Is the market underestimated? "Model inflection point" drives the rise of AI agents, rapidly unlocking enterprise application scenarios

The AI wave is undergoing a critical inflection point, with the rapid leap in model capabilities accelerating last year's enterprise pilot projects into production deployment, while the market continues to systematically underestimate the depth and speed of this transformation.

According to Hard AI, a team of Citi analysts led by Heath Terry stated in their latest report that enterprise-level applications have shifted comprehensively from last year's pilot phase to production deployment, with model capabilities improving faster than ever before, and the industry's demand curve rising sharply.

Citi has raised its total AI industry revenue forecast for 2026–2030 from $2.8 trillion to $3.3 trillion, and capital expenditure forecast for the same period from $8.0 trillion to $8.9 trillion. Their assessment is: while the market is still focused on risks such as the difficulty of building data centers, financing pressures, and intensifying competition, it is ignoring the high returns these investments are already generating, as well as the emergence of an enterprise-driven productivity cycle.

For the software industry, this is a more dangerous moment than most people realize. As revenue curves of AI-native companies steeply climb, the high switching costs, strong pricing power, and high entry barriers that traditional software vendors have relied on are being repriced by AI technology. This repricing has already been reflected in stock valuations—the software sector’s valuation trend has clearly diverged from AI infrastructure targets over the past year—but Citi believes consensus earnings forecasts have far from captured the final impact.

At the infrastructure layer, especially memory, storage, CPU, and power segments, Citi currently believes the risk-reward ratio is optimal. The recent underperformance of hyperscale cloud providers is seen as another window of opportunity.

Model capabilities are rising at an even steeper slope

GPT-5.4, Gemini 3.1 Pro, and Claude Sonnet 4.6—three leading models—were released less than three weeks apart, with the magnitude of capability improvement far exceeding any previous cycle. Measured by ARC-AGI-2 scores, Gemini 3.1 Pro improved 1.5 times over the previous version from three months ago; GPT-5.3-Codex is OpenAI’s first model involved in generating its own code—a milestone that cannot be overlooked.

More noteworthy is that as model capabilities improve, token pricing is also rising. Inference models adopt technologies such as Mixture of Experts (MoE) and Reward Verification Reinforcement Learning (RLVR), consuming more tokens with each response. Although the pricing of Gemini 3.1 Pro remains the same as the previous generation, its intelligence score has already doubled.

Citi believes that the interplay of these two trends means AI service providers have structurally higher unit revenue potential. Capability improvements are already infiltrating specific enterprise decisions. In Block’s recent layoff announcement, AI was explicitly mentioned, signaling the early stages of technology diffusion from development into operations.

The shift from pilot to production deployment is happening faster than expected

System integrators are the main drivers of this acceleration. Leading consulting firms are transforming their own internal operations while helping traditional businesses quickly deploy solutions from Anthropic, OpenAI, and others, acting as the "capillaries" of AI proliferation. Field studies by Citi with CIOs, CTOs, and system integrators indicate that competitive pressure is the core driver of enterprise acceleration—no one wants to let their competitors get ahead.

Numbers back this up: AWS, GCP, Azure, and CoreWeave had a combined backlog order growth of 100% in Q4 2025, while revenue only grew 30%, and capital expenditures grew 70% over the same period.

Addressing external concerns about backlog quality (high concentration of AI lab clients), Citi’s research concludes that growth is already widely distributed among traditional enterprises. Data center lessor DLR even stated that the launch of Claude Opus 4.6 triggered new leasing demand—something that was almost unimaginable a year ago.

The market is still systematically underestimating the scale of capital expenditure

Consensus forecasts for hyperscale cloud providers' capital expenditure in 2024 and 2025 are significantly underestimated. Citi expects this situation to persist for another five years.

In 2026, hyperscale cloud providers' capital expenditure plans are about 70% higher than in 2025. Citi now estimates the combined 2026 capital expenditure of Amazon (AWS), Google, Meta, Microsoft (Azure), and Oracle to reach $678 billion, global AI-related capital expenditures (including private cloud, emerging cloud vendors, and sovereign AI spending) to reach $770 billion, and to climb to about $2.9 trillion by 2030, for a CAGR of 47.5%.

It’s not just device prices—such as increases in memory and storage costs—that are pushing up costs, but also the capitalization of electricity. Hyperscale cloud providers are increasingly shifting power generation costs from operational to capital expenditures, requiring them to build their own power supplies for projects. The "Build Your Own Power Plant" (BYOPP) nonbinding pledge co-signed by Google, Microsoft, Meta, Oracle, xAI, OpenAI, and Amazon directly reflects this structural transformation. As a result, Citi has raised its estimate of capex per GW of data center for 2026–2027 by about 30%, noting that the widely used market average estimate of around $5 billion/GW may be too low.

"No one is using vibe coding to do SAP"—Citi uses this phrase to acknowledge the boundaries of technology diffusion, as productivity gains in code development cannot be directly extrapolated to the entire enterprise. But the bigger logic remains: AI is replacing tools whose costs expand linearly with usage with a technology that scales at zero marginal extension cost—this is a fundamental reconstruction of business models, not just a functional iteration.

For traditional software companies, pressure comes from two directions: First, AI-native competitors (including many VC-backed new entrants) continue to eat into the market; second, shrinking seat counts and pricing pressures, as AI enables fewer users to accomplish more.

Citi believes the logic underpinning the software sector’s premium valuations—high switching costs, strong pricing power, high barriers—are being reevaluated, but consensus earnings forecasts have not yet fully reflected the ultimate impact. From valuation trends, the market is already voting; it’s just that the vote isn’t over yet.

In addition, across the AI technology stack, Citi believes the best risk-reward ratios are in the bottleneck segments of the infrastructure layer: memory and storage, optical interconnect and networking, and power equipment. Hyperscale cloud vendors, who have recently underperformed, are also now regarded as opportunities worth attention.