Month-One AI Stack Cost for a Solo SaaS: $176 Itemized

The premise of "AI stack" articles is usually wrong about the same thing: they price the paid tier of every vendor and assume the founder needs all of them in month one. The free tiers are not vestigial. They are the working tier for the first 100 to 10,000 users at most vendors. The only line that always costs real money on day one is the AI API, and that is because tokens are metered per call regardless of user count. This walkthrough runs a careful month-one stack through the calculator and itemises every dollar.

1. The month-one scenario

The scenario is an early solo SaaS (testing audience, not paying yet) with 12 Claude Sonnet 5.5 API calls per user per day at 1,500 input and 500 output tokens per call. Inputs to the engine: the free tiers of Vercel hosting, Supabase database, Clerk auth, Resend email and Sentry monitoring. Domain runs $14/year. "Other" (Stripe fees, design tools, misc) lands at $25/month. The engine projects fixed user tiers; the month-one read is the 100-user tier.

What the engine returns at the 100-user tier: $314.17 total monthly cost. AI API alone is $288. Hosting, database, auth, email, and monitoring all return $0 because the engine knows the free tier of each covers a 100-user product. Domain costs $1.17 per month ($14 / 12). Other is $25. That is the entire month-one bill if you do not pay for a paid tier on every vendor before you need one. The 99.91% dominant cost driver the engine reports (at its 10,000-user reference tier) is correct: at this volume, everything else is rounding error compared to API.

2. The five vendors that stay at $0

Most month-one cost spreadsheets price every line at the paid tier. The engine inverts this: every vendor below volume threshold returns $0 because the free tier covers it. The thresholds, drawn directly from each vendor's published pricing:

• Vercel Hobby ($0): 100 GB of Fast Data Transfer and 1 million CDN requests a month^[1]. Vercel restricts Hobby to non-commercial, personal use^[8], so it fits a pre-launch test; once the product takes payments, Vercel Pro at $20/month is the compliant plan.
• Supabase Free ($0): 500 MB Postgres, 50,000 MAU on auth, 1 GB file storage, 5 GB egress^[2]. Sufficient for any solo SaaS that has not yet hit product-market fit. Supabase Pro is $25/month and adds 8 GB of disk per project and daily backups kept for 7 days.
• Clerk Hobby ($0): Up to 50,000 monthly retained users per app. Above that, Clerk Pro is $25/month plus $0.02 per additional user^[6]. A month-one product almost never crosses this line.
• Resend Free ($0): 3,000 emails/month, 100/day. Enough for transactional email through 1,000 users at low engagement, or 300 users at heavy engagement.
• Sentry Developer ($0): 5,000 errors a month, 1 user, 30-day lookback^[7]. Sufficient until error volume exceeds 5,000 a month; Sentry Team at $26/month covers 50,000.

The temptation is to upgrade to Pro tiers because "you'll need them eventually." Resist. The marginal cost of upgrading vendors as you need them is identical to the cost of upgrading them on day one — but the cash conservation in those first months can extend personal runway by 4 to 8 weeks at solo scale. The burn rate calculator shows the compounding effect.

3. The $288 AI line item is the whole bill

AI API at 100 users, 12 calls per user per day, 1,500 input and 500 output tokens, Claude Sonnet 5.5 at $2/$10 per million tokens^[3]:

Calls per user per month       12 × 30 = 360
Total monthly calls (100 users) 100 × 360 = 36,000
Input tokens                   36,000 × 1,500 = 54,000,000
Output tokens                  36,000 × 500   = 18,000,000
Input cost                     54 × $2   = $108
Output cost                    18 × $10  = $180
Total AI API                                       $288 (engine returns $288)

Two things matter operationally. First, output tokens drive 62.5% of the AI bill at this split. Reducing average output from 500 to 350 tokens saves $54/month. Second, this is per 100 users: at 1,000 users it becomes $2,880/month, at 10,000 users $28,800. The slope is steep, and free tiers do not blunt it. The startup cost estimator handles the cross-vendor projection.

4. Domain + 'other' = $26.17

The remaining two line items are domain and "other." Domain costs roughly $14/year on a .com registered through Cloudflare or Porkbun (skip the registrar that bundles a paid SSL — Cloudflare gives away free certificates that work fine). That is $1.17/month amortised.

"Other" at $25/month is the line every solo founder underestimates by a factor of 2. It typically includes: Stripe's fixed monthly fee (zero on standard accounts), Figma free tier, GitHub Free, ChatGPT Plus or Claude Pro for the founder's own use ($20), domain privacy and DNS niceties, and the inevitable $5 SaaS that creeps in. The honest range is $20 to $50 in month one; $25 is the careful number.

5. Where the 3x overspend starts

Founders who pay $430+ in month one (the $314 bill plus about $116 of paid tiers) when the careful bill is about $176 are paying for capacity they cannot use. Three patterns recur. First, paying $20 for Vercel Pro while the product is still a free pre-launch test; Hobby's 100 GB of transfer covers that stage, and Pro becomes necessary once you charge customers. Second, paying $25 for Supabase Pro before the 500 MB free tier is anywhere close to full. Third, paying $25 for Clerk Pro before the 50,000-user free limit is even visible.

Add a few more — Sentry Team at $26, Resend Pro at $20, a paid email-deliverability scanner, and a "small" monitoring service — and the month-one bill clears $500 fast. None of those vendors will help the product reach product-market fit. They are insurance against scale that has not arrived. The careful number (about $150 of AI after routing, the output cap and prompt caching, plus $25 other and $1.17 domain, so about $176) holds because it includes only the items that produce value at month-one volume.

The 3x overspend almost always comes from one specific failure mode: copying a "modern SaaS stack" article and pricing every vendor at their published tier without checking which tier you actually need. The vendor pricing pages are written to optimize for revenue at scale; the founder needs to optimise for cash conservation pre-PMF.

6. The scaling tiers: when free tiers break

The engine projects to 100, 1,000, 10,000, and 100,000 users — every tier is in the engine block above. The interesting transitions are at 1k and 10k: AI API rises from $288 to $2,880 to $28,800, total cost tracks it almost exactly ($314.17 → $2,906.17 → $28,826.17), and at 100,000 users Clerk auth finally turns on at $1,025/month for a $289,051.17 total.

Cost per user is essentially flat from 100 to 100,000 ($3.14 → $2.91 → $2.88 → $2.89) because AI API scales linearly and every other line is fixed-step. The step that matters is Clerk crossing 50,000 users into its Pro plan ($1,025/month at 100,000 users). Below that, every fixed vendor cost is absorbed by free or near-free tiers. Cloudflare Workers — not in this scenario but the cheapest viable replacement for Vercel at high scale^[4] — extends the free-tier runway by another 5 to 10x.

The strategic implication is that AI cost is the only variable that matters for product economics at solo scale. Every other vendor decision either stays free or becomes a rounding error compared to model spend. Founders should spend zero time optimising hosting choice at month one and 100% of their cost-optimisation time on prompt design, model routing, and output caps.

7. Month-one actions ordered by impact

Five moves, ranked by dollar impact:

• Route Claude Sonnet calls to Haiku where quality permits. If half the calls run fine on Claude Haiku 4.5 ($1 input / $5 output, half of Sonnet 5.5's price), the AI line drops from $288 to $216: about $72/month saved.
• Cap output tokens at 350 instead of 500. Saves about $54/month on the Sonnet-only baseline, or about $40 once half the calls are on Haiku, before quality testing reveals truncation issues.
• Enable prompt caching on the system prompt. Cuts input cost by 26-40% depending on cache hit rate: $21-$32/month on the $81 of input cost left after routing.
• Stay on free tiers for everything non-AI. Saves about $116/month by not paying for Vercel Pro ($20), Supabase Pro ($25), Clerk Pro ($25), Sentry Team ($26), and Resend Pro ($20) before they are needed.
• Defer the domain privacy/SSL upsells. Cloudflare provides both at $0. Saves $10-$15/month on registrar add-ons that solo founders accidentally subscribe to.

Combined, the month-one bill drops from $314.17 to roughly $176 without changing any product behaviour a user can notice. Track every dollar. Run the calculator again at month three and month six — the cost shape will have changed less than you expect, and the same five levers will still produce most of the savings. See the methodology for the full derivation^[5].

Frequently asked questions

What is the actual month-one cost of an AI SaaS stack at 100 users?

On the free tiers of Vercel, Supabase, Clerk, Resend and Sentry, with Claude Sonnet 5.5 at 12 calls per user per day and 1,500/500 token sizes, the calculator returns $314.17 of monthly cost, $288 of it AI API. At 100 users every vendor is still inside its free tier, so the marginal cost per added user is roughly $2.88 of model spend.

Which vendors actually have to be paid in month one?

Only the model provider and the domain. Vercel Hobby, Supabase Free, Clerk Hobby (up to 50,000 monthly retained users), Sentry Developer, and Resend Free cover everything else through at least 100 users. Vercel limits Hobby to non-commercial, personal use, so budget $20 for Vercel Pro once the product charges customers. The temptation to upgrade to paid tiers in month one is the single largest source of 3x overspend in solo SaaS launches.

When do the free tiers actually break?

It depends on the vendor's metered limit. Clerk's free Hobby plan covers 50,000 monthly retained users per app. Supabase Free caps the database at 500 MB and egress at 5 GB a month; Vercel Hobby includes 100 GB of Fast Data Transfer but is limited to non-commercial use. Sentry's free Developer plan covers 5,000 errors a month; Team at $26 covers 50,000. The AI API line scales linearly from day one; everything else has a step function.