AI Infrastructure Spending to Top $1T by 2030

Two market research reports have come to the same conclusion: the AI infrastructure market. He will soon surpass the $1 trillion mark. The only difference between the two is how far the market will go.

One report by Dell’Oro Group says the worldwide data center IT semiconductors and components market is poised to reach $1.8 trillion by 2030. This forecast has been revised significantly since the January 2026 report, reflecting higher projections for data center capex and power capacity additions.

“AI will remain the primary driver of data center infrastructure spending, with accelerators accounting for the largest share of the component market,” said Baron Fung, vice president of research at Dell’Oro Group in a statement. “Memory, storage, CPUs, and NICs will innovate and scale alongside accelerators, while supply constraints are expected to sustain elevated pricing for memory and storage.”

At the same time, inference, storage, and agentic AI are expected to increase demand for general-purpose servers, supporting growth across the broader component categories.

Significant development is also underway to improve accelerator performance per watt and reduce cost per token by scaling architectures from the package through the rack level. “Heterogeneous computing targeted at various workload domains could further improve efficiency at the data center scale,” according to Fung.

Other highlights from the report include:

• Server and storage components are projected to consume more than 200 GW of power over the next five years, requiring data center power and supporting infrastructure to expand alongside component demand.
• Component prices, particularly for memory and storage, are expected to moderate over the forecast period, helping support more sustainable long-term market growth.
• AI accelerators, CPUs, and NICs will increasingly shift toward custom silicon as hyperscalers seek greater efficiency and lower costs.
• NIC growth will accelerate as front-end demand increases from general-purpose servers supporting inference, agentic AI, and storage, while scale-out AI networking drives stronger back-end demand.

Meanwhile, Futurum’s new five-year forecast finds the global data center semiconductor market reaching $1.2 trillion by 2030 as agentic AI inference reshapes silicon demand. By way of contrast, the global data center semiconductor market reached $241.0 billion in calendar year 2025, up 3.6x from $66.6 billion in calendar year 2023.

Futurum says growth is front-loaded — as a percentage of growth rate it will more than double in CY2026 and peak at 54.0% in CY2027 as committed capacity lands. Future years will have greater dollar figures but slower CAGR.

The forecast’s organizing thesis is that agentic inference, not training, now sets semiconductor roadmaps, which is a common thesis. Inference-focused accelerators will grow from roughly half the CY2025 market to 73% by CY2030 ($884.9 billion), and Agent and Reasoning-first inference alone reaches $546.0 billion, or 45% of all data center semiconductor revenue.

“2026 is the first year of the inference transition that defines the rest of the decade,” said Brendan Burke, research director, semiconductors, supply chain & emerging tech at Futurum in a statement. “High-concurrency agentic workloads will set the architecture for vendor roadmaps through 2027.”

The research found GPUs will hold half the market at $604.5 billion by 2030 (30.8% CAGR), but custom XPUs will grow the fastest at a 44.6% CAGR to $237.2 billion and run roughly 75% cheaper than equivalent GPU deployments. That’s why GPUs account for such a large dollar share: they are very expensive, rumored to run $40,000 each.

Off-chip memory is the single fastest-growing line at a 72.4% CAGR to $260.5 billion, while HBM grows 5.9x to $197.6 billion, making the AI accelerator market almost as big as the DRAM market. With its 30% price premium, it’s easy to see why memory makers have prioritized the market.

“At a 75% cost advantage, every inference workload that can run on an XPU eventually will, and the binding constraint is advanced packaging throughput,” said Burke. “On memory, the supercycle is real, but buyers should build flexibility into every commitment before the cost structure resets in 2028.”