📚 Chapter 2Part I: The Repricing of Venture

Building With AI: What I Learned by Actually Doing It

Investors who build understand what they fund. There is no substitute for skin in the game.

TC
Trace Cohen
3x founder · 65+ investments · Author, The Value Add VC

Key Insight

There is an enormous gap between using AI as a consumer and building production AI systems as a founder. Latency, cost-per-query, context window limits, hallucination rates, and reliability under load are invisible in demos but decisive in production. Investors who build with AI understand these constraints in ways that passive observers cannot — and it changes how they evaluate every AI pitch.

10x
Cost difference: demo vs. production at scale
3 wks
Time to prototype (vs. 6 months pre-AI)
2 eng
Team size for MVP that once took 10
∞
Gap between working demo and production system

Why I Started Building

I had been investing in AI companies for years before I started building with the tools myself. I understood the category intellectually. I had read the papers, evaluated dozens of companies, and developed what I thought was a solid mental model of what was real and what was hype.

Then I started building. Connected to Vercel, Supabase, Google Cloud, and a growing list of other services. Building real products, under real constraints, with real users who would tell me exactly when something was broken. The gap between what I knew from evaluating companies and what I learned from building was enormous.

The Demo is Not the Product

Every AI demo is optimized for the demo. Short inputs, best-case outputs, pre-selected examples, hidden latency. When you build in production, the real constraints emerge immediately.

  • Latency compounds. A single LLM call might take 2-3 seconds. A multi-step agentic workflow with 5 calls takes 10-15 seconds. Users abandon workflows that take 15 seconds to respond.
  • Cost scales faster than you expect. A feature that costs $0.002 per query in testing costs $2,000/day at 1M queries. Unit economics that look fine at demo scale become existential at production scale.
  • Context windows fill up with real data. Production systems have real documents, real conversation history, real tool outputs. These fill context windows fast, requiring careful context management that demos never need.
  • Hallucination rates that are tolerable in demos are not tolerable in enterprise. A 2% hallucination rate sounds small until it's a wrong medical dosage, an incorrect legal citation, or a fabricated financial figure.

What Changes When You Build

When you build with AI tools, you develop pattern recognition that changes how you evaluate every pitch deck. You stop being impressed by demos that work perfectly on stage. You start asking about production reliability, p99 latency, error handling, and what happens when the LLM returns something unexpected.

You also learn what actually matters for defensibility. The model is not the moat. API access to any frontier model gives you capability. What creates durable value is what you build on top: proprietary training data, domain-specific fine-tuning, workflow integrations that create switching cost, and compliance infrastructure that enterprise buyers require before they'll bet their operations on your product.

From the Book

“Building under real constraints, with real users, making real mistakes, taught me more about AI-era company building than any pitch deck I've reviewed. There is no substitute for skin in the game.”

— Trace Cohen, The Value Add VC

The Prototype-to-Production Gap

AI has genuinely compressed the cost and time of building working prototypes. A feature that once required a team of 10 for 6 months can now be built by 2 engineers in 3 weeks. This is real and it matters for startup formation rates, competitive dynamics, and what counts as a meaningful barrier to entry.

But it has not compressed the cost of making a prototype production-ready for enterprise deployment. Security reviews, SOC 2 compliance, HIPAA or FedRAMP requirements, integration with legacy systems, SLA guarantees, on-call rotations, disaster recovery — none of this gets easier because the model underneath is smarter. The time to ship a demo has collapsed. The time to ship something a Fortune 500 will bet its operations on has not.

What This Means for Investment Decisions

Investors who build develop a specific kind of skepticism that makes them better evaluators of AI companies. They can tell the difference between a team that has shipped production systems and a team that has shipped demos. They understand which technical claims are sound and which are optimistic. They know what questions to ask about infrastructure, reliability, and the actual cost structure at scale.

More importantly, they understand why the moat is not the model. The best AI investments of the next decade will not be in companies with proprietary foundation models — training costs are too high and competitive dynamics are too intense. They will be in companies that have accumulated proprietary data, embedded deeply into specific workflows, and built the compliance infrastructure that makes switching genuinely painful.

You can read about this difference in pitch decks. Or you can build and feel it yourself. Only one of those leaves you with real conviction.

Frequently Asked Questions

What's the biggest misconception investors have about AI startups?+
Most investors evaluate AI startups based on demos, which are optimized to hide latency, cost, and failure cases. Production AI systems face real constraints: context windows fill up with real data, latency compounds in multi-step workflows, and hallucination rates that seem acceptable in demos cause real problems in enterprise workflows. Investors who have built understand these gaps; those who haven't often over-fund the demo and under-fund the production reality.
How has AI changed the cost and speed of building software?+
AI has dramatically compressed the time and team size required to build working software. A function that previously required 10 engineers for 6 months can now be prototyped by 2 engineers in 3 weeks. This is real and has real implications for startup formation. But what it has not changed is the cost of making that prototype production-ready for enterprise: security reviews, compliance, reliability infrastructure, and the human trust required to embed deeply in workflows still take years.
What should founders know before building an AI product?+
Founders should understand that the model is not the moat. API access to GPT-4 or Claude gives you capability, not defensibility. What creates durable value is what you build around the model: proprietary data, domain-specific fine-tuning, workflow integration that creates switching cost, and the compliance infrastructure that enterprise buyers require. The prototype is the easy part. The moat is everything after it.
Why should investors build with the tools they invest in?+
Building with AI tools creates pattern recognition that is impossible to acquire any other way. You learn which claims in pitch decks are technically sound and which are demo tricks. You understand what 'production-ready' actually requires. You develop intuition for which teams have shipped real systems versus which are still in prototype. Trace Cohen argues that this hands-on knowledge makes you a meaningfully better investor in the AI category.
←
Previous
Ch. 1: Capital Cost, Liquidity, and Duration
All Chapters
Next
Ch. 3: The Bifurcation: Mega-Funds and Emerging Managers
→
📚

Read the Full Book

22 chapters on how venture capital actually works — the math, the mechanics, and the decisions that compound over time.

Read Free Online →Browse All Chapters