Market underestimated? "Model turning point" drives the rise of AI agents, rapidly unlocking enterprise application scenarios

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

According to Hard AI, Citi Bank analyst Heath Terry's team stated in their latest report that enterprise applications are moving from last year's pilot phase fully into production deployment, with the speed of model improvement outpacing all previous periods and the entire industry 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 its capital expenditure forecast for the same period from $8.0 trillion to $8.9 trillion. Their judgment is: The market is still focused on risks such as the difficulty of data center construction, financing pressures, and intensified competition, but is overlooking the high returns these investments are generating and the formation of a productivity cycle driven by enterprises.

For the software industry, this is a moment more perilous than most realize. As AI-native companies' revenue curves steeply climb, traditional software vendors' previously relied-upon high switching costs, strong pricing power, and high barriers to entry are being repriced by AI technology. This repricing has already occurred at the stock level—over the past year, software stock valuations have obviously diverged from those related to AI infrastructure—but Citi believes consensus earnings forecasts have yet to reflect the ultimate impact.

At the infrastructure layer, especially memory, storage, CPU, and power, Citi currently sees the best risk-reward ratio. The phase lag of hyperscale cloud providers is regarded as another opportunity window.

Model capabilities are improving at a steeper rate

GPT-5.4, Gemini 3.1 Pro, and Claude Sonnet 4.6—three leading models released within less than three weeks—have seen capability jumps far exceeding any previous cycle. Measured by ARC-AGI-2 scores, Gemini 3.1 Pro's score improved 1.5 times over the previous version three months ago; GPT-5.3-Codex is OpenAI’s first model to participate in generating its own code—a milestone hard to ignore.

More noteworthy is that as model capabilities rise, token pricing is also increasing. Inference models employ techniques like Mixture of Experts (MoE) and RLVR (Reinforcement Learning from Verifiable Rewards), which consume more tokens per response. Although Gemini 3.1 Pro’s pricing remains on par with the previous generation, its intelligence score has doubled.

Citi believes the combined effects of these two trends mean there is structural upside for unit income of AI service providers. Capability improvements have begun to permeate enterprise decision-making. In Block’s recent layoff announcement, AI was explicitly mentioned—an early signal that technical diffusion is moving from development to operations.

The transition from pilot to production deployment is faster than expected

System integrators are a key driver behind this acceleration. Leading consulting firms are transforming their internal operations while helping traditional enterprises rapidly deploy solutions from companies like Anthropic and OpenAI, playing the role of AI diffusion's “capillaries.” Citi’s field research with CIOs, CTOs, and system integrators shows competitive pressure is the main force driving enterprises to accelerate—no one wants to let rivals get ahead.

The numbers bear it out: AWS, GCP, Azure, and CoreWeave saw combined backlog order growth of 100% in Q4 2025, while revenue growth was just 30% and capital expenditure growth was 70% for the same period.

Regarding concerns over backlog quality (high concentration of AI lab clientele), Citi’s research concludes that growth is broadly distributed among traditional enterprises. Data center lessor DLR even stated that the release of Claude Opus 4.6 drove new leasing demand—this transmission chain was almost unimaginable a year ago.

The market is still systematically underestimating the scale of capital expenditure

In 2024 and 2025, consensus estimates for hyperscale cloud providers’ capital expenditures are significantly underestimated. Citi expects this to persist for the next five years.

In 2026, hyperscale cloud providers’ capital expenditure plans are about 70% higher than 2025. Citi has raised its combined capex forecast for Amazon (AWS), Google, Meta, Microsoft (Azure), and Oracle in 2026 to $678 billion. Global AI-related capex (including private cloud, emerging cloud providers, and sovereign AI spend) is expected to reach $770 billion, rising to about $2.9 trillion by 2030, with an annual compound growth rate of 47.5%.

Cost drivers include not only equipment prices—rises in memory and storage are key factors—but also the capitalization of electricity. Hyperscale cloud providers are increasingly shifting power generation from OPEX to CAPEX, requiring self-built power supply for projects. The “Build Your Own Power Plant” (BYOPP) non-binding commitment co-signed by Google, Microsoft, Meta, Oracle, xAI, OpenAI, and Amazon directly reflects this structural shift. Citi is therefore raising its capital expenditure per GW for data centers in 2026–2027 by about 30%. Previously, the market generally used an estimate of about $5 billion/GW, which now faces upward bias risk.

Disruption of the Software Industry: Consensus Forecasts Have Not Priced It In

“No one is doing SAP with vibe coding”—Citi acknowledges that technical diffusion has boundaries; productivity gains in code development can’t be directly extrapolated to entire enterprises. But this doesn’t change the bigger logic: AI is using technology with zero marginal scaling cost to replace tools whose costs expand linearly with usage—a fundamental restructuring of business models, not a functional iteration.

For traditional software companies, pressure comes from two directions: first, AI-native competitors (including numerous VC-backed new entrants) continue to eat away at the market; second, seat contraction and pricing pressure, as AI enables fewer users to do more.

Citi believes the logic supporting software valuation premiums in the past—high switching costs, strong pricing power, and high moats—is being re-examined, but consensus profit forecasts haven’t fully factored in the ultimate impact. From valuation trends, the market is already voting; but the vote isn’t finished yet.

In addition, across the AI tech stack, Citi believes the best risk-reward ratios are concentrated in bottleneck links at the infrastructure layer: memory and storage, optical interconnect and networks, and power equipment. Hyperscale cloud providers, having recently underperformed, are also identified as worth watching for opportunity.

This article is from WeChat Official Account “Hard AI”. For more cutting-edge AI news, click here.

Risk Warning and DisclaimerMarkets carry risks. Investments should be made with caution. This article does not constitute individual investment advice, nor does it take into consideration specific users' investment objectives, financial situations, or needs. Users should evaluate whether any opinions, viewpoints, or conclusions in this article are suitable for their particular circumstances. Investing based on this is at your own risk.