AI Labs Tout AGI, But Cash Flow Matters First

No matter how loudly people shout about AGI, they still have to answer one question first: is the cash flow healthy? Recently, overseas SaaS stocks have rebounded; Salesforce's latest earnings report sent its stock price soaring over 22% in a single day. On the other side, GPT-6 Astra was released, with the president saying "Welcome to the AGI era," citing scores of 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4. It's certainly buzzing, but I'm looking at where the money comes from.

From a financial perspective, the easiest battles for AI are indeed tasks like copywriting, retrieval, forms, and code completion. The path is short, delivery is fast, and benchmarks look good. But as model capabilities become more like public utilities (water/electricity), gross margins at the application layer get squeezed easily. I just started using pandas these past few days, ran some expense details and inference usage data through it, and the conclusion wasn't surprising. Looking solely at effectiveness, AI can do the work; but when calculating unit costs, many scenarios are just swapping human hours for GPU hours without generating higher margins.

One-sentence review: The easy battles are almost over; what's left is dirty work. Permissions, logs, data standards, failure cases, customer payment cycles—these aren't sexy, but they determine valuation. Previously, I thought agents lacking governance would take a discount; now my view is even harder. Models can calculate, but accounts might not be auditable. Only auditable cash flows deserve high multiples.

Advice for those evaluating projects: Don't just stare at ARR and demos. Break down inference costs, manual reviews, retention, and payment cycles, and ask: Is the cash flow healthy? If model manufacturers start building coding agents themselves, the moat for wrapper applications needs to be reassessed. That's it.