The AI Wave — Notes & Observations

Collected notes from VC podcasts and industry conversations on the AI investment landscape, LLM economics, vertical applications, and where things seem to be headed.

1. Enterprise Software Evolution

Enterprise software has gone through three distinct generations:

Enterprise 2.0 was commercially successful but never delivered the step-function productivity gains people expected. Most of the value turned out to be moving from on-prem to cloud, which improved software delivery but didn't fundamentally change business operations. The human remained the bottleneck in every process decision.

Productivity did improve in some sectors (financial services) but barely moved in others (healthcare, construction), as measured by TFP (Total Factor Productivity). Revenue and earnings per employee in the S&P 500 have risen steadily over 25 years, so the effect was real, just incremental. Tech diffusion works like that: 2-3% per year, compounding over long periods.

2. The AI Wave: Scale, Speed, and Structure

Growth Velocity

• Top 100 AI companies are growing at 3.6x the median rate of top 100 SaaS companies (per Stripe data).
• Over two years, that compounds to ~13x faster growth.
• Companies are raising multiple rounds per year at significant markups.

The AI Investment Stack

Value accrues differently at each tier. The lower tiers (energy, chips, data centers) are capital-intensive and increasingly dominated by incumbents. The upper tiers (infrastructure, applications) are where most startup activity is concentrated.

3. Are LLMs Commodities?

Right now, no. Model API businesses run 50-60% gross margins. True commodity businesses (metals, agriculture) run 5-15%. That's not a commodity.

But the answer depends on the domain going forward.

Why Opinions Diverge

Everyone's position on this maps directly to their economic interest:

• Model company investors ($3-5B deployed) insist models are worth trillions.
• Data and infrastructure companies (data warehouse CEOs, etc.) openly call models commodities because their own value rises if models are fungible.
• Nvidia wants models somewhat commoditized (so they capture hardware value) but not too commoditized (so labs keep raising money to buy chips).

Show me the incentive, I'll show you the opinion.

Three Scenarios for LLM Value

1. The Singularity Thesis: nothing matters except who reaches AGI/ASI first. Winner takes all. (This one is almost religious in character.)
2. Sustained Frontier Premium: the frontier stays meaningfully ahead through 2032+. Great business for years.
3. Catch-Up Commoditization: open-source and competitors close the gap within months and the frontier advantage shrinks to insignificance.

The Hyperscaler Analogy

Cloud primitives (S3, EC2) are commodities, but AWS/Azure/GCP built managed services on top that created massive switching costs. The parallel question for model companies is whether they can build enough value-added services (memory, personalization, managed infrastructure) to lock in customers.

4. Model Company Business Model Divergence

Each major lab is pursuing a different business, shaped by whatever data asset they happen to sit on:

Building the model alone is not enough. You need a business around the core capability because catching up to the frontier is getting easier over time. Software defensibility has always come from product and data gravity (memory, personalization). Model companies will be no different.

5. Frontier Models vs. Vertical Applications

In most verticals, use-case-specific application companies outcompete the labs. Last-mile delivery of customer experience is a ton of work. The model primitive alone is not enough. Focus, customer understanding, and network effects all matter, same dynamics as SaaS.

The exception is code generation and developer tooling, which sits so close to the labs' core competency that they compete directly. Even here, though, the market seems large enough for multiple winners (Claude Code, Cursor, GitHub Copilot, Cognition/Devin).

6. Where AI Is Working Today

The verticals where AI is landing fastest share a common pattern: the workflows decompose into a handful of core AI skills. If a job maps to these, AI can do it:

1. Multi-turn conversations
2. Document understanding
3. Structured output
4. Search
5. Reasoning
6. Computer use

Logistics & Freight Brokerage

• AI agents handle proof-of-delivery collection, load negotiations, status queries, carrier sourcing.
• Handles ~90-95% of cases autonomously and escalates the hard ones to humans.
• Humans become managers of AIs, handling only the challenging cases. 24/7 availability means 100% of inbound calls get fielded, which is a massive expansion of the addressable market.
• The dynamic flips: software creates prioritized work queues for people, rather than people driving software.

CPG Back-Office

• Automating deductions management for CPG brands: figuring out why a retailer paid $70 for $100 of goods.
• Tight integrations with retail/distributor portals create data access that enables expansion into adjacent back-office workflows.

User Research

• AI agents conduct user/market research interviews at scale, synthesize learnings, deliver insights.
• Breaks the old tradeoff between high-fidelity/expensive manual interviews and low-fidelity/cheap surveys. The cost curve shifts enough to make it economical to interview every user.

Accounting & Audit

• AI agents doing the practitioner's work for CPA firms on financial and risk audits.
• CPAs get fixed fees per engagement; AI gives them 50%+ more capacity in a labor-supply-constrained market (CPA count declining nationally).

Title Insurance

• Automating title search and analysis, which is the core workflow of title agents.
• Title work maps cleanly to AI skills (document understanding, search, structured output, reasoning).

Cybersecurity Red-Teaming

• AI agents trained on elite cybersecurity knowledge that are now outperforming human red-teamers.
• Reasoning-intensive work where AI excels: planning, understanding systems, finding exploits. Agents managing other agents.

7. The Incumbent Software Dilemma

Existing SaaS companies have every advantage here: data gravity, workflow gravity, distribution channels, customer relationships. Yet the vast majority haven't launched meaningful AI products.

The reasons are mostly organizational. AI eats into gross margin profiles. Pricing model changes cannibalize existing revenue. Leadership isn't bold enough. Talent competition is fierce because AI is still an entrepreneurial motion, and companies without that culture can't execute.

The surprising part of this whole wave is that technology isn't the bottleneck. Process and organizational structure are.

8. Market Expansion, Not Just Market Capture

The most interesting AI economics aren't about taking existing market share. They're about expanding the total addressable market.

• Programming: $3T/year global spend on programmers. Even a 30% productivity gain is ~$1T of new value.
• Customer research: it was never economical to interview all bottom-up PLG (Product-Led Growth) customers. Now AI can do it.
• Freight brokerage: 24/7 AI concierge fields 100% of calls vs. whatever fraction humans could handle.
• Accounting: more engagements served with same headcount in a shrinking labor supply.

9. Bubble or Not?

The Case Against a Bubble (Current State)

• Nvidia trades at ~25x forward earnings, which isn't extreme.
• Micron up 250% in 12 months but still at 9x forward PE (Price-to-Earnings ratio).
• Aggregate venture capital deployment is well below 2021 ZIRP (Zero Interest Rate Policy) levels.
• AI capex is ~3-4% of GDP. Historical industrial revolutions saw 20-50% of GDP in investment.
• If this is truly an industrial revolution, investment levels will get much crazier.

The Case for Eventual Bubble Dynamics

• Pockets of bubbly behavior already emerging: investors who miss a round invest 2 months later at 5x the valuation with minimal incremental progress.
• Technology is more cyclical than people credit. The ZIRP era proved that.
• Overcapacity correction is inevitable at some point (semiconductor manufacturers are inherently cyclical).
• Startups indexed to serving other startups are especially vulnerable to venture capital cycles.

Where It Lands

The direction is right but timing is hard to predict. Productivity gains will probably unfold slowly over 10 years. You can be long the trend, rationally optimistic, and still lose money if you're wrong on timing or valuation.

10. Biggest Concerns

Social & Political

• Job displacement friction is real, even if the long-term outcome is net-positive.
• The Luddites had a point about lifestyle disruption even if the economic outcome was better.
• Both left and right want to "manage" the transition for political power, which creates risk of populist capture.
• The US spends $40B/year on worker retraining programs that are largely ineffective.

Financial

• Cyclical overshoot is inevitable. The economy can only absorb so much so fast.
• Technology diffusion creates a gap between what's possible and what's deployed. Macro analysts miss this because they're too far from the ground reality.

11. Sources of Optimism

Key Recurring Themes

• The human is the bottleneck, not the technology. AI's promise is removing that constraint.
• Technology is already ahead of adoption. The frontier models could freeze today and still reshape the economy. Diffusion is the limiting factor.
• Vertical, use-case-specific companies beat generalist model providers in every domain except coding. Focus and last-mile delivery win.
• Every stakeholder's view on commoditization maps to their economic position. Show me the incentive, I'll show you the opinion.
• The most interesting AI economics come from enabling previously uneconomical activities, not just capturing existing market share.
• Incumbents have every advantage (data gravity, customer relationships) and are mostly squandering it.
• Agents create work queues and escalate. Humans supervise. Software drives people now, not the other way around.