CHAPTERS
Why tech dominance + staying private longer reshapes growth investing
David George opens with the fund’s founding premise: tech has become an outsized share of global market cap, and top companies are remaining private much longer. That combination expands the investable opportunity set—but also increases the importance of managing liquidity and DPI expectations.
- •US tech companies make up a large share of the most valuable global firms
- •Companies staying private longer creates more private-market ownership opportunity
- •Staying private is a double-edged sword: more time to invest, but DPI/liquidity matters
- •AI is highlighted as the biggest change versus prior growth-fund eras
The AI infrastructure buildout: unprecedented CapEx and who’s funding it
The discussion shifts to the scale of AI infrastructure investment, with big tech driving a massive CapEx run-rate largely directed to data centers and AI. This buildout is framed as a tailwind for application-layer companies because hyperscalers bear much of the upfront burden.
- •Big tech CapEx run-rate is described as roughly $400B annually, largely AI-related
- •Infrastructure groundwork is larger than prior cycles and still likely underestimated
- •Hyperscalers can tolerate overbuild because of strong cash generation
- •This supply buildout enables startups to build on top without funding the entire stack
Input costs collapsing while model capability rises (faster than Moore’s law)
David outlines two compounding trends: model access costs have dropped dramatically while frontier capability improves at a rapid cadence. Together, these dynamics expand what founders can build and compress time-to-market for new AI products.
- •Model access costs cited as down ~99% over ~2 years
- •Frontier capability improving by ~2x every ~7 months
- •Cheaper + better models increase feasible product scope and experimentation
- •AI may become a utility-like layer (compared to electricity/Wi‑Fi) over time
AI’s market opportunity vs software: surplus creation and GDP-scale impact
They compare AI’s potential to prior platform shifts (mobile + cloud), arguing AI affects a much larger share of the economy than traditional software. The conversation emphasizes value creation (consumer/enterprise surplus) and how only a fraction needs to be captured to build enormous companies.
- •Prior mobile+cloud cycle created ~10T in market value; AI expected to be larger
- •Software spend ~1% of GDP vs white-collar payroll ~20% of GDP
- •Rule of thumb: ~90% of value accrues to customers, ~10% to suppliers—but 10% is huge
- •Examples of consumer surplus: iPhone willingness-to-pay vs price; Google monetization vs value
Why this cycle is different from dot-com: supply risk vs demand signals
Erik raises concerns about overbuild reminiscent of early-2000s infrastructure. David argues this cycle differs because stronger companies are building, leverage dynamics differ, and—most importantly—AI demand is visible and global much earlier due to internet/cloud distribution.
- •Key risk factor to watch: leverage and who finances data centers (banks/insurers via private debt)
- •Demand signal example: ChatGPT reaching massive search volume far faster than Google did
- •AI distributes instantly via existing internet + cloud; no new hardware adoption curve required
- •Broad early usage suggests capacity is more likely to be absorbed than in broadband overbuild
Monetization reset: subscriptions, global price discrimination, and AI commerce
They explore how AI may monetize differently than prior consumer internet platforms, with clearer subscription willingness-to-pay and regional pricing. David also predicts new monetization paths for free users, potentially via commerce/affiliate-like mechanisms as AI replaces traditional search journeys.
- •ChatGPT-style products show consumers paying $200–$300/month for high-end tiers
- •Regional pricing example: lower-cost subscription in India vs higher US tiers
- •AI “deep research” changes shopping behavior and could reduce referral traffic to websites
- •Potential freemium monetization via advertising/affiliate or transaction-driven models
Biggest bottlenecks for AI: energy, construction speed, then cooling
Audience questions pivot to physical constraints: powering and building data centers quickly enough, and then dissipating heat. David points to nuclear and localized natural gas as promising near-term answers, while Erik flags cooling as an underappreciated next bottleneck.
- •Energy is a near-term bottleneck; a16z is investing in nuclear-related opportunities
- •Examples: reopening/repowering nuclear assets; data centers sited near plants
- •Natural gas (e.g., West Texas) can support training clusters near supply
- •Operational bottlenecks include construction/logistics; cooling is highlighted as the next constraint
Unit economics and gross margins in AI apps: what matters most
David lays out how they evaluate AI-native businesses amid margin uncertainty from model costs. He emphasizes enduring customer value and acquisition efficiency over current gross margin, assuming competitive model markets continue pushing inference costs down.
- •Top metrics: gross retention (enduring value) and ease/efficiency of customer acquisition
- •Gross margins matter, but they’re more lenient today if input costs are likely to fall
- •Competitive pressure at the model layer (OpenAI, Anthropic, Google) supports continued cost declines
- •They avoid “zero gross margin” businesses but underwrite improving unit economics over time
Cash burn vs pricing: OpenAI economics, R&D intensity, and monetization upside
David addresses concerns about massive cash burn alongside potential consumer pricing pressure. He argues monetization upside is more likely than downside due to low payer penetration, while burn is largely R&D that should be disciplined by competitive and financial incentives.
- •Only a fraction of users pay today; biggest upside is expanding monetization breadth/depth
- •Quantity (users) may be closer to saturation than price (ARPU), which could rise via segmentation
- •Burn is driven largely by research and future model investment
- •Consumer brand stickiness may be stronger than API/developer stickiness (easy switching via API)
Durability of AI application revenue: what’s sticky vs what commoditizes
They differentiate AI apps with embedded workflow/integrations from more experimental, easily-switched tools. Stickiness is tied to how deeply the product is integrated into enterprise processes, brand requirements, and rule-driven workflows.
- •Sticky areas cited: medical scribe/workflow, customer support, financial analysis use cases
- •Stickiness drivers: integrations, rules engines, workflows, enterprise features, brand voice consistency
- •Less sticky: lightweight prototyping and low-end internal tooling experiments
- •Market likely bifurcates between prototype tools and production-grade deployment platforms
Go-to-market time compression: redefining what “great” growth looks like
David explains that benchmarks like time-to-$10M or $100M ARR are shifting because the fastest AI companies scale far quicker than prior SaaS generations. Their investment debates increasingly compare new companies to current AI leaders rather than legacy SaaS comps.
- •Historical SaaS growth heuristics look “modest” vs today’s fastest AI trajectories
- •Context matters: they benchmark against top AI breakouts (e.g., Cursor, Decagon, Abridge, ElevenLabs)
- •Fast scaling increases outcome variance and raises the bar for ‘best-in-class’
- •Staying close to the market (seeing many companies) is essential to calibrate expectations
Pricing models in AI: seat vs usage vs outcome-based (and why it’s early)
They discuss the industry’s search for a new pricing paradigm—potentially charging for completed tasks rather than seats or consumption. David is skeptical that a universal shift happens quickly, noting customers still prefer familiar models and outcome measurement is hard outside select categories.
- •Past transitions: perpetual licenses → seat-based SaaS → usage-based cloud
- •Hypothesis: AI could monetize “tasks humans do,” but measurement is difficult
- •Customer support is the most advanced example of task/outcome monetization so far
- •Competition + measurement challenges may leave much surplus with customers
Companies staying private longer: implications for exits, liquidity, and DPI
David reviews data showing longer private timelines and a growing share of market value in late-stage private companies. They explain how private markets have evolved to provide partial liquidity (tenders) and why the highest-growth tech is increasingly concentrated pre-IPO.
- •Typical time-to-IPO has extended; private market value above $1B has expanded dramatically
- •Only a small fraction of public software/internet companies forecast >25% growth
- •Private markets now mimic public-market features via tenders to help with talent/liquidity
- •Balancing what’s best for the company with the fund’s DPI/liquidity goals
Portfolio construction and a16z advantage: access, collaboration, and category mix
They close on how the growth team works with early-stage partners to gain access and insights, shaping both follow-ons and selective new investments. David outlines expected opportunity concentration in AI apps/infra and American dynamism, plus selective crypto exposure via the dedicated crypto team.
- •Access advantage: early-stage relationships drive entry into later rounds
- •Insight advantage: early-stage proximity improves market/product judgment
- •Portfolio approach: mix of ‘champion’ companies with strong downside protection and selective high-variance elite teams
- •Likely category weighting: AI infra/apps largest; American dynamism next; growing interest in AI-enabled health; crypto via high-conviction collaboration (e.g., stablecoins)
