Uncapped with Jack AltmanDavid George on Growth Investing, AI, and Why the Power Law Is Stronger Than Ever | Ep. 58
CHAPTERS
- 0:00 – 1:04
Why AI outcomes are additive: the answer is “and,” not “or”
David frames a core mental model for AI investing: most debates (open vs. closed, incumbents vs. startups) won’t be zero-sum. Because demand is expanding so quickly, multiple layers of the stack can succeed simultaneously.
- •AI market debates are often framed incorrectly as either/or
- •In AI, multiple approaches and winners can coexist (“and”)
- •The pace and breadth of change makes this an unusually exciting investing moment
- •Keeping up with the technology and market is itself a challenge
- 1:04 – 1:55
Where we are in the AI buildout: trillions of capex, tiny revenue diffusion
They zoom out to supply/demand fundamentals: AI infrastructure buildout is already historically large, yet revenue is still concentrated in a small set of users. This gap between early monetization and eventual adoption is why David expects massive growth ahead.
- •AI infrastructure spending has reached historically high GDP share benchmarks
- •Data center buildout likely continues for many years, reaching multi-trillion spend
- •Current AI revenue is driven mostly by a surprisingly small paying cohort
- •Enterprise diffusion is still very low (<5%) despite huge awareness
- 1:55 – 4:30
The power law in token demand: coders (and top coders) drive the market
David argues the current AI economy is extremely power-law distributed: a small number of coders account for most paid usage. That concentration implies enormous upside if AI penetrates broader knowledge work.
- •~30M coders generate most enterprise AI revenue today
- •Within that, the top ~1M users drive disproportionate spend
- •There are ~1.5B knowledge workers—huge headroom beyond coders
- •Early category wins (coding, writing) are signals, not endpoints
- 4:30 – 6:59
What could stop token growth: adoption friction vs. algorithmic efficiency leaps
Jack pressures the thesis: would added capacity always be consumed? David identifies two plausible “break” scenarios—enterprise impact failing to materialize broadly, or major efficiency breakthroughs reducing compute needs—but he sees both as unlikely in the near term.
- •Two main downside paths: limited enterprise impact beyond coding, or big efficiency breakthroughs
- •Training efficiency improvements are possible, but inference demand could still dominate
- •Reasoning at inference time often increases value (and compute), reinforcing demand
- •Near-term oversupply is unlikely because bringing capacity online is slow and bottlenecked
- 6:59 – 10:42
Sellers vs. buyers of tokens: proving ROI and where value accrues
They discuss whether today’s booming revenues mainly reflect token sellers rather than buyers achieving ROI. David is bullish that productivity gains are real, but notes many enterprises need extensive implementation and “hand-holding” before value is fully realized.
- •Key question: are buyers getting measurable ROI or just consuming tokens?
- •Coding productivity has flipped sentiment among top engineers post-model improvements
- •Non-native enterprises may lag until integration and change management happen
- •Best companies focus AI on building new products (revenue upside) over cost cutting
- 10:42 – 12:51
Frontier labs vs. open source: revealed preference, pricing, and “N-1” maturity
David lays out why frontier models dominate spend today (quality + product harness), while predicting a large future for cheaper N-1/open-source options as usage scales. The market will segment by required value, cost sensitivity, and product integration.
- •Frontier labs capture the vast majority of dollars today; users pay ~10x for quality
- •First-party products + tight harness/model coupling drive superior UX
- •As usage scales 50x, cost optimization will pull adoption toward cheaper models
- •Applications will increasingly abstract model choice based on task value (e.g., support tickets)
- 12:51 – 14:12
Application investing strategy: the coding blast radius, horizontals, and vertical moats
David explains how a16z thinks about app opportunities given labs building first-party products. He expects labs to prioritize coding and broad horizontals, leaving many verticals to specialized companies where product detail and go-to-market execution are decisive.
- •Labs focus first-party on coding (and nearby workflows) plus broad knowledge-worker tools
- •Vertical apps can win where details matter and enterprise GTM is required
- •Moats shift toward workflow fit, integrations, distribution, and implementation
- •Despite overlap, multiple app winners can thrive—even inside the “blast radius”
- 14:12 – 20:17
Harvey and the legal category: reasoning models unlock real adoption
Using Harvey as a case study, David argues legal is in a breakout phase similar to coding, just lagged by about a year. Hallucination concerns have faded as reasoning improves and clients increasingly demand AI usage from law firms.
- •Legal adoption is accelerating; often described as ~12 months behind coding
- •Client pull: end customers now require firms to use tools like Harvey
- •Reasoning models are particularly well-suited to legal work (structured, documented)
- •Labs are unlikely to verticalize deeply due to required product nuance and GTM
- 20:17 – 21:46
Consumer AI today: a billion users, but still “skeuomorphic” search replacement
They turn to consumer AI, arguing usage is massive but the dominant behavior is still basic—people using ChatGPT like a better search engine. David contrasts this with kids’ more native behavior and expects product form factors to evolve significantly.
- •Consumer AI has >1B users, but most use cases remain simple Q&A/search substitution
- •“Skeuomorphic mode” dominates: old behaviors in a new interface
- •Kids behave more natively—creating, iterating, and producing artifacts
- •Consumer remains under-discussed relative to enterprise/coding despite huge upside
- 21:46 – 25:54
From reactive to proactive assistants: multimodal interfaces and action-taking agents
David argues the major consumer step-function will come when AI becomes proactive and trustworthy, not just reactive. Improvements in voice and agentic action enable more natural interaction and sustained daily utility, supporting large subscription and ad businesses.
- •Big shift: reactive chat → proactive, action-taking assistant
- •Multimodal (especially voice) is now natural enough for real interaction
- •Early agent moments hint at future workflows that run on users’ behalf
- •Business models likely combine subscriptions with new ad formats native to AI consumption
- 25:54 – 26:48
Autonomy, robotics, and AI in health: the next product waves beyond language
David zooms out to concurrent mega-cycles—AI, autonomous driving, robotics, bio/health, and defense modernization—arguing their combined market cap creation will exceed prior eras. AI dominates attention today, but these parallel waves are large and investable.
- •Product cycles (not just capital cycles) drive long-run returns
- •Previous wave (mobile/social/cloud/SaaS) created ~$25T market cap; next wave may be larger
- •Major upcoming/ongoing waves: autonomy, robotics, AI in health, American dynamism/defense
- •These trends reinforce the “everything works” view across multiple sectors
- 26:48 – 30:48
Autonomous driving is underappreciated: safety, cost per mile, and massive elasticity
They unpack why autonomy could expand markets by 10x or more: it is already materially safer, can undercut rideshare economics, and will reshape personal car ownership economics. Diffusion is still early—Waymo feels ubiquitous in select cities but is tiny nationally.
- •Waymo-scale autonomy shows 10–14x safety improvements over human driving
- •Lowering ride-hail cost per mile could trigger explosive demand in an elastic market
- •Autonomy add-ons for personal vehicles could be worth ~$10k+ per car at scale
- •Diffusion is early: Waymo fleet is <10k vehicles in the US despite high visibility
- 30:48 – 33:27
Robotics: factory-floor first, then a “ChatGPT moment” for physical work
David expects robotics to become a market larger than language once it clicks, but acknowledges home robots are further out. Near-term wins are in constrained environments like factories, where ROI is high and real-world feedback loops can train systems quickly.
- •Robotics likely starts in defined, safe, repetitive industrial settings
- •Factory deployments offer immediate ROI and valuable real-world learning loops
- •Example: robots on manufacturing lines as an early wedge into scale deployment
- •A consumer-level “ChatGPT moment” for robotics could arrive within ~5 years
- 33:27 – 37:16
Benchmark’s new growth fund and why growth investing matters more now
The conversation shifts to capital: why Benchmark launched a growth fund and why late-stage private investing has become more important. David argues a large share of private market returns now happens at C+ stages, especially as companies stay private longer.
- •Private markets are large; skipping growth stages leaves returns on the table
- •a16z analysis: roughly half of private returns historically came from C+ stages
- •As IPO timelines extend, late-stage share of value creation likely increases
- •Power law dynamics make owning winners at growth especially attractive
- 37:16 – 41:57
Product cycle vs. capital cycle: why this decade is a 9–10/10 for opportunity
David distinguishes product-cycle strength from capital-cycle attractiveness. Even if valuations aren’t at rock bottom, the magnitude of product waves (AI, autonomy, robotics, bio/health, defense) makes this period unusually favorable for long-term investing.
- •Ideal investing: strong product cycle + cheap capital cycle; they rarely coincide
- •2021: weak product cycle; today: exceptionally strong product cycle
- •AI is unique: more capital can directly improve the product (scaling laws)
- •Growth investors win by underwriting revenue upside more than margin tweaks
- 41:57 – 49:13
Narrative and “vibes” as competitive advantage: fundraising, hiring, valuation, customers
David argues narrative is now a tangible business lever: it affects valuation, talent, fundraising, and even customer trust. In a world where everyone lives online, founders (and key leaders) who communicate directly can compound momentum.
- •“Vibes matter” because they impact valuation, dilution, hiring/retention, and sometimes sales
- •Higher valuation can be a strategic weapon—especially when capital improves outcomes
- •Examples discussed: Palantir/Karp, Elon-led companies, and visible operator voices
- •Rule of thumb: don’t short messianic founders or products people love—narrative compounds
- 49:13 – 53:11
How big can a16z get: scalability limits, services org, and riding the next waves
They close on venture-firm scale: some investing work is not scalable, but platform support and specialization can extend reach. David expects tech’s share of global market cap to keep rising, with new category leaders still to be created across the next decade’s waves.
- •a16z balances scalable parts (platform/support) with non-scalable parts (investing judgment)
- •Large internal services org is positioned as a way to “go far” organizationally
- •Tech’s share of overall market cap may continue rising; new giants are forming
- •The main challenge is keeping up with the pace while staying in top-tier deals
