Economics of AI
Understanding AI Through the Lens of Economics
Ashish Kulkarni
Takshashila Institution | GCPP Program
February 2026
Upstream
Midstream
Downstream
The Central Tension
$100M+
to train a single frontier model
(GPT-4: $100M+ | Llama 3 hardware: $720M)
$0.0001
per inference query (and falling fast)
280x price drop in 2 years (GPT-3.5 level)
This gap between massive fixed costs and near-zero marginal costs IS the economics of AI.
Key question: Who can afford to produce AI, and who captures the value from consuming it?
The Pipeline: How to Think About AI Economics
Think of AI like oil & gas: a value chain from raw inputs to end consumers
UPSTREAM
What does it cost
to produce AI?
MIDSTREAM
Who makes AI, and
why so few?
DOWNSTREAM
Who benefits,
and how?
UPSTREAM
The Cost Structure of Producing AI
What does it actually cost to build a frontier AI model?
Silicon & Infrastructure: The Foundation
Two Laws That Shape AI
Moore's Law: Transistor density doubles every ~2 years. Performance gets better.
Rock's Law: Fab costs also double every ~4 years. Building chips gets harder.
Result: Only TSMC, Samsung, Intel can make cutting-edge chips. A single fab costs $20B+.
The Gigacluster Buildout
Meta's Llama 3: 24,000 H100 GPUs
$720M in hardware alone
A training cluster needs:
Land + facility + cooling + power substations + network interconnects + chips
Sam Altman's vision: 1GW clusters coming online weekly by 2030
Spine: AI's cost structure starts with massive, irreversible fixed costs in silicon and infrastructure.
Data: The New Binding Constraint
$2.6-3.2B
AI training data market
(2024)
$1.5B
Anthropic's copyright
settlement (2025)
2026-32
Projected exhaustion of
high-quality public text
Three Dynamics Reshaping Data Economics
1
From scraping to licensing Reddit: $60-70M/yr to Google & OpenAI. Books: $2,500-5,000/title. Data has a price now.
2
Quality over quantity Microsoft's Phi-1: 1,000-10,000x more cost-effective through data curation, not scale.
3
Synthetic data as wildcard 60-80% cheaper, but 'model collapse' risk. Hybrid approaches (70-80% real) are optimal.
Spine: Data is the new binding constraint — and it's running out.
Training vs. Inference & The Cost Decay Phenomenon
The Economic Split
| Training | Inference |
Type | One-time CapEx | Continuous OpEx |
Cost share | 10-20% | 80-90% |
Analogy | Building a factory | Running the factory |
AWS estimates inference = 90% of total model cost
The Cost Decay Phenomenon
280x
GPT-3.5-level inference cost drop (2022-2024)
Drivers: Better chips, quantization, distillation, MoE architectures, competition
The DeepSeek moment: Competitive performance at a fraction of the cost. Proved efficiency can rival scale.
But: The Jevons Paradox — cheaper AI drives MORE usage, not less total spend.
Spine: Cost decay is dramatic, but the Jevons Paradox means total spend keeps rising.
Energy & Talent: The Hidden Constraints
Energy: The Recoupling
For decades, economic growth decoupled from electricity demand. AI is recoupling them.
By 2030: US data centers may consume 12% of total electricity — more than steel, chemicals, & aluminum combined.
Water: the 'secret footprint' — a single large data center uses 5M gallons/day, rivaling a small city.
Talent: Extreme Concentration
Top ML researchers: $1M-10M+ in compensation. Acqui-hires (Inflection→Microsoft, Character.AI→Google) are market-clearing.
Data labeling: the invisible workforce. Geographic arbitrage — Kenya, Philippines, India. Scale AI avg contract: $93K.
India's role: A major hub for data labeling and RLHF work, less so (yet) for frontier R&D.
Concept: 'Energy gentrification' — data centers compete with homes, schools, and local businesses for finite grid capacity. When AI arrives in a region, it reshapes who gets access to power.
Spine: Energy and talent concentrate power in very few hands.
MIDSTREAM
Market Structure & Competitive Dynamics
Why does AI look like an oligopoly, and will it stay that way?
The Two-Tier Oligopoly (February 2026)
TIER 1: FRONTIER LABS
Anthropic
Claude Opus 4.6 (Feb 2026). 1M token context, Agent Teams. Safety-first positioning.
OpenAI
GPT-5.3 Codex (Feb 2026). Self-debugging training runs. Product velocity.
Google DeepMind
Gemini Ultra. Integration with Search, YouTube, Workspace. TPU advantage.
TIER 2: CHALLENGERS
xAI (Grok + Twitter data), Mistral (European positioning, data sovereignty), others. Differentiation, not direct competition.
OPEN SOURCE
Meta (Llama): 'commoditize the complement.' Give away the model, profit from the ecosystem. Quality gap is narrowing.
CHINESE PARALLEL
DeepSeek, Baidu, Alibaba, ByteDance. Operating under chip export controls. China's manufacturing strength: AI as commodity.
Spine: The market is a two-tier oligopoly — and the barriers reinforce themselves.
Why Oligopoly Persists: Self-Reinforcing Barriers
Capital
Minimum $1B+ to compete at frontier. Training a single model costs $100M+. Infrastructure requires multi-year commitments.
Talent
~1,000 people worldwide can lead frontier model training. $1-10M compensation. Acqui-hires as market clearing mechanism.
Data & Knowledge
Accumulated training techniques, proprietary datasets, RLHF feedback loops. Institutional learning that can't be bought.
Scale Economies
Fixed costs spread over millions of users. First-mover brand recognition. Network effects from user feedback.
Key question: Are these barriers increasing or decreasing? What would it take for a new entrant?
Spine: Each barrier reinforces the others. Capital buys talent, talent creates data moats, data moats require scale.
Open Source: The Most Powerful Competitive Force
Meta's Strategic Logic: 'Commoditize the Complement'
Concept: If your business depends on X, make X's complement cheap. Meta profits from ads + engagement, not model access. Free Llama strengthens Meta's ecosystem.
Effect on closed labs: Open source creates a price ceiling. If Llama is 'good enough' for free, how much can Claude/GPT charge?
The quality gap: Narrowing, but frontier closed models still lead on reasoning, safety, and specialized tasks. How long does this last?
The sustainability question: Can open source close the gap without billions in capital? Only if backed by a tech giant (Meta) with different incentives.
Analogy
Think of it like GPS navigation:
Google Maps is free. It destroyed the market for standalone GPS devices (Garmin, TomTom).
Why? Because Google profits from knowing where you go, not from selling maps.
Meta's Llama follows the same logic.
Spine: Open source is the most powerful competitive force — not charity, but strategy.
Supply Chains, Vertical Integration & Geopolitics
80-90%
Nvidia's share of
AI GPU market
CUDA = software lock-in
Vertical Integration
Google: Makes chips (TPU) + models + apps
Microsoft-OpenAI: Compute + distribution
Anthropic-Amazon/Google: Cloud partners
Who captures value?
Geopolitics
US-China chip controls are creating parallel AI ecosystems
EU: regulation-first (AI Act)
India: largely a consumer, not a producer (yet)
India's Position in the AI Value Chain
India is a major player in the services layer (data labeling, RLHF workforce, IT services integration), but not yet in frontier model production. India's AI policy focuses on deployment and adoption rather than foundational R&D. The question for India: can it move from being a consumer and service provider to a producer? Or is the downstream position strategically valuable in its own right?
Spine: Vertical integration and geopolitics are reshaping who can compete. India's position is largely downstream.
DOWNSTREAM
Markets, Business Models & Economic Impacts
Who captures the value, and what changes for society?
Pricing & Business Models
The Price Convergence Puzzle
$20/month (Plus) $200/month (Pro)
Both OpenAI and Anthropic converged on the same price points. Coincidence? Or tacit coordination?
Three Revenue Layers
Consumer: Subscriptions ($20-200/mo). Viral growth, high churn.
API: Per-token pricing. Volume discounts. The developer economy.
Enterprise: $100K-millions/yr. Long sales cycles but sticky.
Future: outcome-based pricing? Pay for results, not tokens.
The Consumer Surplus Gap
Someone pays $20/month but extracts $500/month in value (saved time, better work). This $480 gap is consumer surplus — and it's enormous. AI may be the most underpriced technology in history relative to value created.
Spine: Consumer surplus is enormous; labs capture only a fraction of the value created.
The Agentic Economy: AI as Worker, Not Tool
$4.5B (2025) → $98B (2033 projected)
1 billion AI agents predicted by end of 2026
Claude Code / GPT Codex
Terminal-native coding agents. Claude Code: decide → do → verify loops. GPT-5.3 Codex debugged its own training run. These aren't autocomplete — they're autonomous developers.
OpenClaw + Moltbook
A personal AI assistant (150K+ GitHub stars) that runs on your devices, talks through WhatsApp/Slack. Moltbook: a social network for AI agents — 1.7M agent accounts. Early, theatrical, but a preview of where things are heading.
MCP & Infrastructure
Model Context Protocol (donated to Linux Foundation, 97M monthly SDK downloads) standardizes how agents connect to tools. Like HTTP for agent-to-service communication. The plumbing of the agentic economy.
Spine: The agentic economy is the next frontier — AI as autonomous worker, not passive tool.
Labor Markets & Productivity: Task by Task
"You can see the computer age everywhere but in the productivity statistics." — Robert Solow, 1987. Are we repeating this with AI?
Acemoglu-Restrepo: Tasks, Not Jobs
Key insight: AI automates tasks, not entire jobs. Jobs are bundles of tasks — some get automated, others get complemented.
Automated: Data entry, basic analysis, scheduling, routine writing, simple code
Complemented: Strategic thinking, complex problem-solving, interpersonal skills, creativity requiring deep context
Example: A lawyer spends less time on research (automated), more on strategy & client relations (complemented).
The Implementation J-Curve
Productivity dips first (learning, reorganization), then rises. We may be in the dip now.
Historical pace:
Electricity: 40 years to productivity gains
Computers: 20-30 years
Internet: 10-15 years
AI?: TBD — but likely faster
Spine: Labor markets will transform task by task, not job by job.
The Central Policy Question: Who Captures the Gains?
Five Dimensions of AI Inequality
1
Capital vs. Labor AI labs and cloud providers capture producer surplus. Workers face task displacement. Capital's share of income rises.
2
AI-adopters vs. laggards Firms and individuals who use AI effectively pull ahead. Those who don't fall behind. A new digital divide.
3
High-skill vs. low-skill AI currently complements high-skill work but may substitute for it with future models. Middle-skill cognitive workers most at risk.
4
Producers vs. consumers US, China produce frontier AI. India, Europe, others primarily consume it. Value accrues to producers.
5
Present vs. future Current AI investment may crowd out other R&D, infrastructure, education. Opportunity costs are real but unmeasured.
For your students: When AI can do most tasks, where does human comparative advantage lie? This is the question that will define careers, firms, and nations.
The Spine: 12 Takeaways from the Economics of AI
UPSTREAM
1
AI's cost structure is dominated by massive, irreversible fixed costs
2
Data is the new binding constraint — and it's running out
3
Cost decay is dramatic, but the Jevons Paradox means total spend keeps rising
4
Energy and talent concentrate power in very few hands
MIDSTREAM
1
The market is a two-tier oligopoly — and the barriers reinforce themselves
2
Open source is the most powerful competitive force, not a charity
3
Vertical integration and geopolitics are reshaping who can compete
4
India's position is largely downstream — consumption, not production
DOWNSTREAM
1
The agentic economy is the next frontier — AI as worker, not tool
2
Consumer surplus is enormous; labs capture only a fraction of value created
3
Labor markets will transform task by task, not job by job
4
The central policy question: who captures the gains?
Thank You
The economics of AI is the economics of concentration:
massive fixed costs, near-zero marginal costs, and a small number
of firms reshaping how the world works.
Ashish Kulkarni
ashish@econforeverybody.com
A companion reading list is available for deeper exploration.