Key Observations on AI Investing: 2025 Perspective
On investment implications and predictions for 2025
This is the raw note I’ve got to jot down on a long-haul flight. I might turn this into a formal post later - or I might not, because the space is evolving so quickly. Pardon the format.
Nonetheless, I do plan to get back to writing more again - part of my 2025 New Year resolution :)
Lay the Stage:
The current AI wave represents a fundamental shift in how we process information:
Internet era (1980 -2000): Analog to digital transformation
SaaS era (2000-2020): Structured data commercialization (~20% of data)
AI era (2020 - 2040?): Unlocking value from unstructured data (~80% of data)
This suggests we're still in the early innings. Just as "internet company" became redundant, "AI company" will likely fade as AI becomes ubiquitous infrastructure.
Saas vs AI-native companies:
Traditional SaaS:
Often linear customer acquisition costs
Predictable but gradual margin improvements
Revenue scaling often tied to sales headcount
AI-Native Companies:
Proprietary model architecture or meaningful model routing customization - AI-naitive companies can build a moat without needing to train or build their own models.
Data flywheel effects that compound over time, as that’s the key to produce the most relevant and high quality output/ action
Results?
Exponential efficiency gains post-PMF
Margin expansion decoupled from headcount
Capital requirements that decrease with scale, especially on the application side
VC implication
We're observing a steeper “return curve” compared to the SaaS era
AI-native companies are showing unprecedented capital efficiency at scale
Revenue acceleration post-PMF is notably stronger than traditional software
Core alpha has concentrated in seed/pre-seed stages, before obvious signals emerge often at series A
This creates an interesting market dynamic: billion-dollar VC funds are essentially forced to compete for a small set of post-PMF companies, driving up late-stage valuations while the real opportunity lies earlier.
“Let’s talk money”: The Emergence of Action-Based AI Economy
As exemplified by most the foundation LLMs models, we are already moving from per-seat-based pricing to token-based.
A critical shift is occurring in how we value and price AI capabilities.
Prediction: in 2025, We're also branching into what I call the "cost-per-reasoning" paradigm. This transition fundamentally changes the unit economics of AI deployment:
Traditional Model: Pay per token (input/output)
Emerging Model: Pay per completed action or reasoning task
Future State: Autonomous AI agents operating as independent economic units
The implications are profound. When you can measure and price AI actions, you can start building real ROI models for AI deployment. A human knowledge worker costs roughly $40/hour in Western markets, with only about 25% of that time spent on actual reasoning tasks. The remaining time splits between rote work and organizational overhead. To further break it down:
Traditional Knowledge Work:
Fixed hourly rates
Linear productivity scaling
High organizational overhead
AI-Enabled Work:
Outcome-based pricing
Exponential productivity gains
Minimal operational friction
Current AI pricing trend suggests we're approaching human-equivalent costs for pure reasoning tasks. For example: A human analyst costs $150k annually to produce 200 reports. An AI agent / system at $100k can generate 20,000 reports, operating continuously.
But that's not the interesting part. The real inflection point comes when one human salary buys you multiple AI agents. This creates an interesting inflection point:
Current State: One AI agent ≈ One human replacement cost
Near Future (12-24 months): One human cost = Multiple AI agents
Long-term: AI agent networks that collaborate, creating multiplicative value
We're seeing early signs of this in research and high-value decision-making roles. The math would compound quickly when you factor in:
24/7 operation capability
Zero organizational overhead
Network effects between collaborating AI agents
Investment areas/ market unfolding predictions
Application Layer
Watch for breakthrough applications in high-value decision-making and tight vertical workflow integration (think healthcare, finance, legal, etc.)
Look for scalable AI agent networks rather than single-agent solutions
Prioritize companies that understand the unit economics of AI vs. human labor
Infrastructure / Dev Tool Layer
Look for companies building the "economic operating system" for AI agents
Focus on platforms that can price and manage action-based AI services
Bet on innovations in AI deployment efficiency
The common themes of capturing value? Three core principles:
Information wants to be free, yet Intelligence wants to compound. Network effects in AI are about “networked intelligence”, not just connections
Value accrues to systems that can self-correct and learn. Economic leverage comes from learning, not just scaling.
Finally, the best companies solve real problems, not just AI problems
The transformation won't be uniform. We'll likely see:
Early wins in high-value, reasoning-intensive sectors
Initial hybrid models where AI agents augment rather than replace
Accelerating adoption as cost-per-action drops below human equivalence
Prediction: three distinct phases of market maturation:
Phase 1 (Current):
High-value, reasoning-intensive applications
Hybrid human-AI workflows
Premium pricing for specialized capabilities (think GPT Pro and SORA)
Phase 2 (2025 -2026):
Broad enterprise adoption
Action-based pricing dominance
AI agent networks emergence
Phase 3 (3-5 years):
Ubiquitous AI integration
Autonomous agent economies
New economic paradigms
The debate over AI company valuations without the timeline horizon present a deep misunderstanding. We're not just witnessing another technology cycle—we're experiencing the emergence of an entirely new economic engine. The companies that understand this will create disproportionate value.
The internet looked expensive and widely misunderstood in 1995 too. But only if you were measuring the wrong things.