"Software will devour AI": Amidst the SaaS doomsday panic, HSBC says you're wrong.

Written by: Naruto Cosmic Wave, Fukami TechFlow

In February 2026, the technology stock market was experiencing a systemic collapse that some media outlets called the "SaaSpocalypse" (the end of SaaS).

Salesforce's stock price has fallen nearly 40% from its 2025 high; ServiceNow's stock price plummeted more than 11% in a single day after the release of its quarterly earnings report, simply because management mentioned in a conference call that "AI agents are complicating visibility into seat growth"; Workday fell more than 22%; and the entire S&P 500 Software & Services index has lost nearly $1 trillion in market value in the first six weeks of 2026.

The market logic is straightforward: AI agents have already replaced a large number of manual operations. Enterprises have used AI to complete tasks that previously required 100 people, so naturally, they no longer need 100 software workstations. The SaaS business model, which charges per workstation, is considered to have reached its end.

Just as this panic trading swept across the industry, Stephen Bersey, head of US technology research at HSBC, released a research report with a highly provocative title: "Software Will Eat AI" .

His core argument can be summarized in one sentence: the market panic is a misjudgment.

A report that comes against the trend

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

He wrote this at the beginning of the report. In his view, AI will not destroy software, but will be absorbed by software, becoming a capability layer embedded within enterprise software platforms. Software is not the opponent of AI; software is the vehicle through which AI reaches the real world.

This logic flips the entire narrative framework of the current market. The market's fear is that "AI will replace software," while Bersey's judgment is that "software will tame AI."

He cited a historical analogy from the internet era: when the internet exploded, the initial value accumulation was concentrated in physical infrastructure—servers, fiber optic cables, and data centers. A large amount of capital poured into hardware infrastructure, while those struggling early internet companies were the ones who ultimately won long-term value. Software, however, is the ultimate source of internet value.

Bersey believes that the evolution of AI is repeating the same script. 2024 and 2025 are the period of infrastructure construction, with computing power, models, and code integration all paving the way for the explosion at the software layer. And 2026 will be the year the engine truly ignites.

“Software will be the primary mechanism for the spread of AI among the world’s largest enterprises. We believe 2026 will be the starting year for software monetization.”

Why can't basic models replace enterprise software?

The report's most compelling argument is its step-by-step breakdown of the logic that "AI will directly disrupt software."

The critics' arguments seem quite convincing: large language models can already write code, vibe coding (generating usable software directly from natural language descriptions) is on the rise, and AI model companies are already making more application-layer attempts, so why do enterprises still need expensive traditional software systems like Oracle, SAP, and Salesforce?

Bersey's answer unfolded on three levels.

First, the basic model has "inherent defects".

The report explicitly points out that the basic models "have inherent flaws" and are incapable of "completely replacing" the core platforms of large enterprises. They perform well in narrow scenarios, such as image generation, small application development, and text processing, but for high-fidelity, enterprise-level core platforms, this is "unrealistic."

The root cause lies in the limitations of training data. LLM is trained on publicly available internet data, while the proprietary architectural knowledge, business logic, and operational standards accumulated by enterprise software systems over decades—these core intellectual property rights are not on the public internet. AI cannot learn from them, nor can it replicate them. The system moats of Oracle and SAP cannot be caught up by writing code; they are built up over time and through business scenarios.

Second, the capability boundaries of Vibe Coding are seriously overestimated.

The report directly points out Vibe Coding's fatal flaw: it places the entire responsibility and burden of design on the developers. You tell the AI, "I want a system that can handle the global supply chain," and the AI ​​can generate code, but the judgments on "how to define the architecture of this system, how to handle abnormal situations, and how to ensure it doesn't collapse under extreme pressure" still require human intervention.

More importantly, Bersey points out that the major AI model companies "have almost no experience in building enterprise-grade software." They are entering an extremely complex environment from scratch. Enterprise software, on the other hand, has evolved over decades to a level of "near-zero errors, high throughput, and high reliability," a benchmark that AI upstarts cannot reach in the short term.

Third, the cost of business switching is a real and significant barrier.

Even assuming that AI can indeed write code of the same quality, the cost for enterprises to replace their core systems remains extremely high. Risks of revenue interruption, loss of productivity, system compatibility issues across IT environments, and the accumulation of trust in supplier brands and service capabilities are all real switching costs that will not disappear just because AI can write code.

Enterprise-level software requires 99.999% uptime, proven over many years, and error-free operation in various complex IT environments. This trust is earned over time, not built with piles of code.

Who will be the real beneficiaries of AI monetization?

If the first half of the report is a defensive argument, then the second half is an offensive strategy.

Bersey's core judgment is that the largest share of the AI ​​value chain will ultimately flow to the software layer, rather than the hardware and chip layer.

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

He also pointed out that hardware scarcity—GPU shortages, power constraints, and data center bottlenecks—will continue to exist for years to come. This scarcity precisely reinforces the strategic importance of software platforms: only software platforms can transform AI capabilities into scalable and repeatable commercial value.

The report points to AI agents as the specific means of monetization.

Bersey predicts that 2026 will see large-scale deployments of task-oriented, workflow-embedded AI agents in Fortune 2000 companies and SMEs. However, his characterization of agents differs significantly from the mainstream narrative in the market. He does not believe that agents are disruptors that replace software, but rather that agents must operate within the parameters and permissions defined by software. It is precisely this "bounded agent" that can meet the needs of enterprises for AI risk management.

In other words, enterprises don't need an omnipotent, free-roaming AI; they need an AI that can be governed, audited, and operate within a compliance framework. And this can only be achieved by intelligent agents deeply embedded in enterprise software systems.

"Software is the key way for enterprises to use AI in a controlled manner." This is the most crucial judgment in the entire report.

Meanwhile, the report also predicts that inference demand will gradually surpass training demand, becoming the main driver of increased computing power consumption. This means that as intelligent agents become more widespread, computing power consumption will not shrink but will continue to grow, further supporting the entire software and infrastructure ecosystem.

Opportunity or trap?

When the report was released, the overall valuation of the software sector had already fallen to a historic low. Bersey's assessment was that the combination of low valuations and the upcoming year of monetization presented an entry opportunity, not a signal to exit.

"Software valuations are at historical lows, even though the industry is on the eve of a massive expansion."

HSBC's logic for recommending specific targets is clear: software companies that have built strong data moats, possess the ability to embed AI agents, and do not rely on a purely per-capita billing model will be the biggest beneficiaries of this wave of AI monetization. The list of companies with buy ratings includes Oracle, Microsoft, Salesforce, ServiceNow, Palantir, CrowdStrike, Alphabet, and others, covering almost all the core players in enterprise software.

It is worth noting that HSBC also downgraded the ratings of IBM and Asana, and placed Palo Alto Networks on the "reduce" list. Not all software companies can weather this storm safely. The key is whether they can become the infrastructure for the implementation of AI agents, rather than a human interface that can be bypassed by the agents.

Bersey's report is logically sound, timely, and its contrarian stance has a strong communicative effect.

However, the report did not directly answer one question: If AI agents can truly operate efficiently within the framework of enterprise software, will the demand for software "seats" still be quietly shrinking? The value of software as a carrier of AI may be valid, but whether the "per-person" business model can support the current valuation remains an open question.

Whether software devours AI or AI devours software, every financial report in 2026 will provide new evidence for this debate.