4 min read

The Software Seat-Eater

The Software Seat-Eater
Photo by Joan Gamell / Unsplash

The Per-Seat Problem

Salesforce, a titan of the cloud software era, has become one of the worst-performing stocks in the S&P 500 this year. The stock has plunged 23%, a dramatic reversal for a company that had soared over 150% in the preceding two years. The slump stands in contrast to the fortunes of tech giants like Microsoft and Oracle. The anxiety is palpable on Wall Street, where the question is no longer about quarterly performance but long-term survival. As Bloomberg put it:

While Salesforce, which makes customer relationship management software, has its own AI products, concerns about its long-term trajectory mean the shares may languish even if the company posts strong earnings when it reports after the bell on Wednesday. … "I’m not expecting a major AI contribution to Salesforce over the next several quarters, and in the meantime there are a lot of concerns that AI can disrupt and replace the legacy software names," said Jake Seltz, who manages the Allspring LT Large Growth ETF, which owns Salesforce shares.

Salesforce's struggles are a canary in the coal mine for the SaaS business model. For two decades, software has run on a simple and incredibly lucrative formula: the "per-seat" subscription. A company with 1,000 salespeople paid for 1,000 Salesforce licenses; if they hired 100 more, their bill went up. It was a model that tied revenue directly to a customer's headcount.

The threat comes from a new paradigm known as "agentic AI." Unlike tools that merely assist human workers, AI agents are designed to function as autonomous "digital workers." If one AI agent can do the work of 10 human sales reps, a company no longer needs to pay for all those seats. In this paradigm, the fundamental link between headcount and software spend will be severed.

This is the great AI divide that Wall Street is now pricing in. Microsoft and Oracle are seen as winners not because they sell simple tools, but because their positions in the tech stack give them a unique and powerful way to capture AI value. Microsoft has an unparalleled ecosystem lock-in; it owns the digital real estate where business workflows already reside (Windows, Office 365, Azure). This makes upselling an AI layer like Copilot a natural extension of an existing, sticky relationship. Oracle, meanwhile, is winning because its high-performance infrastructure is uniquely suited for the intense demands of AI workloads, positioning it as a foundational infrastructure layer for the new economy.

Traditional SaaS companies like Salesforce and Adobe, however, face a brutal dilemma. This is where the tension lies: the immediate, sellable version of AI is not the same as the long-term, existential version of AI, and this creates a trap. The AI features they can offer today are incremental—the "slightly smarter CRM"—which are not transformative enough to justify big price hikes from customers. This makes AI a massive R&D cost that squeezes margins without a clear new revenue stream.

But the truly disruptive agentic AI—the kind that functions as a "digital worker"—is a direct threat to their per-seat business model. A customer won't pay for 1,000 human seats if it can eventually buy 100 AI agents to do the same work for less. Salesforce can't aggressively build this second, more powerful form of AI without cannibalizing its own multi-billion-dollar cash cow. So it's stuck selling the weaker, incremental version, leaving it vulnerable to a disruptor who has nothing to lose.

Salesforce, to be fair, is not standing still. Its "Agentforce" platform and its $8 billion acquisition of Informatica are attempts to pivot from selling seats to selling automated outcomes. But it is a high-wire act. The 23% drop in its stock this year suggests that, for now, investors believe the platform players have a fundamentally stronger hand.


The Nvidia Defiance

One of the basic assumptions of a state-driven economy is that when the state tells its most important companies to do something, they do it. And yet, in the critical arena of AI, that is not quite what is happening in China. The government in Beijing has been strongly discouraging its tech champions from buying Nvidia’s AI chips. But those very champions—Alibaba, ByteDance, Tencent—are still trying to buy them anyway.

This peculiar standoff reveals a fundamental tension at the core of the global tech war: the hard logic of the market is, for now, overriding the top-down logic of industrial policy. Beijing’s goal is to force its national tech industry to become self-reliant by using homegrown chips from Huawei. The goal of Alibaba and ByteDance, however, is to win the brutal commercial race for AI dominance, and to do that, they need the best tools available, regardless of their origin.

The reason for their quiet defiance is simple pragmatism. Here is Reuters on the reality facing China’s tech giants:

Despite that pressure, demand for Nvidia chips remains strong in China due to constrained supplies of products from domestic rivals such as Huawei and Cambricon, the four sources said.

Another three sources who are involved in engineering operations at Chinese tech firms also said Nvidia's chips perform better than domestic products.

The performance gap is apparently significant enough that these companies are willing to navigate the political risks. Sources say they are not only looking to secure orders of Nvidia's current H20 chip, but are already monitoring plans for its next-generation China-specific processor, the B30A, which could be up to six times more powerful. That is a performance delta you cannot easily ignore, even when your government is telling you to.

For Nvidia, this creates a bizarre but advantageous position. It is caught between a US government that restricts its sales and a Chinese government that discourages them, but it still holds the one trump card that matters most: a superior product that its customers believe they cannot live without. Beijing can order its companies to buy the homegrown alternative, but it cannot yet order that alternative to be better.


The Scoreboard

  • Fintech: JPMorgan to launch digital retail bank in Germany in 2026 (Reuters)
  • AI: OpenAI boosts size of secondary share sale to $10.3 billion (CNBC)
  • Software: Salesforce issues weak revenue guidance even as earnings beat estimates (CNBC)

Enjoying these insights? If this was forwarded to you, subscribe to ARPU and never miss out on the forces driving tech:

Subscribe