a16zThe Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z
CHAPTERS
- 0:00 – 0:47
Agentic coding hits product–market fit and exposes the capacity squeeze
Benedict opens with the view that agentic coding has rapidly progressed from “useful” to genuinely transformative. The conversation immediately frames AI’s biggest near-term constraint as not ideas but infrastructure scarcity and cost.
- •Agentic coding is becoming a step-change in what software can do
- •Near-term AI progress is gated by compute capacity, not demand
- •The industry can’t sustainably spend on the order of $10T/year on AI infrastructure
- •Framing: today’s magic becomes tomorrow’s invisible baseline
- 0:47 – 2:51
What’s changed since last year: strategy divergence and a narrower focus
Reflecting on the last year, Benedict argues the biggest change is clearer product strategy divergence among model labs and a major industry focus on coding. Despite rapid scaling, many foundational questions remain unresolved.
- •Competitive tension now includes product choices, not just ‘bigger models’
- •OpenAI’s strategy has iterated; coding became a focal point
- •Coding is the standout use case with clear pull from customers
- •Big questions remain: winners, value capture, daily consumer usage
- 2:51 – 5:40
Why coding emerged first—and what it might mean for engineering jobs
They discuss whether coding’s breakout was predictable and what it implies for teams and careers. Benedict emphasizes it’s too early for firm conclusions; organizational effects will take years to settle, especially given pricing instability.
- •Developers were the earliest experimenters; it’s natural they pushed coding first
- •Timing of agentic coding’s effectiveness wasn’t deterministic
- •Job structure questions (junior hiring, task automation) are now immediate
- •Team/org outcomes are still unknown and will evolve over years
- 5:40 – 8:37
OpenAI vs. Anthropic: focus, execution, and the adoption gap
Benedict contrasts OpenAI’s broad experimentation with Anthropic’s coding focus that ‘worked.’ He highlights a widening gulf between power users building deep workflows and the majority using AI only occasionally.
- •OpenAI tried many directions; Anthropic’s focus on coding paid off
- •Key challenge: bridging ‘all day’ power users vs. casual weekly users
- •Non-tech companies pursue one-at-a-time point solutions (e.g., cash flow forecasting)
- •Enterprise adoption often looks like targeted automation, not a universal chatbot
- 8:37 – 12:08
Platform history lesson: mobile data pricing crunch as the AI mirror
Benedict compares today’s token pricing confusion to early smartphone-era mobile data shocks: bill surprises, capacity limits, and eventual pricing realignment. The analogy sets up a broader point about infrastructure layers enabling value that accrues elsewhere.
- •Early mobile data saw $5–10k bill shocks and network overloads
- •AI has similar mismatch: cheap subscriptions vs expensive marginal usage
- •Pricing models evolve to match costs and perceived value (caps, throttling analogs)
- •Explosive usage growth doesn’t guarantee infrastructure-layer profitability
- 12:08 – 17:49
Where value accrues: infrastructure vs operating system—and why models may not win
They explore whether foundation models become commodity infrastructure (like telcos/ISPs) or capture OS-like power (like Windows/iOS). Benedict argues models lack classic network effects and may struggle to sustain pricing power as competition and efficiency rise.
- •Telcos built world-changing infrastructure yet captured limited value
- •Central question: can models ‘do the whole thing’ or do apps capture value?
- •Models appear to lack OS-style network effects and leverage
- •Upcoming CapEx + efficiency gains could push the market toward commodity dynamics
- 17:49 – 23:53
Why ‘foundation models aren’t the product’: differentiation, UI limits, and leverage up the stack
Benedict lays out a chain of reasoning: model differentiation is hard to sustain, chatbots are a crude UI, and real value needs tooling, data, workflows, and verticalization. This points to an ecosystem where many apps embed models rather than standardize on one model provider.
- •Sustainable ‘better-than-everyone’ model differentiation is unclear
- •Chatbot is a V1 interface; real products need workflow, data, configuration, guardrails
- •Skill/templates resemble ‘file new’ in Excel—useful but limited at scale
- •Enterprises rarely standardize on a model the way they pick an end-user platform; it’s abstracted like cloud
- 23:53 – 29:04
Key open questions: edge/on-device models, industry-specific disruption, and unknown constraints
They move to what to watch next: when smaller/older/open models become “good enough,” how AI reshapes professional-services pyramids, and how generative AI differs from past shifts because core technical limits are less knowable in advance.
- •‘Good enough’ models may move workloads to open source or on-device compute
- •Professional services (law/consulting/finance) may face pyramid restructuring
- •Unlike past platforms, AI’s future capabilities and cost curves are less bounded
- •Next breakout could come from a non-coding domain—uncertain which
- 29:04 – 31:35
From automation to new possibilities: Jevons paradox and ‘things you couldn’t do before’
Benedict frames AI’s impact not just as doing the old work faster, but enabling new categories of work and business models. He stresses that industry-by-industry outcomes will vary, as with the internet’s uneven effects on newspapers vs. movie studios.
- •Automation triggers price elasticity: do less for cheaper or far more for same spend?
- •AI can remove barriers to entry and unlock new competitive dynamics
- •Most important shifts come from new capabilities, not ‘old tasks but more’
- •Industry outcomes will diverge—analogies help, but don’t predict
- 31:35 – 38:09
Commerce and advertising: richer product understanding and better matching
Benedict argues LLM-like systems can ‘understand’ products and preferences differently from today’s metadata + co-purchase heuristics. This could change discovery, recommendations, and conversion—already hinted at in ad platform performance improvements.
- •Today’s ad/retail systems are constrained by shallow product representations
- •LLMs enable more semantic queries: identify items, find alternatives, compare pros/cons
- •Personalized shopping agents could translate taste signals into concrete recommendations
- •Early evidence: improving conversion and ad performance as AI is integrated
- 38:09 – 43:11
Enterprise software after coding: more software, different interfaces, new competition
Asked about SaaS consolidation, Benedict expects ‘more software’ as building becomes cheaper and new capabilities emerge. He outlines enterprise software’s three-bucket reality (big horizontals, many SaaS/vertical apps, and Excel/email glue) and how LLMs add another option across them.
- •Cheaper/faster building increases competition and product experimentation
- •Enterprise stack: big horizontals (SAP/Workday), many SaaS apps, and Excel/email middle layer
- •LLMs can be embedded as features (bottom of stack) or orchestrators across systems (top of stack)
- •Investors fear a ‘SaaS apocalypse’ but it’s unclear which categories are most exposed
- 43:11 – 48:20
Workflow reality: implicit organizational knowledge and the limits of ‘just add an agent’
Benedict explains why adoption depends on messy, undocumented processes and exception handling—areas where outside consultants often add value. This complicates visions of software with minimal UI or fully agent-driven systems of record.
- •Software and consultants both formalize better workflows—but many workflows are implicit
- •Critical organizational knowledge isn’t documented or in training data
- •Automation bottlenecks are exceptions and judgment calls, not routine steps
- •LLMs excel at ‘average’ answers; harder where differentiation and novel judgment matter
- 48:20 – 52:17
The CapEx problem: financial gravity, existential competition, and where spending tops out
They examine whether ‘overinvesting is safer’ has limits. Benedict argues big tech CapEx levels are historically extreme relative to revenue, and while strategic fear drives spending, there are hard financial ceilings and unresolved ROI math.
- •Big tech’s CapEx can approach >50% of revenue—far above telecom-style intensity
- •$700B/year is already ‘global infrastructure’ scale; growth must eventually taper
- •AI is existential for incumbents, creating FOMO-driven investment behavior
- •ROI is hard to model because key parameters (efficiency, demand, model cadence) are shifting
- 52:17 – 55:08
Token maxing, ROI ambiguity, and consumer surplus: why benefits may be competed away
Benedict predicts pricing and usage will realign, but measuring ROI early is inherently difficult. Much benefit shows up as hard-to-quantify productivity and quality improvements, and even measurable gains may turn into consumer surplus that doesn’t translate into higher prices or profits.
- •Early-stage misuse and bill-shock stories mirror past platform rollouts
- •ROI is often indirect (analytics, support quality, productivity) and slow to quantify
- •Productivity improvements can get competed away (do more work for same price)
- •A likely path: AI becomes a competitive necessity rather than a profit center for many users
- 55:08 – 1:00:32
Advice for model labs and the long-view takeaway: every platform becomes ‘obvious’ later
Benedict reframes commoditization as a challenge question, not a certainty: explain why models won’t become commodity-like. He ends with a historical reminder—transformative technologies feel chaotic early on, then become invisible infrastructure that people take for granted.
- •Commoditization is a plausible chain of logic; not guaranteed but must be addressed
- •Model labs face the ‘then what?’ problem beyond coding as the killer app
- •Expanding into the broader economy requires product/workflow integration, not just bigger models
- •Long arc: today’s ‘magic’ becomes tomorrow’s assumed baseline
