CHAPTERS
- 0:01 – 1:32
Why a16z published this analysis: AI demand is surging and the cycle is still early
David George frames the piece as a new, data-driven format drawn from a16z’s internal “exhaust” of market observations. He lays out the core thesis: AI demand and company quality are extraordinary, the supply side is mostly healthy but worth monitoring, and private markets are where the most exciting action is. He emphasizes the biggest takeaway—this is the very beginning of a 10–15 year product cycle.
- •Motivation for publishing: consolidate internal analysis into a public, coherent view
- •Demand side is “crazy”: strong uptake and compelling products
- •Supply side looks healthy but shows early signs of stretch
- •Most exciting AI activity is happening in private markets
- •AI product cycle is in its earliest innings (multi-decade arc)
- 1:32 – 2:16
a16z’s vantage point: investing across stages and building an internal growth dataset
David explains a16z’s investment activity across private stages, emphasizing growth-stage exposure to AI infrastructure and applications. He then pivots to how the team collects large volumes of data by seeing most growth-stage companies as prospects or portfolio. This dataset underpins the revenue, efficiency, and retention conclusions that follow.
- •a16z invests across all private stages; growth focus has been AI/infra/apps
- •Firm’s market visibility enables a unique comparative dataset
- •Dedicated data analysis team supports the growth investing process
- •Transition from “portfolio showcase” to market-wide benchmarking
- 2:16 – 3:16
2025 revenue acceleration: the reversal of the slowdown and the rise of extreme outliers
Using internal benchmarking, David argues 2025 marked a broad re-acceleration in revenue growth after the post-rate-hike slowdown. The acceleration is most pronounced among the top decile/quartile performers. He highlights how AI outliers are hitting scale (e.g., $100M revenue) much faster than prior SaaS cohorts.
- •2022–2024 slowdown gives way to 2025 acceleration
- •Outliers drive the biggest step-function improvement
- •Fastest AI companies reach $100M revenue far faster than SaaS-era leaders
- •Growth uplift appears across deciles, not only the very top
- 3:16 – 4:16
Why AI outgrows SaaS while spending less: demand-led growth over sales-and-marketing burn
David’s key claim is that AI leaders are not growing because they outspend on go-to-market; they grow because end-customer pull is unusually strong. He contrasts AI vs non-AI growth rates and calls out the eye-catching 693% YoY growth for top AI performers. The broader implication is that distribution efficiency is improving because product value is driving adoption.
- •AI companies grow ~2.5x+ faster than non-AI companies in the dataset
- •Top AI performers show ~693% YoY growth (validated against anecdotes)
- •Faster scaling is not driven by higher S&M spend—often the opposite
- •Demand strength and product compulsion explain much of the growth gap
- 4:16 – 5:17
Margins and what they signal: inference cost as a ‘badge of honor’
David notes AI companies often show slightly lower gross margins, driven by inference costs. Rather than treating this as purely negative, he argues it can be a positive signal: customers are actually using AI features. He also expresses a belief that inference costs will decline over time, improving margins structurally.
- •AI gross margins are currently a bit worse than non-AI peers
- •Low margins can indicate real AI feature usage (high inference cost)
- •Skepticism when margins are “too high” (AI may not be the real driver)
- •Expectation that inference costs trend down over time
- 5:17 – 6:42
ARR per employee (ARR/FTE): what the efficiency numbers do—and don’t—mean yet
David introduces ARR per FTE as a holistic efficiency metric capturing S&M, R&D, and overhead. He cites leading AI companies at ~$500K–$1M ARR per FTE versus a prior SaaS rule-of-thumb around ~$400K. He cautions that much of this may reflect extraordinary demand and a best-of-the-best dataset rather than fully reimagined org design—though early signs are emerging.
- •ARR/FTE captures total operating efficiency, not just GTM efficiency
- •Best AI companies: ~$500K–$1M ARR per FTE vs ~$400K SaaS benchmark
- •Efficiency partially explained by rapid demand-led scaling
- •Not yet proof that most companies have reinvented how they operate
- 6:42 – 11:18
Adapt or die for pre-AI incumbents: rebuilding products and organizations around AI tools
In response to Jen’s question about non–AI-native companies, David argues incumbents must aggressively “disrupt themselves” both in product and operations. He describes rapid improvements in coding tools (Codex/Cursor/others) enabling 10–20x faster development in at least one anecdote, forcing org redesign questions about product, engineering, and design. The mindset shift is summarized by the ‘electricity vs. blood’ framing—use AI where possible instead of human labor.
- •Pre-AI companies must reimagine products natively with AI, not bolt-on chatbots
- •Operational transformation: roll out best coding models/tools across functions
- •Anecdote: 10–20x faster rebuild with modern coding assistants, influencing org structure
- •Cultural reframing: “electricity vs blood” for task execution
- •Risk: laggards will move materially slower than AI-forward peers
- 11:18 – 15:32
Business model evolution: from seats to consumption to outcome-based pricing
David places AI disruption in a broader sequence of enterprise software monetization shifts: licenses → SaaS seats → consumption/usage → outcome-based pricing. He argues the most disruptive scenario is when technology/product change coincides with business model change. Outcome-based pricing is still hard to implement broadly today, but customer support is a promising early area because outcomes can be measured more objectively.
- •Business model spectrum: license → seat-based SaaS → consumption → outcome-based
- •Most disruptive when product + business model shift together
- •Consumption models already reshaping many cloud-adjacent categories
- •Outcome-based pricing likely starts in support/success where resolution is measurable
- 15:32 – 17:37
What sustains AI revenue: retention, renewals, and deep engagement signals
David addresses skepticism about whether hypergrowth is fleeting by describing how a16z diligence focuses on retention, renewals, and behavioral usage depth. He uses Harvey as an example where improved product and better reasoning models lead users to spend roughly double the time in the product. The core point: sustainable AI businesses show durable engagement, not just initial hype-driven expansion.
- •Primary durability checks: retention, renewals, time-in-product, activity patterns
- •Harvey: reasoning improvements align with legal workflows; time-in-product ~doubles
- •AI may not reduce headcount immediately but can raise professional productivity
- •Engagement data used as a proxy for value realization and renewal likelihood
- 17:37 – 21:37
Portfolio case studies: Abridge, ElevenLabs, Navan, and Flock’s measurable ROI
David quickly highlights multiple companies to illustrate sustainable growth and tangible ROI. Abridge shows stable or rising engagement even as user count scales, reducing fears of dilution. Navan demonstrates AI-driven operational leverage—AI handles ~50% of travel-related interactions and gross margin expands ~20 points—while Flock’s ROI is framed as public safety outcomes (crimes solved and clearance rate improvements).
- •Abridge: engagement holds/expands even as users scale (durability signal)
- •ElevenLabs: staggering usage growth; voice as a core interface for many AI tools
- •Navan: AI handles ~50% of complex travel workflows; ~20pt gross margin expansion
- •Flock: outcome-driven ROI (crime solving; +~10% clearance where deployed)
- 21:37 – 25:09
Fortune 500 AI adoption: executives talk urgency, but change management is the bottleneck
David contrasts CEO rhetoric (‘we must become AI companies’) with slower real-world adoption, attributing the gap primarily to change management. He notes that some functions (coding, customer support) are easier entry points, and cites concrete productivity/cost wins from non-AI-native companies like Chime and Rocket Mortgage. The next 12 months are positioned as a critical period when many more measurable examples should emerge.
- •C-suite intent is high; execution lags due to change management complexity
- •Easy first wins: coding tools and customer support automation
- •Examples: Chime support costs down ~60%; Rocket saved 1.1M underwriting hours (~$40M run rate)
- •Prediction: widening productivity gap between adopters and laggards over 5 years
- 25:09 – 28:12
Public markets context: AI is driving returns, but fundamentals look sound
David argues AI winners account for the bulk of recent S&P 500 returns, yet public market pricing is anchored in earnings and earnings growth rather than speculative, loss-making growth. While multiples are elevated, he says they are far from dot-com extremes, and adjusted for margins they look more reasonable. He also underscores that long-term returns tend to follow growth, and markets are rewarding high-growth, especially when paired with high margins.
- •AI winners drive ~80% of S&P 500 returns (as framed here)
- •Multiples elevated but not dot-com-like; pricing tied to earnings quality
- •Leading tech companies show long-run margin improvement and durability
- •High growth is rewarded; high growth + high margin receives the largest premium
- 28:12 – 35:52
AI infrastructure buildout: massive CapEx, emerging debt signals, and ‘no dark GPU’
David describes the scale and concentration risk of AI CapEx while arguing it differs from prior bubbles because it’s largely financed by extremely profitable hyperscalers. He flags the growing role of debt—especially in parts of the ecosystem like data centers and specific players (e.g., Oracle’s aggressive buildout)—as an area to monitor. On utilization, he emphasizes strong demand: GPUs get used immediately (‘no dark GPU’), and older chips still show robust secondary-market pricing and high utilization.
- •CapEx buildout is huge; concentrated investment is inherently risky
- •Unlike dot-com, spending is largely supported by strong cash flows
- •Debt is entering the picture (private credit/data center financing); monitor counterparties
- •Utilization thesis: ‘no dark GPU’; demand exceeds supply per hyperscalers
- •Depreciation worries tempered by strong utilization of older TPUs/GPUs and stable rental prices
- 35:52 – 41:08
How big can AI get? Revenue trajectory, trillion-dollar math, and ‘heard it on the street’ tracking
David quantifies the pace: generative AI revenue has scaled rapidly, and OpenAI + Anthropic’s net-new revenue already approaches a significant fraction of all public software’s net-new revenue. He shares rough market-sizing math suggesting ~$1T annual AI revenue by 2030 might be needed for a 10% hurdle on cumulative hyperscaler CapEx, while noting the payoff horizon likely extends beyond 2030. He also estimates today’s AI revenue at roughly ~$50B and describes an internal system that tracks AI mentions and themes from public company earnings calls to inform portfolio CEOs.
- •AI app/category revenue scaling faster than prior tech categories
- •OpenAI + Anthropic net-new revenue is already near half of public software’s net-new (2025 framing)
- •CapEx payback napkin math: ~$1T annual AI revenue by 2030 for ~10% hurdle; longer horizon may matter
- •Current AI revenue estimate: ~ $50B (imprecise; big tech attribution is hard)
- •a16z tracks AI discussion across public earnings calls and shares insights with CEOs
- 41:08 – 45:06
Private markets and power laws: value concentration, staying private longer, and the public-private tradeoff
David argues private markets are now a mature asset class as companies stay private longer and most $100M+ revenue companies remain private. He highlights power-law concentration: a small number of unicorns represent a disproportionate share of total value, and concentration has increased since 2020. He also discusses ‘volatility laundering’—the debate over private markets smoothing price discovery—while noting merits on both sides and expecting notable IPO activity over the next ~18 months.
- •Most $100M+ revenue companies are private (~86% cited) as public company counts decline
- •Unicorn value concentrates: top 10 represent ~40% of $5.5T total (as framed)
- •Public market disruption is accelerating: shorter average S&P 500 tenure over decades
- •‘Volatility laundering’ debate: private markets can reduce perceived volatility and aid talent retention
- •Expectation of major long-private companies going public in the next ~18 months
- 45:06 – 47:32
Databricks as a pre-AI-to-AI transition blueprint: leadership, product iteration, and AI-native customers
Answering a question on Databricks, David attributes successful transition to leadership from the top and deep technical-commercial fluency. He notes Databricks’ architecture (data lake/warehouse foundation) positions it well for AI workloads and that the company has iterated aggressively into new AI products like Agent Bricks. He also emphasizes customer quality: serving cutting-edge AI-native companies validates the tech and creates durable growth as those customers scale.
- •AI transition requires CEO-led push and strong technical + commercial instinct
- •Databricks’ data platform is a natural substrate for AI workloads
- •Aggressive new AI product iteration (e.g., Agent Bricks)
- •AI-native customers validate capability and provide growth tailwinds
- •Preference for ‘modern technologist’ customers who evaluate and choose best-in-class tools
