Enterprise Software Companies, Not AI Labs, Will Capture the Agent Economy's Real Value

What's New

The long-term value creation in AI will concentrate in the application layer, specifically in enterprise software companies that own workflow context, not in the semiconductor or cloud infrastructure layers that have absorbed most of the capital so far. Monti Saroya, Co-Head of Vista Equity Partners' flagship fund, argues this in a conversation on Alt Goes Mainstream. Saroya draws a direct parallel to the internet era: chips and cloud providers captured early returns, but the software companies that built on top of them generated the largest sustained value. The same pattern is repeating with AI, and the enterprise software companies that embed agents into their existing workflows hold the structural advantage because they already know what the end user should be doing.

Why It Matters

The prevailing allocation of AI capital has favored infrastructure: NVIDIA, hyperscalers, and frontier model labs. Saroya's thesis challenges the assumption that model superiority determines who wins. If workflow context matters more than model intelligence, then vertically specialized software companies with existing customer bases can swap between models as commoditization accelerates, while frontier labs compete on a narrowing performance gap. For allocators, this reframes AI exposure from a chip and cloud infrastructure bet to an enterprise software underwriting question.

Big Picture Drivers

• Workflow context as moat: Enterprise software companies know exactly what a knowledge worker should be doing because the work happens inside their systems. A claims processor at an insurance company logs into Duck Creek, a Vista portfolio company. That system captures the workflow, making it the natural surface for agent deployment.

• Token economics hitting enterprise budgets: Until Q1 2026, token costs were an abstraction. Saroya cites one company that generated a $500 million token bill, forcing enterprises to confront the unit economics of AI for the first time. This is shifting procurement conversations from "deploy AI everywhere" to "deploy AI where the ROI math works."

• Model routing as cost lever: Enterprises are using frontier models for trivial queries. Saroya notes the pattern of asking the most intelligent model to answer questions that a basic lookup could handle. Routing simple tasks to cheaper, smaller models while reserving frontier models for complex reasoning can cut costs by orders of magnitude.

• Custom silicon reshaping inference economics: GPUs are optimized for training, not inference. Vista's investment in SambaNova reflects a thesis that purpose-built inference chips will deliver dramatically better cost performance. Intel processors, at roughly 100x lower cost than NVIDIA GPUs, can handle inference workloads that don't require GPU-scale parallel processing.

• Sovereign compute as infrastructure buildout: Governments, particularly in Europe, are demanding domestic AI compute capacity. This creates a new infrastructure cycle where countries build their own AI processing facilities, driven by data sovereignty requirements and the realization that dependence on U.S. hyperscalers is a strategic vulnerability.

By The Numbers

• $500 million token cost bill at a single enterprise, the inflection point that made tokenomics a C-suite priority

• 100 portfolio companies across Vista's platform, providing real-time deployment data on what AI strategies actually produce returns

• 100x cost difference between Intel processors and NVIDIA GPUs for inference workloads, highlighting the arbitrage opportunity in right-sizing compute

• Q1 2026 when enterprise token costs first became a material budget line item, shifting AI from experimentation to unit economics discipline

Key Trends to Watch

• Vertical agent specialization: The next wave of AI companies won't be horizontal platforms. They will be workflow-specific: not "insurance AI" but "personal lines claims processing AI." Saroya predicts this granularity will define where agent ROI concentrates over the next 12 to 18 months.

• Open source model convergence: As open source models like Llama close the gap with proprietary frontier models for most enterprise tasks, the premium for proprietary models compresses. Enterprise software companies benefit because they can swap underlying models without losing their workflow advantage.

• Cloud-to-on-premise reversal for inference: The economics of running inference workloads in public cloud are pushing enterprises back toward on-premise and sovereign compute. Saroya frames this as a structural shift, not a temporary cost-cutting measure, because inference at scale requires different hardware than training.

Memorable Quotes

• "We're using the most intelligent models for the simplest questions which cost the most." Saroya crystallizes the misallocation of AI compute across enterprises, where frontier models serve as overqualified lookup engines.

• "The thing that people forget is the specific things that you do as a knowledge worker are done in some system. That system tends to be an enterprise software system." This frames why software companies, not model providers, hold the structural position in agent deployment.

• "Agents that the humans do a bunch of stuff offscreen that is important that is not captured by software, that's a different thing." Saroya identifies the boundary condition for agent deployment: the gap between what software tracks and what humans actually do remains the hardest automation challenge.

The Wrap

Saroya's thesis succeeds if enterprise software companies demonstrate that workflow context, customer relationships, and vertical specialization matter more than model capability in driving agent ROI. The test is whether these companies can maintain pricing power as token costs fall and open source models become interchangeable. If model commoditization accelerates faster than vertical software companies can embed agents, the value may instead concentrate in the infrastructure layer or in new entrants purpose-built for agentic workflows. The next two to three quarters of enterprise AI deployment data, particularly from large portfolio platforms like Vista's, will reveal whether the app layer thesis is playing out or whether the value chain is settling differently than the internet analogy suggests.