There is an insatiable demand for AI compute.

AI is not just software. It is a massive physical infrastructure buildout.

Every major AI breakthrough requires more compute. Larger models need more compute to train. Better models need more compute to run. More users create more inference demand. More AI applications create more workloads. And as AI moves from simple chatbots into agents, software development, drug discovery, robotics, autonomous vehicles, defense, industrial automation, video generation, and enterprise workflows, the amount of required compute expands dramatically.

Every model needs chips. Every chip needs memory, networking, cooling, power, land, and data center capacity. Training large models requires enormous compute. Running real-time inference requires even more. And once AI agents, image generation, video generation, robotics, autonomous vehicles, and enterprise automation become part of everyday life, the amount of compute needed will only grow.

AI is not a normal software cycle. In traditional software, once the product is built, the marginal cost of serving another user can be low. AI is different. Every prompt, every generated image, every line of code, every agent task, every robot decision, and every model improvement consumes real compute.

Basically, the more useful AI becomes, the more compute it consumes.

This creates a powerful feedback loop. Better models unlock better products. Better products attract more users. More users create more revenue. More revenue funds more data centers, more chips, more memory, more networking, more cooling, and more power capacity. That infrastructure then enables even stronger models and even more use cases.

There is almost no limit to how much intelligence the world can use.

This is why AI infrastructure is becoming the foundation layer for the next decade of technology and why it remains one of the most important investment themes in the market. The demand is showing up everywhere: semiconductors, cloud providers, data centers, optical networking, memory, power equipment, cooling systems, electrical infrastructure, and manufacturing tools.

But you cannot build data centers overnight. You cannot add gigawatts of power with a software update. You cannot instantly manufacture enough advanced chips, HBM memory, optical components, transformers, cooling equipment, and networking hardware. The physical world moves slowly.

Only a small group of companies can supply the picks and shovels needed to build it.

You can check them in the AI infrastructure portfolio under Data Centers.

But I wanted to highlight a few of them that are standing out.

Nebius $NBIS

Nebius is the clear leader in this space right now. And even more so now that Leopold Aschenbrenner took a position. It’s been discussed at length here over the last year.

Applied Digital $APLD

Applied is offering a good entry spot as highlighted here.

And these are previous crypto miners that have pivoted to HPC (high-performance compute) for AI that are standing out and setting up.

HUT8 $HUT

IREN $IREN

Cipher $CIFR