A new unit economics debate: from licensed software to inference-priced capability
Stripe CEO Patrick Collison is putting a sharp, market-moving idea on the table: the AI era may not simply “add features” to software—it may reprice what software is. In his framing, the industry is drifting away from pre-built products that are manufactured once and licensed endlessly, toward on-demand software—capability generated at the moment of use, shaped by context, data, and intent.
This is more than a product design argument. It is an economic thesis about where value accrues when marginal costs re-enter the picture. In classic SaaS economics, the dream is near-zero marginal cost: ship once, sell forever, expand margins with scale. Collison’s “non-Walrasian software regime” suggests a different center of gravity—one where inference costs, customization costs, and runtime governance become the new marginal costs that matter. If every meaningful interaction triggers compute, model calls, retrieval, and policy checks, then “software” starts to resemble a metered utility rather than a static asset.
That proposition helps explain why markets have been jumpy. When enterprise AI capabilities appear to encroach on the traditional software stack—especially workflows historically monetized through subscriptions—investors quickly translate technical possibility into multiple compression risk. The February sell-off following enterprise-oriented developments around Anthropic’s Claude, alongside steep declines in software equities and a historically sharp one-day drop for IBM, reflects a market trying to price an uncomfortable question: if AI can generate workflows on demand, what happens to the rents embedded in packaged software?
Just-in-time applications and the emerging architecture of “runtime software”
The technological shift implied here is not merely “AI inside apps.” It is a move toward apps as ephemeral outputs—assembled dynamically, executed, and then discarded or revised continuously. The closest analogy is serverless computing’s pay-for-execution model, but with a crucial twist: serverless executes known code; AI increasingly creates or adapts code and workflows at runtime.
Several technical implications follow, each with direct business consequences:
- Just-in-time code generation and orchestration
Generative systems can produce UI components, integrations, queries, and automation flows on demand. This competes with template-driven platforms and reduces the advantage of large, static feature catalogs—especially for customers who value “fit-to-purpose” over “feature-rich.”
- Inference-driven architecture becomes a first-class constraint
When AI is invoked at runtime, organizations must manage:
– Latency (user experience and operational throughput)
– Cost per call (unit economics and pricing strategy)
– Model/version control (reproducibility and change management)
– Security and prompt injection risks (new attack surfaces)
- Toolchain convergence across MLOps, DevSecOps, and API management
The “prompt-to-production” pipeline becomes operationally real. Teams will need integrated systems for evaluation, policy enforcement, observability, and rollback—because the “software” being delivered is no longer only a compiled artifact, but also a governed inference process.
This is where Collison’s thesis becomes concrete: if software is increasingly a runtime event, then the operational layer—compute, governance, monitoring, and data access—becomes inseparable from the product. The competitive moat shifts from “who has the most features” to “who can deliver reliable outcomes under variable inference costs and constraints.”
SaaS margins meet metered reality: valuation shocks and monetization redesign
The economic implications are already visible in investor behavior. Traditional subscription software is prized for predictable revenue and high gross margins. But on-demand AI introduces variable costs that scale with usage, complexity, and model choice. That doesn’t make software unprofitable; it makes it less purely financial-engineering-friendly and more operationally exposed.
Key pressures and adaptations are emerging:
- Downward pressure on pure subscription rent models
If customers can obtain tailored workflows via AI agents or dynamic generation, the willingness to pay for broad, static bundles may soften—particularly in categories where differentiation is thin and switching costs are falling.
- A repricing of “AI readiness” in public markets
The February drawdown signals that markets are scrutinizing:
– Whether incumbents can defend margins while embedding AI
– Whether new entrants can undercut pricing with more flexible delivery
– Whether AI features are additive revenue or cannibalization accelerants
- Hybrid monetization becomes the likely equilibrium
The most credible near-term models blend stability with metering, such as:
– Base subscription + usage-based AI fees
– Outcome-based pricing tied to measurable business value
– Tiered governance and compliance packages (auditability as a premium feature)
This is also where NVIDIA CEO Jensen Huang’s pushback lands: calling the idea of software obsolescence “illogical” is less a denial of AI’s impact than an assertion that enterprises will still demand dependable tools, predictable workflows, and accountable vendors. In that view, AI becomes an accelerant—improving productivity and expanding capabilities—without eliminating the need for robust software platforms and long-lived systems of record.
The truth may be that both perspectives can coexist: AI can compress value in commoditized layers while expanding value in infrastructure, vertical expertise, and governed delivery.
Strategic signals for CIOs and CEOs: vertical moats, compliance-at-invocation, and infrastructure pull
For business leaders, the actionable question is not whether AI “replaces software,” but where the value chain re-anchors when software becomes more dynamic and inference-metered.
Several non-obvious strategic connections stand out:
- Vertical specialization as the new defensibility
As on-demand generation lowers the cost of building generic workflows, differentiation migrates to domain-tuned models, proprietary data, and industry-specific compliance. Horizontal incumbents may face pressure from vertical players that ship faster and fit better.
- Procurement shifts toward a software supply chain
CIOs may increasingly manage a portfolio that includes:
– Model providers
– Data marketplaces and retrieval layers
– Cloud and inference hardware partners
– Policy engines and audit systems
This resembles a supply chain more than a single-vendor software suite.
- Regulation becomes runtime engineering
Real-time inference across jurisdictions raises data sovereignty and privacy constraints that cannot be handled only in contracts. Organizations will need policy-as-code and auditability embedded at invocation time—turning compliance into an architectural requirement, not a legal afterthought.
- Inference hardware demand becomes a strategic lever
If software consumption shifts toward AI calls, then GPUs, TPUs, and edge accelerators become part of the software business model. Capital allocation and vendor leverage may tilt toward those controlling efficient inference capacity.
The Collison–Huang contrast captures a pivotal tension in enterprise technology: software as a durable product versus software as a metered, generated service. The companies that navigate this best will be those that can price and govern inference intelligently, build vertical credibility, and deliver reliability at scale—because in an on-demand world, trust and performance become the features customers notice first.
By
By
By
