The true cost of building an AI product in 2025: data from 30 startups
I surveyed 30 AI startups about their monthly costs. Median API spend: $8,400. Median total infra: $23,000. But the distribution is bimodal. Some spend $500/month with open source. Some spend $200,000 on API calls alone.
Everybody asks "what does it cost to build an AI product?" Nobody gives a real answer with real data.
I surveyed 30 AI startups. Not the unicorns. Real companies with 2-30 employees building real products. Here's what they actually spend.
Monthly cost breakdown (30 startups)
| Cost category | Median | Mean | Min | Max | |--------------|--------|------|-----|-----| | AI API costs | $8,400 | $22,100 | $120 | $198,000 | | Cloud infrastructure | $4,200 | $7,800 | $0 (free tier) | $42,000 | | Database/storage | $1,800 | $3,100 | $0 (free tier) | $18,000 | | Developer tools | $1,200 | $1,900 | $200 | $8,500 | | Monitoring/observability | $600 | $1,100 | $0 | $6,000 | | Other SaaS | $2,400 | $3,800 | $400 | $15,000 | | Total infra | $23,000 | $39,800 | $1,200 | $287,500 |
Source: My survey of 30 AI startups, August-September 2025.
The median total infrastructure cost is $23,000/month. But the mean is $39,800, pulled up by a few high spenders. The distribution is wildly skewed.
The bimodal distribution
| Cluster | Companies | Median monthly spend | Approach | |---------|-----------|---------------------|----------| | "Lean" (open source + cheap APIs) | 12 (40%) | $3,200 | Self-hosted Llama/Qwen, minimal API usage | | "Balanced" (mix of API + self-hosted) | 10 (33%) | $24,000 | API for flagship, self-hosted for volume | | "API-heavy" (mostly cloud APIs) | 8 (27%) | $86,000 | Claude/GPT for everything |
The market isn't a bell curve. It's two humps. One group spends under $5K by running open source models on rented GPUs. Another group spends $50K+ by using cloud APIs for everything.
The "balanced" group uses APIs for quality-sensitive tasks (customer-facing responses, complex reasoning) and self-hosted models for volume tasks (classification, extraction, embeddings).
Cost as percentage of revenue
| Revenue range | Companies | AI costs as % of revenue | |--------------|-----------|------------------------| | Pre-revenue | 8 | N/A | | Under $50K/mo | 10 | 38% (median) | | $50K-200K/mo | 7 | 14% (median) | | Over $200K/mo | 5 | 8% (median) |
For early-stage companies making under $50K/month, AI costs eat 38% of revenue. That's brutal. It improves at scale: companies over $200K/month spend only 8% on AI.
The scaling curve is favorable. AI costs grow sub-linearly with revenue because bigger companies are better at optimizing (caching, model routing, cheaper models for simple tasks).
API provider distribution
| Provider | % of startups using it | Avg monthly spend | |----------|----------------------|-------------------| | OpenAI | 73% | $14,200 | | Anthropic | 60% | $11,800 | | Google | 33% | $4,200 | | Self-hosted (various) | 47% | $3,800 (infra only) | | Together AI | 27% | $2,100 | | DeepSeek | 20% | $800 |
Sources: Survey responses.
73% use OpenAI. 60% use Anthropic. Many use both (different models for different tasks). The 47% running self-hosted models surprised me; open source adoption for production use is higher than I expected.
DeepSeek at 20% adoption is notable for a Chinese provider serving non-Chinese startups. The cost advantage is a strong pull.
Engineering team size and AI spend
| Team size | Median AI spend | Spend per engineer | |-----------|----------------|-------------------| | 2-5 engineers | $4,800 | $1,600 | | 6-15 engineers | $28,000 | $2,800 | | 16-30 engineers | $82,000 | $4,100 |
Larger teams spend more per engineer. This makes sense: bigger teams build more complex products that use more AI features. But the per-engineer cost growth suggests that AI costs don't just scale with users. They scale with product complexity.
What I learned
| Finding | Implication | |---------|-----------| | Bimodal cost distribution | The "lean vs premium" choice is strategic, not just financial | | AI costs = 8-38% of revenue | This will compress as prices keep falling | | Multi-provider is the norm | 63% of startups use 2+ AI providers | | Open source is production-ready | 47% self-host for some workloads | | Cost optimization is a skill | The best companies spend 3-5x less than the worst for similar products |
The biggest insight: cost optimization is as much a competitive advantage as product quality. Two startups building similar products can have a 10x difference in AI costs depending on their model routing, caching, and provider choices.
My survey is small (30 companies) and biased toward companies willing to share their costs. The real median is probably higher because companies struggling with costs are less likely to respond.
I'll repeat this survey annually. Tracking how these numbers change as model prices fall will be one of the more useful datasets in my collection.
If you found this interesting, you might also like:
- I counted every AI startup that raised money in Q1 2021. The numbers are strange.
- The cost of running an AI startup in 2022: a data breakdown
- AI funding in Q1 2023 is absolutely bonkers. Let me show you the numbers.
- The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
- 5 charts that explain why GPU prices went insane in 2021
-- dataku