Fund III Investment Thesis

The central question for AI investment: If AI delivers 66% individual productivity gains and 10x+ team output potential, why do 74% of companies fail to capture tangible value?

The answer isn't technical debt or legacy systems—though those matter. The answer is human. AI adoption follows a predictable progression, and each stage is progressively more threatening to the people asked to implement it. This structural conflict of interest explains why AI transformation fails in large organizations—and why AI value concentrates in private markets.

The Opportunity: Why Gaming, Why Now

Gaming is a $250+ billion annual market—larger than film and music combined—and the fastest-growing segment of entertainment. But gaming's significance for AI investment isn't just market size.

Gaming is where AI agents become viable first. The structural reasons:

• Zero real-world risk: Unlike healthcare (where AI errors cost lives) or finance (where errors trigger regulatory action), gaming offers a safe sandbox. Agents can experiment, fail, and iterate without consequence.
• Instant feedback loops: Games provide immediate, measurable outcomes—retention rates, ARPU, engagement metrics. Agents learn faster when feedback is unambiguous.
• Fully digital workflows: Game development is already fully digital. No physical supply chains, no regulatory bodies, no compliance hurdles to agent deployment.
• Proven path to generalization: Google DeepMind trains AI systems in gaming environments because games transfer to real-world applications—a capability years away in other industries.

The adoption is already happening: 73% of game studios are using AI (a16z, 2024). 84% of that adoption is concentrated in teams under 20 people—the AI-native startup profile. This document explains why that value concentrates in private markets, and how Fund III captures it.

The Value Multiplier Framework: Context → Skills → Agents

AI capability matures through three distinct stages. Each stage represents an order-of-magnitude increase in value—but also an order-of-magnitude increase in organizational threat.

Stage 1: Context (1x Value)

What it is: Structured access to institutional knowledge. AI-powered search, knowledge bases, chatbots that answer "what happened?" or "what did we learn?"

Example: "What were the results of our Indonesia beta test in 2019?" — The AI retrieves historical data; a human still interprets and decides what to do.

Threat level: Low — This is just better search. Non-threatening. Everyone adopts.

Stage 2: Skills (10x Value)

What it is: Packaged, reusable capabilities with defined inputs and outputs. Not just information, but how to do something. Transferable expertise.

Example: A "Beta Test in Southeast Asia" skill that takes a game build, budget, and timeline as inputs and produces market selection, UA configuration, analytics setup, and a go/no-go recommendation as outputs.

Threat level: Medium — Skills encode a way of doing things. Adopting them means admitting your way isn't the best way. Teams resist: "We already know how to do this" / "Doesn't fit our process."

Stage 3: Agents (100x Capability)

What it is: Autonomous systems that orchestrate multiple skills to achieve goals. Goal-oriented, self-directing within constraints, monitoring and adjusting without human intervention for each decision.

Example: A UA Campaign Manager agent that targets $X monthly bookings at Y% ROAS, autonomously adjusting bids, reallocating budget, pausing underperforming creatives, and escalating only when thresholds are breached.

Threat level: Existential — Agents do the work. If an agent manages campaigns, adjusts bids, rotates creatives... what does the campaign manager do?

Why the Progression Dies: The Human Problem

The structural and technical barriers to AI adoption are well-documented: technical debt (up to 40% of enterprise technology estates), legacy systems (70% of Fortune 500 systems built 20+ years ago), integration complexity (60% cite this as primary barrier to agentic AI). But these explain why AI is hard—not why it fails.

The Threat Escalation Table

Evidence from the Field

This pattern plays out consistently across industries. In gaming specifically, the data reveals a deeply polarized workforce:

Large/old studios are optimized for what they do. EA's 2024 AI pilot sparked a 200-employee petition and walkouts. Ubisoft's AI integration for Assassin's Creed faced "fierce resistance" requiring months of town halls. CD Projekt Red paused AI expansion after 55% of staff felt "unsettled." Asking these organizations to transform is asking them to become different organizations—and that usually destroys morale and fails anyway.

Why This Matters: The Structural Divide

The conflict of interest creates a structural divide between who can capture AI value and who will:

The Math of Team Economics

When AI is native to operations—not retrofitted onto legacy processes—the team economics are transformational:

• Individual productivity: 66% increase from AI tools (Nielsen Norman Group, 2024)
• Team output multiplier: AI-native teams achieve 4-5x productivity—a 20-person team delivers what traditionally required 80-100
• Revenue per employee: AI-exposed firms grew 27%—3x faster than less-exposed companies (2018-2024)
• Org structure: AI-native startups maintain flat, fluid structures because AI handles standardized work

Traditional headcount scaling becomes a competitive disadvantage when capital-efficient alternatives exist.

The Exit Paradox: How Acquirers Integrate What They Can't Build

A natural objection: if incumbents can't adopt AI due to cultural conflict, why would they successfully integrate AI-native acquisitions? This apparent paradox has a 40-year answer.

The Innovation Conveyor Belt (1985-2025)

Gaming has operated a "conveyor belt" model for four decades: private markets innovate → public companies acquire → acquired studios operate autonomously. This structural inability to innovate internally is why Microsoft (Mojang, Activision), Sony (Bungie), and Tencent (Supercell) rely entirely on the M&A conveyor belt to fuel growth. The pattern predates AI—it's how the industry has always worked:

Source: Wikipedia list of largest video game M&A; SuperJoost "Three decades of games industry consolidation" (2024)

Why Autonomous Operation Works

Sophisticated acquirers understand they cannot integrate innovation—they can only acquire and protect it:

The Counter-Example: What Happens When Publics Build Internally

THQ (1991-2013) attempted the opposite strategy—building 11 internal studios for original IP rather than acquiring proven assets:

• uDraw tablet experiment: $100M+ loss on in-house hardware innovation
• Stock collapse: 86% decline before bankruptcy
• Assets auctioned: Relic sold to Sega for $26.6M at distressed prices

THQ's failure wasn't bad management—it was attempting to do internally what the industry structure rewards doing externally. Public companies are structurally penalized for risky in-house R&D.

J.P. Morgan's Private Markets Thesis

This dichotomy between public stagnation and private innovation is not unique to gaming—it is the core of J.P. Morgan's 2025 thesis. Their Private Markets Outlook (November 2024) identifies private markets as a key investment theme—alongside positioning for the AI revolution. The connection is direct:

The question isn't "can big companies use AI?" They can, and they will—at the Context layer. The question is "can they transform how they operate using AI?" The structural answer is: rarely, slowly, and at enormous cost.

The AI-Native Formula

What distinguishes a company that is AI-native from one that uses AI:

Missing any factor significantly reduces probability of success. An AI-bullish team without domain expertise builds impressive demos that don't ship. Domain experts without AI-native processes will always find reasons why "our situation is different."

Fund III: Investing at the 100x Layer

Fund III's thesis positions at the top of the AI value curve—backing AI-native gaming studios that capture the structural advantages outlined above.

The Economics: Margin Transformation

AI-native studios don't just produce more—they fundamentally restructure the margin profile:

Capital efficiency attracts acquirers. When revenue growth decouples from headcount growth, AI-native studios become premium acquisition targets—lower integration costs, immediate margin accretion, and transferable AI capabilities that acquirers cannot build internally.

*Traditional gaming content trades at 1.8-2.8x EV/Revenue (Drake Star 2024). AI-native content studios achieve 40-50% margins vs traditional 20-25% through ~70% cost efficiency gains. At equivalent EBITDA multiples, this margin improvement implies 3-5x EV/Revenue (Transcend analysis). Note: Platform fees historically 30%, but declining via web stores and regulatory pressure (Epic v. Apple, EU DMA)—an additional tailwind not modeled in base case.

Revenue upside (not modeled): AI also enables revenue improvements through personalized content, dynamic difficulty, and optimized live ops. We model only cost savings to avoid compounding speculative assumptions—meaning higher company profits can accelerate growth (more UA investment) or flow to earlier distributions.

Why Not Big Tech?

A natural question: if agents are so valuable, why won't Microsoft, Google, or Meta dominate?

• They build infrastructure, not applications: Big Tech's AI strategy is to sell the shovel—cloud compute, foundation models, developer tools. They capture value horizontally, not vertically.
• Platform revenue conflicts: Microsoft can't build a game studio workflow that obsoletes Azure's gaming customers. Google can't ship an agent that competes with studios using Vertex AI.
• Speed mismatch: DeepMind's Genie 2 generates 3D game worlds at 1 FPS with sub-minute stability. R&D timelines measured in years. AI-native startups ship in weeks.
• Agility vs. scale: Value accrues to the application layer that fine-tunes models for specific game loops—a specialization Big Tech structurally cannot pursue. In the AI era, the game studio IS the application layer—content and generation engine fused into one.

Compute Cost Collapse Enables Unit Economics

The objection "agents are expensive to run" is increasingly obsolete:

• 280x cost reduction in GPT-3.5-level inference (Nov 2022 → Oct 2024) per Stanford AI Index
• 40x annual decline in cost to achieve GPT-4 performance on complex tasks (Epoch AI)
• 1,000x reduction in GPT-3 pricing over three years ($60/M tokens → $0.06)
• 88% spot price decline for H100 GPUs in some regions during 2025

The consensus among AI researchers: inference costs are declining 10x annually. By 2027, compute cost for most agent workloads will be negligible—a rounding error on the P&L.

The Investment Strategy

Why Not Invest in AI Transformation of Incumbents?

The structural conflict of interest makes incumbent transformation a poor risk/reward:

• Large studios will not adopt agents that threaten their teams
• Forcing transformation destroys morale and usually fails
• Even successful transformation takes 3-5 years and enormous capital
• AI-native startups will reach the same capability in 18-24 months at 10% of the cost

The better bet is backing AI-native teams who build without the conflict of interest.