Coffee With David Zeng: Sell the Shovels, Charge for the Service

I had coffee with David Zeng, co-founder of Beacons AI, and walked out with a notebook full of sharp opinions and one homework assignment. These are the notes — lightly cleaned up, mostly in the order the conversation went.

A quick framing note: this is one founder's working theory of where software value is heading, told over coffee. Treat it as a strong opinion to think with, not a verdict.

"I want to cancel a lot of our AI subscriptions"

David runs a company that ships AI features — and he's actively looking at the stack of AI tools they pay for and asking which ones still earn their keep.

The clearest target: evals. He's not convinced most AI eval products are worth the money. His take was blunt — for a lot of what these tools do, you can vibe-code the equivalent yourself in an afternoon. The moment a paid product is a thin wrapper around something you could stand up internally, the math stops working.

The bar for paying has moved: if a competent engineer can rebuild your product over a weekend, you're not selling a product — you're selling the time it takes to rebuild it.

Most AI companies aren't selling the shovels

This was his sharpest critique, and it stuck with me. In a gold rush, the durable business is selling shovels — the thing everyone needs, no matter who strikes gold.

David's point: a lot of AI companies say they sell shovels, but they actually leave the hard part on your side of the table. You still have to implement the real work yourself — wiring up the instrumentation, building the dashboards, producing the graphs that tell you how your company is actually behaving. They hand you a box of parts and call it a shovel.

My takeaway: "Selling the shovels" isn't about handing over a tool. It's about removing the hard part. If the customer still has to do the hardest 80% themselves, you sold them a brochure for a shovel.

The sexy job attracts the crowd; the boring job solves the problem

We riffed on this for a while. The glamorous problems — the demo-friendly, tweet-friendly, "look what AI can do" problems — pull in enormous crowds. Everyone wants to build the sexy thing, so that market gets crowded fast and margins evaporate.

The boring jobs are different. They're unglamorous, nobody's racing to build them, and they're the ones that actually solve a painful problem for a real customer. The crowd self-selects away from the durable opportunities. It's the same reason I think go-to-market is an engineering problem — the unglamorous, instrumented, iterative work is exactly where the leverage hides.

The market is crowded where the work is fun and empty where the work is valuable.

Where we agreed: overpriced SaaS is on borrowed time

Here's where we found common ground. A lot of SaaS charges genuinely crazy prices for what amounts to a feature. David's position — and mine — is the same: if a tool is overpriced and the underlying thing is simple enough to vibe-code, he'll just vibe-code it. Open-source-able, rebuildable software has no pricing power anymore.

But — and this is the important half — that is not how he feels about real productivity tools.

Where he'll happily pay: Notion, Granola, Linear

Productivity software is a different category to him. Notion, Granola, Linear — these save a lot of people a lot of time, every single day. And when he runs the numbers, the decision is obvious:

The salaries of my employees are so much higher than what these tools cost. If a tool genuinely saves their time, paying for it is trivially worth it.

The point isn't "cheap vs. expensive." It's leverage relative to a salary. A $20/seat tool that saves an engineer a few hours a week isn't a cost — it's one of the best trades a company can make. He'll pay for that all day. What he won't pay for is a marked-up wrapper around something free.

The real product is the service, not the software

This was the line that tied it all together, and it matched something I'd been circling for a while. The point of SaaS is the Service — not the Software. It rhymes with the "Service as a Software" idea I wrote about in Stop Treating AI as a Tool: once the software itself is easy to reproduce, the value moves to everything wrapped around it.

The software is increasingly commoditized. What you're actually buying is the service wrapped around it: the SLA, the SLO, the guarantee that it's up, that it's maintained, that someone is on the hook when it breaks. That's the part a weekend vibe-coding session can't replicate. You can clone the features; you can't clone the on-call rotation and the contractual promise that it works at 3am.

You can vibe-code the software. You can't vibe-code the SLA.

That's the moat now. Not the code — the commitment.

How we actually met

I told David my background and the slightly improbable story of how we ended up here. He'd reached out by email with a thoughtful note about how my product could be better — not a pitch, just genuinely useful feedback. The moment I read it I connected with him on LinkedIn, we had a few conversations, and somewhere along the way it stopped being a networking thing and we just became friends.

He also shared his own path: from Stanford PhD student to founder — leaving the program to start a company. Objectively a wild decision. But hearing him talk about it, it sounded less like a gamble and more like the obviously fun thing to do.

The assignment

He left me with homework, and I like it enough to actually do it. Since I live right next to Hayes Valley, the assignment is simple:

Go talk to three different people in San Francisco.

No agenda beyond that. Just talk to people. I'll report back on how it goes.

Frequently Asked Questions

What does "sell the shovels" mean in the AI context?

It's a reference to the gold-rush idea that selling tools to prospectors is more durable than prospecting yourself. David's twist: many AI companies claim to sell shovels but actually leave the hardest work — instrumentation, dashboards, real integration — for the customer to do. Truly selling shovels means removing the hard part, not packaging it.

Why is David willing to pay for Notion or Linear but not for AI eval tools?

Leverage relative to salary. Productivity tools like Notion, Granola, and Linear save expensive employees real time every day, so the cost is trivial against payroll. AI eval tools, in his view, are often thin enough that an engineer can rebuild them internally — which removes their pricing power.

What's the difference between selling Software and selling Service?

The software (the features, the code) is increasingly easy to clone or vibe-code. The service — uptime guarantees, SLAs, SLOs, maintenance, someone accountable when it breaks — is the part you can't replicate over a weekend. David's argument is that the real, defensible product is the service, not the software.