One of the biggest questions surrounding the AI boom has been whether customer demand would eventually justify the vast investment in data centers, GPUs and cloud infrastructure. New research suggests that AI revenue is now beginning to cover the cost of the infrastructure that powers it.
According to a report from research firm Exponential View, global AI revenue outside China reached $25 billion during the first quarter of 2026. That figure exceeded the industry’s estimated $21 billion in depreciation costs tied to AI infrastructure for the second consecutive quarter.
The finding matters because depreciation represents the ongoing cost of the enormous investments that tech companies have made in AI data centers, networking equipment and GPUs. If revenue cannot cover those costs, the long-term economics of the AI market become difficult to sustain.
Tech giants including Amazon, Microsoft, Alphabet and Meta are collectively expected to spend as much as $725 billion on capital expenditures this year, much of it directed toward AI infrastructure. The spending spree ranks among the largest waves of corporate tech investment in history.
Yet the economics remain challenging. While AI revenue has moved above depreciation costs, those charges still consume more than two-thirds of industry revenue. That leaves relatively little room to absorb other major expenses such as electricity, data center operations, labor, financing costs and cloud infrastructure management.
Analysts have frequently pointed to the gap between infrastructure spending and end-user demand as one of the sector’s biggest risks. LLM developers such as OpenAI and Anthropic continue to spend heavily on compute resources, while cloud providers rapidly expand capacity in anticipation of future demand.
Data From More Than 1,000 Companies
Exponential View’s analysis draws on spending and revenue data from more than 1,000 companies across the AI ecosystem. Researchers combined public filings, executive disclosures, cloud-provider information and industry reporting while adjusting figures to avoid counting the same revenue multiple times across the supply chain.
A key finding concerns the durability of AI hardware investments. Some critics have argued that GPUs depreciate faster than conventional IT equipment because new chip generations arrive rapidly, potentially rendering older hardware obsolete. The report assumes a six-year depreciation cycle for AI infrastructure, including GPUs.
However, market data suggests older chips continue to hold value. Rental pricing for NVIDIA’s H100 accelerator remains close to 80% of its original launch-level pricing, indicating sustained demand even as newer Blackwell-based systems enter the market. AWS has similarly reported ongoing customer demand for older NVIDIA A100 servers.
The report also highlights a shift in how developers consume AI models. Usage data shows growing adoption of open-weight models and Chinese-developed alternatives such as DeepSeek. On multi-model development platform OpenRouter, the combined share of requests going to models from OpenAI, Google and Anthropic fell substantially over the past year.
That trend may not threaten the largest AI model providers, but it could increase pricing pressure. Premium vendors may need to differentiate through enterprise services, workflow integration, ecosystem lock-in and other value-added offerings rather than relying solely on model performance.
However, if demand growth slows, pricing declines sharply, or hardware assets depreciate faster than expected, the economics could quickly become more challenging.
