As agentic AI moves into a phase where it can execute work—not just assist with it—the shock to the software industry is beginning not with features, but with how software is priced. The market’s argument is straightforward: companies are increasingly paying not for “software to use,” but for work that gets done—and that shift may bring restructuring pressure to the SaaS sector first.

The reason talk of the “end of SaaS” is spreading is also simple. As AI evolves from a tool that supports people into an agent that performs tasks on their behalf, the 기준 for enterprise spending changes. The first model to feel that impact is the seat-based subscription—the long-standing approach where customers pay based on the number of human users or accounts.

Capital markets are reflecting the same shift. The Financial Times reported that combined AI infrastructure investment by major tech firms—such as Alphabet, Amazon, Meta, and Oracle—has been discussed at roughly $660 billion for 2026. Individual estimates cited in market discussions run into the hundreds of billions of dollars, reinforcing the view that the center of gravity is moving from “software spending” toward “buildout spending.”

In this environment, the SaaS categories most exposed to near-term restructuring share a clear trait: they are workflow products that depend on humans opening a screen and repeatedly inputting, approving, and processing information. If agents begin performing those steps, contract sizes can shrink—not because the products are inferior, but because the economic rationale for seat-based pricing weakens.

The first impact zone is CRM and sales operations. CRM is, in principle, a system of record. In practice, day-to-day performance has long depended on human time spent updating fields, logging activity, and managing follow-ups. If an agent can read emails and meeting notes, automatically log interactions, and generate next actions, companies may evaluate seat reduction before feature expansion.

The second is customer support SaaS, including helpdesk and contact-center platforms. Tier-1 inquiries—delivery, refunds, account issues—are well-suited to automation. As agents move beyond answering to also handling ticket triage and resolution, industries under strong labor-cost pressure may first reduce the number of human agents, and seats can decline in tandem.

The third is marketing operations, SaaS, and execution agencies. Ad setup, creative variations, A/B testing, and reporting are repetitive and well-structured tasks—exactly the kind of work an agent can run “every day” with minimal friction. Even without a clear deterioration in performance, if the operating layer thins out, renewals may increasingly start with seat cuts.

The fourth is back-office software—expense management, accounting, AP/AR, settlement, and HR operations such as HRIS and ATS. Receipt classification, journal entry creation, matching, approval routing, candidate screening, and scheduling are heavily rule-driven. However, auditability, internal controls, and privacy constraints are substantial. In the near term, the dominant pattern may be less “full replacement” and more redesigning operations into “one person plus multiple agents.”

The fifth is project management and collaboration tools. The category can appear resilient, but much of daily work is “status labor”—updates, summaries, and reporting. If agents absorb that layer, teams may achieve the same output with fewer people. At that point, license optimization begins, and pressure rises to shift pricing away from seats and toward work volume and outcomes.

By contrast, the next five years of capital expenditure are expected to focus on addressing the infrastructure bottlenecks necessary to run AI at scale. According to McKinsey, total data-center demand could reach 219 GW by 2030, with AI workloads accounting for as much as 156 GW. The same outlook suggests annual incremental AI capacity could rise from 13 GW in 2025 to 31 GW by 2030.

Electricity is now the critical constraint. The IEA projects data-center power consumption could reach 945 TWh by 2030, roughly approaching a doubling over the period, with growth from 2024 to 2030 averaging about 15% per year. As AI-optimized data centers proliferate, early access to power delivery, cooling, suitable sites, and grid interconnection increasingly defines who can scale.

That is why investment is spreading from chips to the physical layer of the economy—grids and equipment. Reuters reported that Siemens Energy announced a $1 billion plan (Feb. 3, 2026) to expand U.S. production for grid and gas-turbine components, including a large grid-equipment facility in Mississippi targeted for completion in 2028. As data-center loads rise, bottlenecks can emerge simultaneously across transformers, switchgear, cables, power electronics, and cooling—making “on-the-ground infrastructure” a central arena for competition.

In the end, what some describe as a “software collapse” looks less like disappearing demand and more like a change in the revenue formula. Seat-based workflow SaaS may face early pressure through down-seating and pricing renegotiations, while investment expands beyond GPUs into data centers, power interconnection, cooling, and grid equipment. The market’s center is shifting from app-subscription competition toward a new race: who can secure compute and electricity capacity first—and deploy it reliably at scale.

Yeon Choul-woong More by this reporter