CHAPTERS
Why a16z published this data-heavy AI markets piece (and what stood out)
David George sets context for a new “put it on paper” style of analysis built from a16z’s internal growth-stage dealflow data. He previews the core conclusions: extraordinary demand, generally healthy supply with some stretched areas to watch, and the biggest action happening in private markets. He emphasizes how early the AI product cycle still is—more like the beginning of a decade-plus wave than the middle.
- •Genesis of the report: packaging internal analysis “exhaust” into a public artifact
- •Demand side: uptake and company quality look unusually strong
- •Supply side: broadly healthy, but early signs of stretch worth monitoring
- •AI is most exciting in private markets right now
- •AI cycle is still early in a 10–15 year product-cycle framework
a16z’s growth-stage vantage point: what the dataset represents
George explains why his team can observe market-wide patterns: they evaluate a large share of growth-stage companies as investors or prospects. This creates a high-frequency, real-world dataset on growth, margins, and operating efficiency. He briefly references a16z’s breadth across stages and verticals as the backdrop for the analysis.
- •Growth team sees a wide swath of late-stage private companies
- •Internal data analysis team aggregates portfolio + prospect metrics
- •Activity spans stages; recent emphasis on AI infrastructure and apps
- •The dataset is skewed toward high-performing companies (important for interpretation)
2025 revenue re-acceleration: AI outliers growing at “unicorn is real” speeds
The conversation turns to the headline numbers: 2025 marks a clear reversal from the post-rate-hike slowdown. The fastest AI companies are hitting $100M revenue far faster than prior SaaS-era comps, with extreme outliers showing ~693% YoY growth. George stresses the growth is driven by product pull and demand—not by heavier sales & marketing spend.
- •2025 shows accelerated revenue growth after 2022–2024 slowdown
- •Top AI performers exhibiting ~693% YoY growth (validated against anecdotes)
- •Fastest AI companies reach $100M revenue faster than SaaS predecessors
- •AI cohort grows ~2.5x+ faster than non-AI in the dataset
- •Growth is not coming from higher S&M spend; often the opposite
Margins in AI: lower gross margin as a usage signal (inference costs as a ‘badge of honor’)
George explains why AI companies often show worse gross margins than traditional SaaS: inference costs are real and reflect actual feature usage. Counterintuitively, very high gross margins can be a warning sign that “AI” isn’t what customers are truly buying or using. The thesis is that inference costs should decline over time, improving margins.
- •AI gross margins trend lower vs. classic SaaS due to inference costs
- •Lower margins can indicate customers are truly using AI features
- •High AI gross margins can be a red flag (AI not the value driver)
- •Expectation: inference costs fall over time, expanding margins later
ARR per employee and the efficiency debate: demand vs. AI-driven org redesign
They introduce ARR/FTE as a holistic efficiency metric that captures overhead and R&D—not just sales efficiency. Best AI companies show ~$500K–$1M ARR per employee versus a prior SaaS rule-of-thumb around ~$400K. George argues much of today’s efficiency is driven by intense demand and post-2021 discipline, with ‘run the company totally differently’ still early.
- •ARR/FTE used as a broader efficiency measure than S&M metrics alone
- •Top AI companies: ~$500K–$1M ARR per FTE vs. ~$400K SaaS-era benchmark
- •Current efficiency partly reflects extreme demand and faster scaling
- •Also influenced by broader tech belt-tightening after the 2021 era
- •Full organizational reimagination is starting but not yet widespread
‘Adapt or die’ for pre-AI companies: product reinvention + AI-enabled operations
Responding to how non–AI-native companies compete, George frames a two-sided mandate: rebuild products to be AI-native (not just bolt-on chat) and rebuild internal operations with the latest models/tools. He highlights rapid progress in coding tools and describes founders seeing 10–20x faster build velocity when teams adopt modern AI coding stacks with “unlimited budget.” The shift forces new org design questions about boundaries between product, engineering, and design.
- •Pre-AI companies must aggressively self-disrupt: adapt product and operations
- •AI-native product means reimagined workflows, not a chatbot add-on
- •Coding is the leading edge of adoption; biggest leaps happening recently
- •Anecdote: 2 engineers rebuilding with AI tools, seeing 10–20x speed gains
- •Org design implications: rethink roles, team structure, and workflows
From seats to usage to outcomes: the next business-model disruption
George lays out a business-model evolution: perpetual licenses → SaaS/seat-based subscriptions → consumption/usage-based pricing → outcome-based pricing. He argues the most dangerous scenario for incumbents is simultaneous product and business-model disruption. Outcome-based pricing is still limited today (customer support is most measurable), but could expand as models improve and outcome measurement becomes feasible.
- •Business-model spectrum: licenses → seat-based SaaS → consumption → outcomes
- •Biggest incumbent risk: tech/product shift + pricing model shift together
- •Consumption pricing already disruptive for many seat-based software vendors
- •Outcome-based pricing currently most viable in customer support
- •Model capability + measurability will determine how far outcomes-based goes
Sustainable growth proof: retention, renewals, and deep engagement signals
To address whether AI revenue is “fleeting,” George describes how a16z diligences sustainability: retention/renewals, engagement frequency, and time-in-product. He uses portfolio examples to illustrate: Harvey’s time-in-product roughly doubling as reasoning improves; Abridge maintaining (and slightly growing) engagement even while user counts expand; ElevenLabs showing staggering usage growth and strong operational efficiency. Nevan demonstrates margin impact from automation, and Flock highlights ROI framed as crime reduction.
- •Sustainability checks: revenue retention, renewals, usage frequency, time-in-product
- •Harvey: improved product + reasoning models → ~2x time spent in product
- •Abridge: user growth with steady/high engagement (no drop-off in marginal users)
- •ElevenLabs: rapid usage growth; voice as a core interface layer
- •Nevan: AI handles ~50% of complex travel interactions; ~20-pt gross margin expansion
- •Flock: ROI framed as solving crime; ~700k crimes solved annually; higher clearance rates
Fortune 500 AI adoption: CEO intent vs. change-management reality
George contrasts what big-company CEOs say—urgent desire to become “AI companies”—with what’s actually happening: slow adoption driven by process inertia and change management. He notes coding and customer support are easier early wins, while broader process transformation requires backend readiness and organizational commitment. Concrete examples (e.g., Chime and Rocket Mortgage) show pockets of real savings and productivity gains, hinting at a coming multi-year divergence between fast and slow adopters.
- •CEOs express urgency; actual transformation lags due to change management
- •Coding and customer support are clearer near-term adoption pathways
- •Broader transformation requires systems/data readiness beyond “chatbot usage”
- •Examples: Chime reports ~60% support cost reduction
- •Rocket Mortgage reports 1.1M hours saved in underwriting; ~$40M run-rate savings
- •Next 12 months likely to produce more visible enterprise case studies
Public markets snapshot: AI leaders driving returns, but ‘dot-com bubble’ signals look limited
George argues AI winners account for a large share of S&P 500 returns, yet fundamentals appear strong. He notes multiples have not reached dot-com extremes and are supported by earnings and earnings growth rather than loss-making growth. He highlights that high-growth, high-margin companies earn a clear valuation premium, reinforcing the idea that growth drives long-term returns.
- •AI winners drive a disproportionate share of S&P 500 returns (~80% cited)
- •Market performance largely supported by earnings and earnings growth
- •Multiples elevated vs. history but not dot-com-like, especially adjusted for margins
- •Investors paying for profits more than speculative growth (vs. 2021 dynamics)
- •High growth (especially with high margin) is rewarded most in valuations
AI infrastructure buildout: CapEx scale, who funds it, and where debt introduces risk
The discussion shifts to the supply side: massive AI CapEx, concentrated among hyperscalers—seen as positive for startups needing capacity. George says fundamentals differ from prior bubbles because the spend is backed by historically profitable businesses, though debt is increasingly entering the picture. He flags monitoring training economics, paybacks for model builders, and counterparty differences (with specific attention to players like Oracle and private credit’s role).
- •CapEx buildout is massive and concentrated; concentration increases systemic risk
- •Unlike dot-com, spend is largely financed by highly profitable incumbents
- •Hyperscalers bearing most CapEx is beneficial for AI ecosystem capacity
- •Debt is entering: cannot fund all forecast CapEx via cash flow alone
- •Counterparty quality varies; monitoring credit signals (e.g., CDS widening)
- •Tracking model training economics and payback periods closely
No ‘dark GPUs’: utilization, depreciation worries, and why demand still exceeds supply
George addresses finance debates about depreciation and hardware obsolescence, arguing older chips still retain strong utilization and secondary-market pricing. He cites disclosures like older TPUs running at full utilization and notes A100/H100 rental prices holding up. The broader takeaway: as tokens get cheaper, consumption rises, and hyperscalers report demand still outstripping supply—suggesting the infrastructure is being used immediately.
- •Key paradox: cheaper tokens → higher consumption
- •Hyperscalers report demand > supply; ‘no dark GPU’ analogy vs. dark fiber
- •Depreciation/obsolescence fears not yet reflected in utilization or pricing
- •Older TPUs reportedly at ~100% utilization; A100/H100 rentals holding up
- •Early adopters stay on models longer; later adopters switch faster (model lifecycle nuance)
How big does AI revenue need to get? Payback math and where we are today
George walks through a simplified payback framework: cumulative hyperscaler AI CapEx near ~$5T by 2030 could require ~ $1T annual AI revenue by 2030 to clear a ~10% hurdle rate, roughly ~1% of global GDP. He cautions against limiting the horizon to 2030, expecting payback to extend into 2030–2040. When asked where we are now, he estimates roughly ~$50B in AI revenue today, with growth well above 100% YoY.
- •Napkin math: ~$4.8–$5T CapEx by 2030 implies ~ $1T annual AI revenue for ~10% hurdle
- •$1T annual AI revenue ~1% of global GDP (context framing)
- •Payback likely extends beyond 2030 into 2030–2040
- •Current AI revenue estimate: ~ $50B (imperfect measurement)
- •Big Tech AI revenue attribution is hardest to measure; clouds can obscure reporting
- •AI revenue growth characterized as far above 100% YoY
Private markets and power laws: value concentration, staying private longer, and volatility tradeoffs
George argues private markets are now a “real asset class,” with many $100M+ revenue companies remaining private and the public-company count shrinking over decades. He shows power-law concentration: the largest unicorns represent a large fraction of total unicorn value, and concentration has increased since 2020. He also notes faster disruption (shorter S&P 500 tenure) and discusses the founder debate around staying private to ‘launder volatility’ versus the benefits of going public, expecting a notable wave of IPOs in the next ~18 months.
- •Companies staying private longer; public-company count down significantly over decades
- •Claim: ~86% of $100M+ revenue companies are private (scale of private markets)
- •Unicorn value concentration: top 10 comprise ~40% of ~$5.5T total value
- •Public markets also show power laws; definition of large-cap has expanded since 2019
- •Average S&P 500 tenure has declined ~40% over 50 years (faster disruption)
- •Debate: private-market volatility smoothing vs. benefits of being public
- •Expectation of major late-stage private IPOs over the next ~18 months
Case study: Databricks’ shift from pre-AI to AI-embedded platform strategy
In Q&A, George attributes Databricks’ AI transition to leadership from the top—similar to Shopify—and to a strategic position: housing data in a way that’s well-suited for AI workloads. He highlights aggressive AI product iteration (e.g., Agent Bricks) and the validation of serving cutting-edge AI-native customers. Customer quality is emphasized as a key diligence signal, because sophisticated technologists pick the best tools.
- •Transformation driven by CEO leadership and technical-commercial execution
- •Databricks’ data platform position supports AI workloads naturally
- •Aggressive iteration on AI products (Agent Bricks called out)
- •AI-native companies as customers validate product strength and positioning
- •Customer sophistication used as a key indicator of durable technology advantage
