Cost of building an MVP (with AI)
In last week’s post, I explained why I took time off from Duolingo to build Lazyweb.com. This one is about the less romantic part: what it actually cost to get the MVP live with just me and a bunch of AIs.
Short version: I spent more than I needed to, on purpose.
The Non-Negotiables
Domain (Lazyweb.com): $13,000
Why spend that much on a domain?
Because descriptive names compound.
If you’ve read First1000, you already know this is a pattern for me. Descriptive names are easy to remember, easy to explain, and easy to search for. Even if someone forgets the exact name, they remember something adjacent. “First something newsletter.”
That matters a lot early on for word of mouth for nascent products that are not meant to be used daily.
Made-up names were a hard no for me for that reason.
Also, SEO is core to Lazyweb, so I wanted:
A name that had never been used before
A .com, because Paul Graham said so.
A descriptive, unused .com with virtually zero SEO competition is rare. It took six months of negotiation, impersonating different buyers, and $13k to make it happen.
Stripe Atlas: $500
Company setup via Stripe Atlas. I got incorporation, banking, legal boilerplate, and a few other perks without much thinking or effort. Could I have done it cheaper? Sure. Would I do it again? Yes.
Infrastructure & Backend
Render: $152
Render is hosting without the AWS anxiety.
I have personally been burned by misconfigured AWS services turning into four-figure bills overnight (that AWS generously waived after a few days of back and forth with support). Render gives me:
Predictable pricing
Great performance and monitoring
Enough abstraction that I can focus on the product
Big fan (:
Supabase: $0 (Startup Program)
After Render, Supabase is my second true love. Auth, database, storage, realtime, all in one place. Thanks to their startup program, it’s free for now.
I do not want to imagine building Lazyweb without it.
Twilio (SendGrid): $119
Used for welcome emails and internal alerts. I’m not thrilled with it and will give Resend a spin soon.
This is mostly inertia from First1000 days since I have used SendGrid for some email flows related to First1000.
Coding & Dev Tools
Cursor: $1,101
My daily driver for a coding agent is Cursor (Ultra $200/mo)
I tried Claude Code, Codex, and Replit agents. Cursor won.
It keeps getting better, feels native to how I think, and saves me hours every week. I still use Codex occasionally for parallel tasks, but Cursor is where real work happens.
Replit: $138
I experimented a little bit with Replit.
For instance Lazyweb’s landing page is built on Replit. It’s great for small throwaways and quick prototypes with managed hosting.
For a serious product, I found it expensive and limiting.
AI Models & Research
OpenAI (ChatGPT Team (x2 seats) + API ): $3,057
This is the biggest line item after my domain.
I pay for a two-seat ChatGPT Team plan. That gets me:
Access to the Pro models (limited to 10 requests/month)
Generous limits for Codex and GPT-5.2 Thinking
Pay-as-you-go API credits for Deep research, GPT-5.2 Pro and Codex (which I occasionally but use their code review a ton)
GPT-5.2 Pro + Pay as you go is the best ROI on the money I spent so far.
When things break, and they occasionally do, I export all my code into a GPT-5.2 Pro in Deep Research inside of ChatGPT and things start working working. It’s my magic trick that I suspect not a ton of people know about.
xAI (Grok): $5
I use grok-4-fast-reasoning for runtime search queries on Lazyweb.com.
Grok 4 Fast has the best latency-to-accuracy tradeoff I’ve tested. It’s not the smartest model, but it’s fast and good enough.
It’s only $5 spend because it’s a variable usage cost and I am currently the only user/tester.
Anthropic (Claude): $20
I experimented with Claude via Anthropic.
Strong model, but I kept defaulting back to Cursor and Codex. With ChatGPT Team already paid for, Claude became a secondary option. Also reasoning over image input for Claude is subpar.
Gemini: $40
Using Nano Banana primarily for images and icons in the app.
Nano Banana is great for one-shot outputs and consistent visual styles. It’s also stubborn once it goes off track, which makes iteration painful.
Still useful in narrow cases but its not the breakthrough it is hyped to be. In fact, try editing any text inside an image and you will quickly lose your mind.
Parallel Research Agents: $165
This is a new startup from Parag (briefly the CEO of X, formerly Twitter). It uses multiple deep research agents running in parallel.
This is useful for scraping + reasoning-heavy tasks, but easy to overspend on and accuracy and evals still not there 100%.
Powerful in narrow use cases, but dangerous if you’re not careful, or let it run in the wild without supervision.
Productivity & Ops
Duckbill: $204
AI + human executive assistance via Duckbill.
I was skeptical. I’m not anymore.
It frees up 2 to 5 hours a week by handling annoying but necessary personal tasks. Filing insurance claims. Calling restaurants. Life admin.
All the stuff that steals focus without adding leverage, for almost nothing.
Wispr Flow: $0
A new favorite flow for me is using voice dictation via Wispr Flow when writing prompts or drafts for a newsletter.
I get a year free through Lenny’s subscriber bundle.
I was not a believer in voice input. I am now.
Deepgram: $0 (Startup credits)
Similar to Wispr Flow, Deepgram is Speech-to-text. Only this time it is an API.
Given my new found belief in voice dictation, I took a little detour to implement these voice features in the Lazyweb.com MVP and it puts a smile on my face every time I use it.
I got Startup credits covering the first $200.
CapCut: $192
I’m doing a decent amount of video editing in CapCut, mostly for unhinged AI-generated ads to support the Lazyweb.com’s launch on X.
CapCut has a horrendous Mac app. It is also not a great traditional editor.
But it’s the closest thing to a video editor with deeply integrated AI, which matters for launch content.
My days with CapCut are numbered but I have bigger fish to fry.
Google Workspace: $181
Email plus Gemini access via Google Workspace.
This covers business email accounts and Google’s AI features. Not exciting, but necessary.
ZeroBounce: $45
ZeroBounce is one of these obscure products that I have been relying on very heavily for years.
It’s an email validation tool to clean the Lazyweb waitlist.
It works and I moved on.
Miscellaneous: $92
Dogfooding competitor apps and random one-offs that don’t deserve their own line item.
Final Tally
I set aside $30k to get Lazyweb off the ground and so far, I’ve spent about $19,100, mostly on the domain and AI credits.
Could I have built a strong MVP for less? Absolutely. I overspent in a few places because speed, learning, and leverage mattered more to me than optimization.
The goal now is simple: make every future dollar work harder than the last one.
Ciao,
Ali Abouelatta