Three founders asked us last month for our 'real numbers' on AI startup costs. The pitch decks all say 'lean'. The reality is rarely lean. Here's what's actually being spent.
These are aggregated from real teams we've shipped products with at three stages: pre-revenue, early revenue ($10K MRR), and scale ($1M ARR).
- LLM inference: $200-800. Mostly Claude / GPT-5. Cache aggressively or this triples.
- Infra: $100-300. Vercel + Supabase + Upstash covers most needs.
- Vector DB: $0-70. pgvector free until you scale. Pinecone Standard is $70/mo if you must.
- Tools: $200-500. Linear, Slack, GitHub, Posthog, Cursor - this adds up.
- Founder salary: $0-3K. Most ramen-budget founders pay themselves nothing.
- Contractor or co-founder: $5-10K. One technical hire or design contractor.
- Inference scales 10x: $3-8K/month. Real users hammer the model.
- Compliance prep: $2-5K. Drata or Vanta + a fractional security person.
- Salaries (4-6 people): $30-60K. CTO + 2-3 engineers + 1 GTM hire, mostly below market.
- Sales tools: $2-4K. Apollo, Clay, HubSpot, Outreach.
- Customer support: $1-2K. Intercom or Front + a part-time CSM.
- Inference + infra: $15-40K. Major spend now. FinOps engineer pays for themselves.
- Engineering team (8-12): $120-200K. Mid-market salaries, mix of senior and mid.
- GTM team (4-8): $40-80K. Sales reps + marketers + CSMs.
- Compliance + ops: $5-10K. SOC 2, security tools, ops headcount.
- Marketing + ads: $10-30K. Variable but real now.
Where lean AI startups beat the bloated ones
Defer the data team to $5M ARR
PostHog Free + Stripe + a CSV export will get you to $1M. Real data engineers cost $200K+ and you don't need them yet.
Outsource the v1 build
A $120K fixed-price MVP from a team that's done it 50 times beats a $800K in-house build that takes 9 months. Even if you hire eventually, you save 6 months.
Pay sales reps less + more commission early
$60K base + 12% of new ARR motivates better than $100K base + 6%. Lean teams need owners.
Buy compliance, don't build it
Drata is $10K/year. A compliance engineer is $150K/year. Buy until you have 50+ enterprise customers.
"AI startups are not magically cheap. The leverage is in compressed time-to-market, not in lower run-rate cost."
Mostly inference + tooling + observability. Real, not magical.
Where the money goes by stage
| Bucket | Pre-revenue | Early ($10K MRR) | Scale ($1M ARR) |
|---|---|---|---|
| Inference | $200-800/mo | $3-8K/mo | $15-40K/mo |
| Infra + hosting | $100-300/mo | $500-1,500/mo | $5-15K/mo |
| Vector DB / search | $0-70/mo | $200-1K/mo | $2-8K/mo |
| Compliance tools | $0 | $2-5K/mo | $5-10K/mo |
| Engineering salaries | $0-3K/mo | $30-60K/mo | $120-200K/mo |
| GTM salaries | $0 | $5-15K/mo | $40-80K/mo |
| Marketing + ads | $0-500/mo | $2-8K/mo | $10-30K/mo |
Lean AI startup playbook
Buy primitives, don't build
Auth, billing, email, compliance - $2-5K/mo replaces 3-5 engineers.
Defer the data team
PostHog + Stripe + CSV exports gets you to $1M ARR. Real data engineers are a $5M+ ARR problem.
Outsource the v1 build
Fixed-price MVP from a team that's done it 50 times beats $800K in-house build that takes 9 months.
Negotiate sales comp aggressively
$60K base + 12% of new ARR motivates better than $100K base + 6%. Lean teams need owners.
We ship paying-customer-ready MVPs in 10-14 weeks.
Auth, billing, admin, eval harness, observability - all wired on day one.
The honest framing
- AI startups cost ~30% more than equivalent SaaS at the same stage.
- Compressed time-to-market is the real leverage, not lower run-rate cost.
- Inference scales with users - architect for routing on day one.
- Buy compliance from day one if you'll ever sell to enterprise.
- Pay yourself nothing as a founder until product-market fit is real.
The honest framing
AI startups are not magically cheap. Inference costs scale with usage, infra grows with users, and salaries are the same as any other startup. The 'AI tax' is roughly 30% on top of normal SaaS costs - mostly inference and tooling.
What's different: you can ship a product with 2 people in 12 weeks. That part is real. The leverage is in compressed time-to-market, not in lower run-rate cost.
The IRPR engineering team ships production software for 50+ countries. Idea → Roadmap → Product → Release. 200+ products live.
About IRPR