Y CombinatorAaron Levie: Why Enterprise Buys Outcomes, Not Models
By treating intelligence as a commodity layer like storage or compute; B2B AI startups capture Jevons paradox gains as enterprise adoption stays under 1%.
CHAPTERS
- 0:00 – 1:15
AI as an abundance engine (Jevons paradox framing)
The conversation opens on an optimistic thesis: AI-driven automation can increase output, lower costs, and raise living standards. Garry frames this as a Jevons paradox-style dynamic—efficiency leading to more usage and broader prosperity rather than a dystopian outcome.
- •AI automation can compound into more building, lower costs, and higher lifestyles
- •Jevons paradox as a lens for understanding AI-driven abundance
- •Rejecting a "Black Mirror" inevitability; choosing an optimistic timeline
- 1:15 – 3:15
Why “GPT wrapper” is mostly a misleading meme
They unpack why calling startups “wrappers” misses where value is created: the workflow software, data handling, and business logic around the model. The only real risk is being directly overlapped by a model provider’s consumer app features, not by the mere existence of a foundation model.
- •There’s a small kernel of truth, but most of the "wrapper" critique is wrong
- •Enterprise value is in workflow, integrations, and proprietary logic—not tokens
- •Real danger: model providers folding features into their consumer products
- •Analogy to early cloud: "wrapping Amazon" misunderstood the real software required
- 3:15 – 5:10
Enterprises buy outcomes, not models: “Just get the workflow done”
Aaron explains that enterprises don’t want an LLM—they want customer support resolved, documents processed, contracts routed, and systems updated. As models improve, B2B products benefit because they can do less “hacky” work and more reliably deliver end-to-end outcomes.
- •B2B customers purchase outcomes (support, EHR transcription, contract workflows)
- •Model improvements increase the value of workflow apps
- •Winning approach: abstract the model away from the customer value proposition
- •Integrations (ERP, support systems) and reliability are the product
- 5:10 – 8:18
Models converge; differentiation shifts to enterprise software packaging
They predict a future where most business use cases won’t strongly distinguish between models as capabilities converge. Model companies increasingly resemble software vendors, selling governance, compliance, privacy, uptime, SLAs, and account support as much as raw intelligence.
- •Temporary model preference differences exist (e.g., developer tastes)
- •In ~5 years, most enterprises won’t perceive large quality gaps
- •“Pure model company” is rare; most are software businesses with models inside
- •Enterprise value-add: security, compliance, governance, SLAs, support
- 8:18 – 10:27
Intelligence becomes a commodity: the startup playbook when tokens go to ~zero
With open source and competition (Meta, DeepSeek) pushing costs down, Aaron argues token pricing will trend toward underlying compute costs. Startups still win by building robust software products—vertical AI, horizontal orchestration, and agentic workflows—using intelligence as a cheap input.
- •Open source creates a permanent pricing counterbalance; intelligence cost trends downward
- •Tokens converge toward GPU/bare-metal economics plus thin margin
- •Opportunity shifts to vertical AI and workflow “stitching” layers
- •AI company fundamentals still look like classic software execution
- 10:27 – 12:47
Reasoning models and agents: higher intelligence unlocks mission-critical workflows
Aaron describes internal benchmarking: reasoning models can be better in some areas and oddly worse in others, but net intelligence gains expand enterprise use cases. Adoption is still early—assistants are nascent, and true “agents” are barely deployed—especially in regulated industries like banking.
- •Reasoning models: improvements are real but not uniformly better across tasks
- •Rising intelligence enables chaining agents and more agentic workflows
- •Mission-critical workflows adopt later due to nondeterminism and risk
- •Enterprise adoption estimate: ~10% assistants, ~1% agents (or less)
- 12:47 – 15:27
Who in the Fortune 500 cares about the underlying model (and why pricing converges)
Technical stakeholders (CTOs, AI leads, IT practitioners) track model differences, while line-of-business execs and end users typically do not. Because models are increasingly fungible, frontier providers must competitively match prices—similar to cloud storage and compute markets.
- •CTO/AI/IT teams care about model choice; business leaders mostly don’t
- •Models are “directionally fungible,” enabling easy switching for many use cases
- •Frontier pricing pressure comes from slightly-inferior cheaper alternatives
- •Analogy: hyperscaler storage pricing convergence; lock-in shifts to workflows/data
- 15:27 – 20:49
From pilot skepticism to usage-based reality: new AI pricing and margin dynamics
They discuss how successful AI companies can swap models underneath without customers noticing, improving margins as token costs fall. The conversation turns to pricing experimentation—per outcome, per qualified outcome, or usage-based—especially when AI replaces elastic “service-like” work (BPO analogs).
- •Enterprises often don’t notice model swaps; they enforce outcome/accuracy SLAs
- •Token deflation can drive massive gross margin expansion
- •Pricing models: per lead, per qualified lead, per outcome, or per resource usage
- •AI enables elastic scaling that historically required hiring and operational buildup
- 20:49 – 27:04
What enterprise execs think right now: AI-first becomes competitive necessity
Aaron contrasts early cloud-era skepticism with today’s AI urgency, citing Goldman’s AI-assisted S-1 prep as a signal of aggressive adoption. Leaders now see AI as affecting competitiveness and talent acquisition: AI-native workers won’t join firms with archaic tooling.
- •Example: Goldman using AI to accelerate S-1 drafting work
- •Unlike cloud (efficiency), AI is perceived as a competitive differentiator
- •AI-native workforce raises baseline expectations for tools and productivity
- •Firms fear losing talent and market share if they’re not AI-first
- 27:04 – 28:17
How Box is going AI-first internally (engineering, support, knowledge management)
Aaron outlines Box’s internal AI investments: coding productivity, improved customer support workflows, and AI-driven knowledge/HR Q&A over internal documents. A key shift is that documents become queryable knowledge rather than static content to be read manually.
- •Rolling out AI coding tools to increase engineering output
- •Applying AI to customer tickets to improve response rate and quality
- •Internal knowledge management: employees ask questions over HR/benefits/docs
- •Documents shift from passive storage to interactive, interrogable knowledge
- 28:17 – 34:48
Build vs buy in the enterprise: the “context vs core” framework
They apply Geoff Moore’s context/core idea: companies should buy “context” (necessary but non-differentiating systems like HR/ERP/CRM) and focus internal building on “core” differentiators (e.g., wealth personalization, drug discovery IP). This explains why many AI solutions will be purchased externally, even as select high-value systems remain homegrown.
- •Context: box-checking systems should be bought, not reinvented
- •Core: differentiating capabilities should be built and owned
- •Misallocating effort (building context internally) wastes time and focus
- •Most knowledge-worker AI in 2030 likely arrives via ISVs, with critical exceptions
- 34:48 – 37:18
Enterprise security and on-prem AI: trust is rising, segmentation persists
They revisit early enterprise fears about sending data to hosted models and how comfort is increasing as vendors mature on privacy, compliance, and controls. Some industries will still maintain on-prem “enclaves,” but cloud-era patterns of trust-building and standardization are repeating in AI.
- •Regulated sectors may keep on-prem or enclave deployments
- •Trust increases as OpenAI/Anthropic invest in enterprise-grade controls
- •Security/compliance maturity correlates with broader adoption
- •Cloud adoption created organizational muscle memory for evaluating hosted services
- 37:18 – 45:24
From on-prem → cloud → AI: compounding prerequisites and a bigger software TAM
Aaron argues AI’s enterprise adoption required prior shifts: SaaS, cloud infrastructure, and consumer tech normalization. He draws a parallel to SaaS-era TAM expansion (e.g., Salesforce) and predicts AI will expand software spend by enabling software to do the work—not just provide tools—often creating entirely new demand rather than merely replacing labor.
- •AI adoption depended on cloud/SaaS and consumer familiarity with modern UX
- •SaaS expanded TAM by slashing deployment complexity and cost barriers
- •AI expands TAM by making software perform work, unlocking new categories of spend
- •Demand is frequently additive: companies automate things they previously didn’t do
- 45:24 – 49:28
Microeconomics of automation: reinvestment, competition, and the abundance upside
They close on why automation gains tend to be reinvested in competitive markets rather than simply banked as profit, leading to more products, growth, and jobs over time. Aaron returns to the abundance vision—AI can broaden access to intelligence, improve services like healthcare and education, and raise welfare, contingent on broader societal constraints (e.g., regulation and housing).
- •Efficiency gains are reinvested to compete, grow, and improve products
- •Zero-sum “payroll to software” framing misses demand expansion dynamics
- •Consumers benefit as quality and access improve (education, healthcare, services)
- •Abundance depends on broader policy/regulatory environment enabling surplus
