In 2018, Crusoe started as a cryptocurrency infrastructure company. Eight years later, it looks very different.
In September 2026, the company raised $3.9 billion at a $30.9 billion valuation. Its focus now sits squarely on AI infrastructure, including data centers, compute capacity and energy systems.
Crusoe’s evolution is a useful way to understand where part of the AI market is heading.
Early AI investment centered on foundation models and applications. As those products scale, more attention is moving to the infrastructure beneath them: compute, networking, storage, cooling and power.
The investment thesis is becoming less about finding the next AI application and more about understanding what could stop AI from scaling efficiently.
AI workloads place very different demands on data centers than traditional cloud software.
U.S. data centers consumed about 176 TWh of electricity in 2023. The U.S. Department of Energy estimates that figure could reach between 325 and 580 TWh by 2028.
Power demand is only one part of the problem. AI servers also require denser compute, faster networking, more sophisticated cooling and better access to large datasets.
The important question for investors is where those constraints become severe enough to support a new market.
Several areas already stand out.
GPUs remain the backbone of most AI infrastructure, but the market around them is expanding.
Specialized chips are being designed for training, inference and narrower workloads. At the same time, memory, packaging and interconnects are becoming more important as compute clusters grow.
Inference deserves particular attention.
Training requires large amounts of compute upfront. Inference happens every time a model responds to a user, processes an image or runs an agentic workflow.
As AI products move into production, inference becomes a recurring operating cost.
That puts more value on systems that can lower latency, energy use and cost per query, even when they do not deliver the highest raw compute performance.
Large AI systems need constant access to enormous amounts of data.
VAST Data, for example, raised roughly $1 billion in 2026 at a reported $30 billion valuation. Its software is designed to help organizations store and process the datasets required by modern AI systems.
Networking faces a similar challenge.
Large AI clusters can involve thousands of processors communicating at extremely high speeds. Performance depends not only on the processors themselves, but also on how quickly data can move between them.
That brings switches, optical connections, interconnects and network management software further into the AI investment landscape.
There is also a software opportunity here. As companies spend more on physical infrastructure, tools that improve utilization, scheduling and performance become more valuable.
Higher compute density has turned energy and heat into central design constraints.
The IEA estimates that the power density of AI servers increased elevenfold between 2020 and 2025 and could increase another fourfold by 2027.
Traditional cooling systems were not designed for this environment.
Companies such as ZutaCore are building direct to chip liquid cooling systems for higher density workloads. ZutaCore raised $100 million in 2026 from investors including Mitsubishi Electric, Carrier Ventures and Samsung Ventures.
Power availability can be even more restrictive.
Building a data center means securing enough electricity, grid capacity and, in some cases, energy storage. In markets where those resources are limited, power can determine whether a project moves forward at all.
That makes energy infrastructure relevant to AI investors in a way that would have seemed unusual a few years ago.
Not every AI workload needs to run in a hyperscale data center.
Robotics, factories, vehicles, cameras and consumer devices often benefit from running models closer to where the data is generated.
Local inference can reduce latency, lower data transfer requirements and allow systems to operate without constant cloud connectivity.
The investment case depends on the use case. Edge infrastructure makes the most sense where local processing materially improves cost, speed or reliability.
That puts specialized processors, embedded systems and edge orchestration software in a distinct part of the infrastructure market.
AI infrastructure does not look like traditional SaaS.
Hardware companies often need more capital and take longer to reach commercial scale. Manufacturing, supply chains, technical validation and customer deployment matter much more.
Customer concentration can also be high, particularly when hyperscalers and large AI labs account for a meaningful share of demand.
The upside is that technical barriers can be much harder to replicate.
A new SaaS product may be copied in months. A semiconductor architecture, cooling system or networking platform can require years of engineering, manufacturing relationships and infrastructure integration.
For VCs, the areas worth tracking include:
Compute and inference efficiency
Memory and storage
AI networking
Power infrastructure
Cooling systems
Edge inference
Semiconductor design
Critical supply chains
Infrastructure management software
The useful question is not which category looks most exciting.
It is where AI growth is creating a constraint that customers urgently need to solve.
AI software depends on physical systems.
More usage requires more compute. More compute raises power and cooling requirements. Larger clusters need faster networks. More inference increases pressure on cost and energy efficiency.
Those dependencies are becoming easier to see as AI moves from experimentation into large scale deployment.
Crusoe is one example of what can happen when a company positions itself around one of those constraints.
The next important AI company may still sit at the application layer. But a growing share of the venture opportunity now sits deeper in the stack.
