Investment in artificial intelligence is shifting from a race focused solely on purchasing GPUs. Dell’Oro Group estimates that the global market for semiconductors and IT components for data centers will exceed $1.8 trillion by 2030, a forecast revised upward due to the expected growth in CAPEX and electrical capacity. Accelerators, memory, storage, CPUs, and network cards will participate in an expansion that is beginning to transform the entire data center infrastructure.
The key points of the data center chip market in 30 seconds
- Dell’Oro raises the total projected market for semiconductors and IT components to over $1.8 trillion by 2030.
- AI accelerators will account for most of the expenditure, but memory, storage, CPUs, and NICs will also grow.
- Servers and storage could require more than 200 GW of power over the next five years.
- Inference and AI agents will also boost demand for conventional servers and networks.
This forecast helps explain why AI expansion is simultaneously impacting industries that were previously analyzed separately. The challenge is no longer just building enough accelerators. Each GPU needs memory, each server requires storage and connectivity, and every new rack demands electricity, cooling, and a network capable of handling enormous data flows.
The International Energy Agency (IEA) observes the same phenomenon from the electrical perspective. Data center consumption increased by 17% during 2025, and the organization expects it to roughly double by 2030. Demand from AI-oriented data centers will grow even faster.
The $1.8 trillion is not just GPU spending
The most striking data point in the Dell’Oro report is the projected market size, but perhaps more interesting is how that money will be allocated.
AI accelerators will remain the dominant category, according to the consultancy. This includes GPUs and other specialized processors used for training and running models.
But an accelerator doesn’t work alone.
It needs high-bandwidth memory, CPUs managing specific loads, SSD and HDD units for data sets and models, and network adapters capable of connecting servers that may contain multiple GPUs each.
Dell’Oro anticipates simultaneous growth in:
- AI accelerators;
- CPUs;
- memory;
- storage units;
- network interface cards (NICs).
This evolution is especially significant in memory and storage. Dell’Oro forecasts that supply constraints will keep prices high in the short term, but expects gradual moderation over the forecast period.
This pressure is already visible outside data centers. Demand for AI memory competes for industrial capacity and capital with other market segments, as manufacturers try to ramp up production of advanced memories.
The data center thus becomes a system whose growth depends on many supply chains moving roughly at the same pace.
Inference and agents are changing data center composition
During the initial phase of the current AI race, much attention was focused on training.
It was easy to associate AI infrastructure with massive GPU clusters used for training ever-larger models.
Inference is partially changing that picture.
Once a model is trained, each user request, application, or agent requires computations to generate a response. The more these systems are used, the greater the infrastructure needed to handle these requests.
AI agents can further amplify this activity.
A chatbot typically responds to a request. An agent may chain numerous steps, query databases, execute code, use external tools, and repeatedly invoke other models before completing a task.
The IEA has already warned that, although energy consumption per AI task is decreasing rapidly thanks to efficiency improvements, the number of users is growing, and heavy uses—including agents—are increasing.
Dell’Oro expects this evolution to benefit also general-purpose servers.
Not all of the infrastructure needed for an agent must run on a GPU. Databases, storage, application servers, monitoring systems, and numerous auxiliary services continue to use conventional CPUs.
The expansion of AI might end up increasing CPU demand, not decreasing it.
NICs are no longer secondary components
The network is another major beneficiary.
Dell’Oro expects network interface cards to accelerate in growth for two different reasons.
On one hand, front-end: conventional servers used for inference, storage, and agent-based applications need connectivity.
On the other hand is the back-end of large AI clusters.
Training or running certain distributed models requires connecting a large number of accelerators. In these systems, the network ceases to be just a way to enter and exit data and becomes part of the compute system itself.
An extremely fast GPU may wait idly if data doesn’t arrive quickly enough.
That’s why manufacturers and operators are increasing Ethernet and InfiniBand speeds and developing architectures aimed at reducing latency and improving communication between accelerators.
The AI race is also a race for networks.
More silicon designed by hyperscalers themselves
Dell’Oro also anticipates another trend with significant market implications: accelerators, CPUs, and NICs will increasingly use custom chips.
Major cloud providers have been working in this direction for years.
Google developed their TPU for machine learning workloads. AWS has Trainium for training and Inferentia for inference, along with their Graviton CPUs. Microsoft has developed Maia accelerators and Cobalt processors. Meta is working on its own MTIA accelerators.
The economic rationale is becoming more compelling.
When a company installs millions of chips, reducing a few watts per device can translate into significant savings in electricity, cooling, and infrastructure.
The same goes for the cost per inference.
Dell’Oro points out that technological development is shifting from encapsulation to entire rack-level designs. The goal is no longer just to create the fastest chip, but to increase performance per watt and lower cost per token.
This favors heterogeneous architectures.
A company can deploy different accelerators depending on whether they need training, inference, video processing, or specific internal loads, instead of trying to run everything on the same type of GPU.
200 GW reveals the other side of the problem
There is a figure from the report that helps illustrate the physical scale of these forecasts.
Dell’Oro estimates that components for servers and storage deployed over the next five years could consume more than 200 GW of power.
This doesn’t mean all that consumption will happen all at once overnight, nor that it is solely for AI. It’s an accumulated capacity forecast related to the equipment the consultancy considers.
But it forces us to look beyond semiconductors.
The IEA estimates that data centers consumed around 415 TWh of electricity in 2024, roughly 1.5% of global electricity. Its baseline scenario projects consumption reaching about 945 TWh in 2030, close to 3% of worldwide demand.
The problem is also unevenly distributed.
Data centers are geographically concentrated. Therefore, a relatively small share of the global electricity consumption can translate into significant regional network challenges.
The IEA estimates that around 20% of planned data center projects could face delays if certain grid integration issues aren’t addressed.
This is one of the real limits to Dell’Oro’s projected $1.8 trillion growth.
More GPUs can be manufactured.
But they also need to be connected somewhere.
AI infrastructure begins before the server
The race to build data centers is extending bottlenecks to components that traditionally received less attention.
Transformers.
Substations.
Generators.
Uninterruptible power systems.
Liquid cooling.
Switches.
Fiber optics.
Memory.
Storage.
In April, the IEA noted that supply chains for gas turbines, transformers, advanced chips, and IT components had tightened over the past year, while new projects increased pressures on planning and electrical connection procedures.
This also explains why Dell’Oro’s forecast has changed since January.
The firm states that its new estimate significantly raises previous projections due to higher forecasts of CAPEX for data centers and new electrical capacity.
The market is beginning to price in the fact that the necessary infrastructure will be greater than what was expected just a few months ago.
From buying GPUs to designing complete data centers
The $1.8 trillion figure reflects a deeper shift.
In the early years of the generative AI boom, Nvidia and its GPUs dominated discussions on infrastructure. They still hold a central position, but the next phase demands much more.
Inference requires servers.
Agents need compute and storage.
Clusters need networks.
Accelerators need memory.
Racks need cooling.
And all of this requires electricity.
The IEA forecasts that worldwide electricity demand from data centers will continue to grow much faster than overall consumption. In the US alone, these facilities are expected to account for around half of the increase in electricity demand through 2030.
Therefore, the $1.8 trillion projected by Dell’Oro does not simply represent another semiconductor cycle.
It describes the construction of a new layer of global IT infrastructure.
The GPU remains one of its most costly components. But it’s increasingly clear that having an accelerator isn’t very useful if you lack memory, storage, network, cooling, or megawatts to power it.
Frequently Asked Questions
How much will the data center component market grow?
Dell’Oro Group estimates that the global market for semiconductors and IT components for data centers will surpass $1.8 trillion by 2030.
Which components will grow due to AI expansion?
In addition to accelerators, Dell’Oro anticipates increased demand for CPUs, memory, storage, and network interface cards. AI inference and agents may also boost the utilization of conventional servers.
Why do data centers need faster networks?
Large AI systems distribute work across numerous accelerators. This requires moving data with high bandwidth and low latency between servers, making the network a critical part of the overall cluster performance.
Can electricity limit AI growth?
Yes. The IEA warns of capacity and connection issues in certain networks and estimates that around 20% of planned data center projects could experience delays if these issues aren’t addressed.
via: delloro
