Let me be clear about something: I'm not bearish on AI as a technology. I've watched transformer architectures quietly eat the world, seen inference costs drop by orders of magnitude, and debugged enough production RAG pipelines to know this stuff is genuinely powerful. What I'm skeptical about is the idea that current stock valuations have any reliable relationship to what these companies will actually be worth in five years.

The Hype Cycle Is Doing What It Always Does

Every few years, a legitimately transformative technology shows up and immediately gets priced as if every company touching it will become the next Standard Oil. We did this with the internet (mostly wrong in the short term, right in the long term), we did it with blockchain (mostly just wrong), and now we're doing it with AI. The difference this time is that the underlying technology actually works—which somehow makes the valuation math even harder, not easier.

When the product is real, separating signal from noise becomes a genuine technical problem. You can't just dismiss it all as vaporware. But you also can't assume that because the technology works, every company slapping 'AI-powered' onto their pitch deck deserves a 40x revenue multiple.

The Moat Problem Nobody Wants to Talk About

Here's the question I keep asking that nobody in the financial press seems interested in: where are the durable competitive advantages? Real moats in AI are actually pretty narrow:

• Proprietary data at scale — not just 'we have data,' but data that's genuinely hard to replicate and meaningfully improves model performance
• Inference infrastructure — owning the compute stack and being able to serve at costs competitors can't match
• Distribution lock-in — embedding deeply enough into enterprise workflows that switching costs become prohibitive
• Foundational model capability — being one of the handful of labs with the talent and capital to train frontier models

Most AI companies being bid up right now have none of these. What they have is a fine-tuned model sitting on top of OpenAI's API and a good sales team. That's a services business, not a tech platform. The moment a cheaper API ships, their margins compress.

Benchmark Theater vs. Production Reality

The press releases love to cite benchmark scores. Model X achieves state-of-the-art on MMLU! Company Y's system beats GPT-4 on HumanEval! What they don't tell you: benchmarks are increasingly gamed, and the gap between benchmark performance and production reliability is a canyon.

I've shipped enough AI features to know that hallucination rates, latency at the 99th percentile, and cost-per-query under real load are the numbers that actually matter. Those numbers don't make it into the investor decks. The companies that are quietly solving those problems—consistent reliability, predictable costs, low failure rates on edge cases—those are the ones worth watching. They're usually not the ones getting the biggest headlines.

The Capex Elephant in the Room

Training and running large models is expensive in ways that don't show up cleanly in quarterly earnings until they suddenly do. The hyperscalers are spending hundreds of billions on AI infrastructure. That spend has to generate returns somewhere. Either the cloud AI services they're building become enormously profitable, or those impairments hit the balance sheet eventually. We're in the investment phase, which means current earnings look detached from the actual capital being consumed.

When the depreciation schedules start biting and we're in a higher-rate environment, the math changes. I'm not predicting a crash—I'm saying the timing is genuinely unknowable, and 'the fundamentals will catch up to the price eventually' is a sentence that has preceded a lot of painful holding periods.

So What Would Actually Change My Mind?

I'm not saying never. I'm saying not at current prices without a clearer picture of who captures the value. Show me a company with genuine proprietary training data, improving gross margins as they scale (not degrading ones), enterprise contracts with real switching costs, and a product that's solving a specific high-value problem rather than being a general-purpose AI wrapper. That's a business I can get interested in.

Until then, I'll keep building with these tools, watching the landscape carefully, and letting other people's FOMO do its thing. The technology is real. The valuations are a different conversation entirely.