter · KI-Plattformen
Applied AI software, agents & data infrastructure
Raw model technology helps nobody until it gets built into actual workflows, and that is exactly where applied AI software comes in. Companies that successfully weave AI into accounting, customer service or software development can displace existing, often decades-old software and earn a lasting place in customers' daily work. At the same time, this is the area with the biggest risk: whoever's own product becomes replaceable by AI ends up on the losing side here, not the winning one.
Vertikale KI-Anwendungen · Datenplattformen und Datenbanken · Entwickler-Werkzeuge · IT-Dienstleister mit KI-Implementierung
AI at the edge
Not every AI request should or can be sent over the internet to a distant data center, whether for cost, latency or privacy reasons. So part of the computation moves directly onto the device: smartphones, PCs and robots get their own specialized chips for it. This triggers a classic replacement cycle where customers specifically swap older devices for AI-capable new ones, a reliable revenue driver for manufacturers and suppliers.
Smartphone- und PC-Prozessoren · Sensorik und Kameras · Robotik-Wahrnehmung · Endgeräte-Hersteller
Materials & inputs
At the very start of every supply chain sit raw materials without which no chip and no data center can be built: copper for cables, rare earths for magnets, ultra-pure silicon for wafers. This demand grows automatically with every new data center, regardless of which individual AI company ultimately wins. Because some of these materials are geographically concentrated or hard to substitute, they create their own bottlenecks that can slow the entire buildout.
Kupfer · Seltene Erden · Spezialgase und Reinstchemie · Reinstsilizium und Wafer · Leiterplatten und Substrate