While everyone is selling off software stocks, HSBC says you're wrong

Market Panic Is a Misjudgment

By: Uchu Bōmei, Deep Tide TechFlow

In February 2026, the tech stock market is experiencing a systemic crash that some media are calling the “SaaSpocalypse” (the end of SaaS).

Salesforce’s stock price has fallen nearly 40% from its 2025 peak; after its quarterly earnings report, ServiceNow plummeted over 11% in a single day, mainly because management mentioned during a conference call that “AI agents are complicating visibility into seat growth”; Workday dropped over 22%; the entire S&P 500 Software and Services Index lost nearly $1 trillion in market value within the first six weeks of 2026.

The market logic is straightforward: AI agents can already replace many manual operations. Companies using AI have completed tasks that previously required 100 people, so naturally, they no longer need 100 software seats. The SaaS business model, which charges per seat, is considered to have reached the end of its lifecycle.

Amid this panic sweeping the industry, Stephen Bersey, head of US tech research at HSBC, published a provocative research report titled “Software Will Eat AI.”

His core point, summarized in one sentence: Market panic is a misjudgment.

A Contrarian Report

“Market concerns that AI will replace enterprise software are mistaken.”

He states this at the beginning of his report. In his view, AI will not eliminate software but will be absorbed by it, becoming an embedded capability layer within enterprise software platforms. Software is not AI’s rival; software is the vehicle through which AI reaches the real world.

This flips the current market narrative. The fear is “AI replacing software,” but Bersey’s judgment is “software will tame AI.”

He draws a historical analogy from the internet era: initially, the value was concentrated in physical infrastructure—servers, fiber optic cables, data centers. Massive capital flowed into hardware infrastructure, and the early struggling internet companies that invested heavily in hardware ultimately became the ones with long-term value. Software is the endpoint of internet value.

Bersey believes AI’s evolution is replaying this same script. 2024 and 2025 are infrastructure-building years—computing power, models, code integration—all paving the way for a software explosion. 2026 is the year the engine truly ignites.

“Software will be the main mechanism for AI to diffuse across the world’s largest enterprises. We believe 2026 will be the year software monetization begins.”

Why Can’t Foundational Models Replace Enterprise Software?

The most compelling argument in the report is a layered dismantling of the logic that “AI directly disrupts software.”

Critics’ arguments seem convincing: large language models can write code, vibe coding (generating usable software directly from natural language) is emerging, AI model companies are experimenting more at the application layer. So why do enterprises still need costly traditional software systems like Oracle, SAP, Salesforce?

Bersey’s response unfolds in three levels.

First, foundational models have “inherent flaws.”

The report clearly states that foundational models “have intrinsic limitations” and cannot perform “comprehensive replacement” of core enterprise platforms. They perform well in narrow scenarios—image generation, small app development, text processing—but for high-fidelity, enterprise-grade core platforms, this is “not realistic.”

The fundamental reason is the limitations of training data. LLMs are trained on publicly available internet data, but the private architecture knowledge, business logic, operational norms accumulated over decades in enterprise systems—these core intellectual properties are not online and cannot be learned or replicated by AI. The moat of Oracle and SAP’s systems is not something that can be caught up with by coding; it’s built over time and through complex business scenarios.

Second, the capabilities of vibe coding are seriously overestimated.

The report directly points out the fatal weakness of vibe coding: it shifts the design responsibility entirely onto developers. If you tell AI “I want a system that handles global supply chains,” AI can generate code, but “how to define the system architecture, handle exceptions, ensure stability under extreme stress”—these judgments still require human input.

More critically, Bersey notes that major AI model companies “have almost no experience in creating enterprise-grade software.” They are entering a highly complex environment from scratch. Enterprise software has evolved over decades to achieve “almost zero errors, high throughput, high reliability,” standards that AI newcomers cannot meet in the short term.

Third, the cost of switching for enterprises is a real high wall.

Even if AI can generate code of comparable quality, the cost of replacing core systems remains extremely high—disruption of revenue, loss of productivity, compatibility issues across IT environments, trust built over years with suppliers and service providers. These are real switching costs that won’t disappear just because AI can write code.

Enterprise software demands proven 99.999% uptime, error-free operation in complex IT environments. This trust is earned over time, not just by code.

Who Will Truly Benefit from AI Monetization?

If the first half is defensive reasoning, the second half is an offensive layout.

Bsey’s core judgment: the greatest value in the AI value chain will ultimately flow into the software layer, not hardware or chips.

“We believe AI is the primary source of value creation in the software stack, and the largest long-term value share will belong to software, not hardware.”

He also points out that hardware scarcity—GPU shortages, power constraints, data center bottlenecks—will persist for years. This scarcity reinforces the strategic importance of software platforms: only software platforms can convert AI capabilities into scalable, repeatable business value.

The specific monetization vehicle he points to is AI agents (agentic AI).

Bersey predicts that in 2026, task-oriented, workflow-embedded AI agents will see large-scale deployment in Fortune 2000 companies and SMEs. However, his characterization of agents differs sharply from mainstream narratives; he does not see them as software disruptors but as bounded entities operating within defined parameters and permissions. Only such “boundary-aware” agents can meet enterprise needs for AI risk management.

In other words, enterprises don’t need omnipotent, free-running AI; they need AI that can be governed, audited, and operate within compliance frameworks. Only deeply embedded enterprise software systems can provide this.

“Software is the key pathway for enterprises to control AI use.” This is the most core conclusion of the report.

Additionally, the report predicts that inference demand will gradually surpass training demand, becoming the main driver of computing power growth. As agents become more widespread, computing consumption will not shrink but continue to grow, further supporting the entire software and infrastructure ecosystem.

Opportunity or Trap?

At the time of the report’s release, the software sector’s overall valuation had already fallen to historic lows. Bersey’s view: Low valuations combined with the upcoming monetization year present an opportunity, not a signal to exit.

“Software valuations are at historic lows, even as the industry is on the cusp of large-scale expansion.”

Regarding specific stocks, HSBC’s logic is clear: companies with deep data moats, embedded AI agent capabilities, and business models not solely reliant on headcount billing will be the biggest beneficiaries of this AI monetization wave. Buy ratings include Oracle, Microsoft, Salesforce, ServiceNow, Palantir, CrowdStrike, Alphabet, covering nearly all core enterprise software players.

It’s worth noting that HSBC also downgraded IBM and Asana, and put Palo Alto Networks on a “reduce” list—meaning not all software companies will safely weather the storm. The key is whether they can serve as foundational infrastructure for AI agents rather than being bypassed by intelligent agents as manual interfaces.

Bsey’s logical, timely report, with its contrarian stance, has a strong propagative effect.

But one question remains unaddressed: if AI agents can truly operate efficiently within enterprise software frameworks, will enterprise demand for software “seats” quietly decline? The value of software as an AI vehicle may hold, but whether the “per seat” business model can sustain current valuations remains an open question.

Will software devour AI, or will AI devour software? This debate will find new evidence in every financial report of 2026.