Interview, Fireside Chat, Conference Presentation
The Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z
- Agentic coding has transitioned from a "useful" utility to a product with definitive product-market fit, causing the tech sector's focus to narrow significantly toward this specific use case within the last year.
- Software developers are the primary early adopters, creating a deterministic feedback loop where the first effective application of LLMs is software development itself, similar to how early PCs were used to build more computers.
- Current market dynamics are defined by an "extreme scarcity" and disequilibrium where demand for compute capacity vastly outstrips supply, leading to a severe imbalance in pricing and CapEx.
- Benedict Evans predicts that the current period of massive investment (estimated $700B to $1T annually by major tech firms) is unsustainable long-term and will inevitably settle into a lower-growth equilibrium.
- OpenAI's strategy has shifted from a "everything all at once" approach to a concentrated focus on coding, contrasting with Anthropic's similar pivot which allegedly achieved product-market fit by stumbling into coding solutions.
- A significant divergence in user adoption exists between the "20 percent" of users who deeply integrate AI into daily workflows and the "40 percent" who view it as merely useful for specific, occasional tasks.
- Corporate adoption is increasingly moving away from general chatbots toward point solutions for back-office processes, such as commodities companies using LLMs to forecast cash flow and invoice timing.
- The technology is currently in an early adoption phase comparable to mobile data in 2009-2010, characterized by user confusion over capacity limits and unexpected "usage shock" bills.
- Foundation models are unlikely to capture value as end-products due to the lack of network effects; instead, they will likely become low-margin commodities similar to mobile infrastructure or internet bandwidth.
- The chatbot interface is considered a limited "v1" UI that lacks the necessary tooling, data integration, and control required for complex enterprise tasks, which require dedicated application layers.
- Model providers face a difficult pricing dilemma: while current demand appears infinite, historical precedents (mobile data, telecom) suggest prices will eventually compress as supply scales and competition increases.
- Evans notes that while mobile data traffic grew 1,500x to 2,000x over 15 years, the mobile network operators themselves captured little value compared to application-layer companies.
- The "Jevons Paradox" may apply to AI: as tasks become cheaper to perform, organizations will consume significantly more of that computational power rather than reducing overall costs.
- New economic questions are emerging regarding the "pyramid structure" of professional services (law, consulting, finance), where automating entry-level tasks forces a reconfiguration of hiring models and client value propositions.
- Unlike previous platform shifts, AI lacks known physical or economic constraints on speed and cost, making it impossible to accurately predict the rate of future performance gains or price reductions.
- Advertising and e-commerce are identified as high-potential sectors for AI-driven automation, specifically through the ability to understand why consumers buy rather than just what they buy, potentially revolutionizing recommendation engines.
- The software industry may not consolidate further but instead fragment, with a proliferation of "AI-native" tools that allow users to build custom workflows, effectively creating a new layer of "Excel-like" improvisation on top of SaaS.
- A key distinction in AI adoption is between "tasks" (which change frequently) and "jobs" (which remain stable); AI will excel at automating the former while the core professional output often remains unchanged.
- Major tech firms (Google, Meta, Microsoft) are driven by existential FOMO to spend heavily on CapEx, fearing they will suffer the same fate as IBM in the 90s or Microsoft in the 2000s if they miss the AI shift.
- Current ROI measurements for AI are often intangible (e.g., "better analytics," "faster slides"), making it difficult for CFOs to justify massive CapEx until tangible revenue lines or direct cost savings are realized.
- Consumer surplus will likely absorb many productivity gains, turning AI tools into competitive necessities where cost savings are competed away rather than translating to higher margins.
- The long-term trajectory suggests AI will become "magic" and invisible, eventually reaching a state where users in 20 years will assume computers have always performed these complex tasks automatically.
- Benedict Evans advises foundation model companies to consider the mobile infrastructure analogy: they may drive the frontier forward and generate massive value for the ecosystem while capturing only a fraction of the total economic return.