The AI wave is undergoing a critical inflection point, where rapid leaps in model capabilities are accelerating last year's enterprise pilot projects into full production deployment. The market, however, continues to systematically underestimate the depth and speed of this transformation.
According to analysis, a Citi research team led by analyst Heath Terry stated in a recent report that enterprise-grade applications are transitioning en masse from last year's pilot phase to production deployment. The pace of improvement in model capabilities is faster than ever before, causing the entire industry's demand curve to rise sharply.
Citi has raised its total AI industry revenue forecast for 2026-2030 from $2.8 trillion to $3.3 trillion. Over the same period, its capital expenditure forecast has been increased from $8.0 trillion to $8.9 trillion. Their assessment is that the market remains focused on risks such as data center construction challenges, funding pressures, and intensifying competition, while overlooking the high returns these investments are generating and the emergence of an enterprise-driven productivity cycle.
For the software industry, this is a more precarious moment than most realize. As AI-native companies experience steep revenue growth, the high switching costs, strong pricing power, and high barriers to entry that traditional software vendors have long relied upon are being repriced by AI technology. This repricing is already evident in stock valuations—over the past year, software stock valuations have clearly diverged from those of AI infrastructure-related assets—but Citi believes consensus earnings forecasts are far from reflecting the ultimate impact.
Within the infrastructure layer, particularly in memory, storage, CPUs, and power, is where Citi currently sees the most favorable risk-reward profile. The recent underperformance of hyperscale cloud providers is viewed as another window of opportunity.
Model capabilities are improving at a steeper rate. Three leading models—GPT-5.4, Gemini 3.1 Pro, and Claude Sonnet 4.6—were released within less than three weeks, with capability leaps far exceeding any previous cycle. Measured by the ARC-AGI-2 benchmark, Gemini 3.1 Pro's score is 1.5 times higher than its predecessor from three months ago. GPT-5.3-Codex is OpenAI's first model to have participated in generating its own code—a significant milestone.
More notably, as model capabilities improve, token pricing is also rising. Inference models utilizing techniques like Mixture of Experts (MoE) and verifiable reward reinforcement learning (RLVR) consume more tokens per response. Although Gemini 3.1 Pro's pricing remains consistent with the previous generation, its intelligence score has doubled.
Citi argues that the combination of these two trends implies a structural upside for AI service providers' revenue per unit. Capability improvements are already influencing specific corporate decisions. Block's recent layoff announcement explicitly cited AI factors, an early signal that technological diffusion is extending from the development layer to the operational layer.
The shift from enterprise pilot projects to production deployment is happening faster than anticipated. System integrators are key drivers of this acceleration. Leading consulting firms are simultaneously overhauling their own internal operations and helping traditional enterprises rapidly deploy solutions from companies like Anthropic and OpenAI, acting as the "capillaries" for AI diffusion. Field research by Citi involving CIOs, CTOs, and system integrators indicates that the core driver for acceleration is competitive pressure—no company wants to let a competitor gain an edge.
Data supports this: the combined backlog growth rate for AWS, GCP, Azure, and CoreWeave reached 100% in the fourth quarter of 2025, while revenue growth was only 30% and capital expenditure growth was 70%.
Addressing concerns about backlog quality—specifically, high concentration among AI lab clients—Citi's research concludes that growth is now widely distributed across traditional enterprises. Data center lessor DLR even stated directly that the release of Claude Opus 4.6 stimulated new leasing demand—a chain of causality that was almost unimaginable a year ago.
The market continues to systematically underestimate the scale of capital expenditures. Consensus forecasts significantly underestimated capital expenditures for hyperscale cloud providers in both 2024 and 2025. Citi expects this pattern to persist for the next five years.
In 2026, hyperscale cloud providers' capital expenditure plans are approximately 70% higher than in 2025. Citi has raised its combined capital expenditure forecast for Amazon (AWS), Google, Meta, Microsoft (Azure), and Oracle in 2026 to $678 billion. Global AI-related capital expenditures—including private cloud, emerging cloud providers, and sovereign AI spending—are projected to reach $770 billion in 2026, climbing to approximately $2.9 trillion by 2030, representing a compound annual growth rate of 47.5%.
The factors driving up costs are not limited to equipment prices—increases in memory and storage prices are significant—but also include the capitalization of power. Hyperscale cloud providers are increasingly shifting power generation from operational expenditure to capital expenditure, requiring them to build their own power supplies for projects. The non-binding "Build Your Own Power Plant" (BYOPP) commitment signed by Google, Microsoft, Meta, Oracle, xAI, OpenAI, and Amazon is a direct manifestation of this structural shift. Consequently, Citi has raised its capital expenditure assumption per gigawatt (GW) of data center capacity for 2026-2027 by approximately 30%, suggesting that the commonly used rule-of-thumb estimate of around $50 billion per GW risks being too low.
The disruption to the software industry is not yet priced into consensus forecasts. "No one is using vibe coding for SAP"—Citi uses this phrase to acknowledge that there are indeed boundaries to technological diffusion; productivity gains in code development cannot be directly extrapolated across the entire enterprise. However, this does not change a larger logic: AI, as a technology with near-zero marginal cost of scaling, is replacing tools whose costs scale linearly with usage. This represents a fundamental business model restructuring, not merely a functional iteration.
Pressure on traditional software companies comes from two directions: first, AI-native competitors, including many VC-backed new entrants, are continuously eroding their market share; second, there is pressure from seat contraction and pricing, as AI enables fewer users to accomplish more.
Citi believes the logic that previously supported software valuation premiums—high switching costs, strong pricing power, and high moats—is being re-evaluated, but consensus earnings forecasts have not yet fully incorporated this ultimate impact. Judging by valuation trends, the market is already voting, but the vote is not yet complete.
Furthermore, across the entire AI technology stack, Citi identifies the best risk-reward profile as concentrated in bottleneck areas of the infrastructure layer: memory and storage, optical interconnects and networking, and power equipment. Hyperscale cloud providers, due to their recent underperformance relative to the broader market, are also listed as opportunities worth watching.
Comments