Every founder building a product on top of AI eventually asks: what will this cost me per customer? It's the right question. Ask it before you have customers, though, and it can be the most expensive question you'll ever answer, because the real cost to build a SaaS startup usually sits somewhere else.
Picture the planning meeting. There's a spreadsheet with a tab called "unit economics", a cell that says $50 per customer per month, and a developer explaining, a little apologetically, that the current design comes in over. So the next two months go into caching, prompt trimming and a cheaper model. Nobody launches. Nobody pays. The cell turns green.
The cost to build a SaaS startup is the total spend needed to design, build and ship a working software product to paying customers and keep it running until revenue covers it. Before product-market fit, most of that spend is people's time, not hosting or AI tokens.
Key takeaways
Per-customer cost is worth tracking, and you'll need it eventually. The mistake is treating it as a hard engineering constraint before you know whether anyone wants the product. Before product-market fit, the cost that should keep you up at night is the one that doesn't fall on its own: people's time, and the months it takes to get something real in front of buyers.
Look at where the money actually goes. The median B2B SaaS company spends 27% of revenue on R&D, and companies under $5M ARR spend 32% [C15]. Engineering headcount is the main driver of that number [C15]. Hosting bills rarely look like that in year one.
Here's the part that surprises most first-time founders: the model is usually the cheapest line in the stack. On Together.ai, the open-source gpt-oss-120B model costs $0.15 per million input tokens and $0.60 per million output tokens [C25]. Llama 3.3 70B costs $1.04 per million tokens, input or output [C25]. A million tokens is a lot of text. For many early products, the monthly model bill is smaller than the coffee budget.
The value, and most of the cost, lives in the software around the model: the screens your customers use, the data it connects to, the checks that stop it telling the wrong person the wrong thing, and the bit that lets someone pay you. That's what takes engineering time. And engineering time is the expensive part.
On our illustrative model, a lean four-FTE US team costs about $745k a year, or roughly $62k a month. Nine months to first paying customers comes to about $559k.
The base salaries come from the US Bureau of Labor Statistics and Payscale. To get from a salary to what an employee actually costs you, we used the BLS figure that wages make up 70.0% of private-industry employer compensation, with benefits the other 30.0% [C3]. Divide by 0.70 and you get a "loaded" multiplier of about 1.43. That multiplier is our derivation, not a BLS number.
Illustrative maths (our calculation from the cited inputs):
| Role | Base salary (source) | Loaded (×1.43) | FTE | Annual cost |
|---|---|---|---|---|
| Software developer | $135,980 (BLS, May 2025) [C1] | ~$194k | 2 | ~$389k |
| QA analyst / tester | $104,300 (BLS, May 2025) [C2] | ~$149k | 1 | ~$149k |
| Product manager* | $175,140 (BLS, May 2025) [C32] | ~$250k | 0.5 | ~$125k |
| DevOps engineer | $115,420 (Payscale avg base, 2026) [C5] | ~$165k | 0.5 | ~$83k |
| Total | 4 FTE | ~$745k/yr (≈ $62k/month) |
*BLS has no "product manager" line, so we used "computer and information systems managers" as the closest proxy [C32].
Eighteen months of runway on those numbers is roughly $1.1M. That's our arithmetic, not a benchmark, and it's a lean team: four FTE, where the average seed-stage team in 2025 was 6.2 employees [C28]. For context, Carta data shows a median seed round of $4M and a median of 2.1 years between seed and Series A [C28]. Don't expect developer pay to ease off, either. BLS projects software developer jobs to grow 10% from 2025 to 2035, against 3% for all occupations [C1].
So the biggest line in your MVP development cost is people. Every month you don't ship costs about $62k.
The obvious response is to hire offshore, and per hour it really is cheaper. Accelerance's 2026 survey of offshore developer rates puts junior developers in Asia at $24–31 an hour and seniors at $31–41, with Latin America and Europe higher [C7].
Run the same four-FTE team at $30 an hour:
Illustrative maths (our calculation from the cited inputs):
| US team (loaded) | Offshore at $30/hr | |
|---|---|---|
| Cost per FTE-year | ~$186k average | $62.4k (2,080 hrs × $30) [C7] |
| 4 FTE, per year | ~$745k | ~$250k |
| Plus hidden costs (avg +25%) [C9] | n/a | ~$312k |
| Per month | ~$62k | ~$26k |
| Annual difference | ~$430k cheaper offshore |
The 25% uplift comes from Info-Tech, which found "seven hidden costs of outsourcing that, on average, add 25% to the overall cost of an engagement" [C9]. It also found IT leaders underestimate each of those costs "at least 20% of the time" [C9]. Even Accelerance, which sells outsourcing, argues that "hourly rates are a poor measure of the true cost of software development" [C7].
Here's the pattern that worries us. A founder hires a low-cost team, five months go by, and there's still nothing customers can use. At about $26k a month, that's roughly $130k spent, which hurts. What hurts more is the five months of runway and market window, because those don't come back.
The useful number isn't cost per hour. It's cost per shipped, working outcome. On our figures, the offshore team costs about 40% as much as the US team. If it takes 2.5 times as long to ship the same thing, you've paid the same money and lost the time. That's arithmetic, not a statistic, and the multiplier depends entirely on the team you hire.
Now back to the planning meeting and that green cell. Say your design costs $70 per customer per month to run and your target is $50. Here's what the gap is actually worth, in round numbers.
Illustrative maths (our calculation):
And that's before the price of the AI itself moves.
Very little in your budget gets cheaper on its own. AI inference does, and fast. a16z calls it "LLMflation": "For an LLM of equivalent performance, the cost is decreasing by 10x every year." The price of hitting a given score on a standard benchmark fell from $60 per million tokens in November 2021 to $0.06, a 1,000x drop in three years [C23]. Epoch AI found the price of reaching GPT-4-level performance on a PhD-level science benchmark fell 40x per year [C24]. Engineering salaries, by comparison, rose roughly 2.7% to 3.5% year on year in Levels.fyi's 2025 data [C4].
Big models are still expensive. Epoch estimates frontier training costs are growing 2.4x a year [T8]. SemiAnalysis estimated in February 2023 that ChatGPT cost about $694,444 a day to operate in compute hardware [T10]. At the small end, though, training costs have collapsed:
| Model / project | Reported training cost |
|---|---|
| GPT-4 (Stanford AI Index estimate) | ~$78M compute [T7] |
| DeepSeek-V3 (final run only) | $5.576M [T15] |
| GPT-3 (final run; estimates vary) | ~$2.1M (Epoch AI) to $4.6M (Lambda) [T4][T5] |
| MPT-7B (MosaicML) | ~$200k [T17] |
| GPT-2 reproduction (Karpathy, llm.c) | $672 [T12] |
| Sky-T1-32B fine-tune (UC Berkeley) | under $450 [T16] |
| nanochat (Karpathy) | $92.40 [T14] |
(OpenAI never published compute figures for Whisper, and there's no credible public dollar estimate, so it isn't in the table [T1].)
You almost certainly won't train your own model. The direction is the point: the AI bill you're optimising today will probably shrink without you touching it. The salary bill won't.
None of this is new. Amazon's first profitable year was 2003, a year after it lost $149M [C10]. Airbnb's first profitable full year on a GAAP basis was 2022 [C11]. Uber lost more than $3B on operations in each of 2016, 2017 and 2018 [C13]. It posted its first full-year profit as a public company in 2023 [C12]. Spotify launched in 2008 and didn't turn a full-year profit until 2024 [C14].
AI products run thinner margins than classic software, too. a16z found AI companies' gross margins "often in the 50-60% range – well below the 60-80%+ benchmark for comparable SaaS businesses" [C19]. Cursor ran at negative gross margins, paying compute bills to Anthropic on individual subscribers [C20]. It only recently reached "slight gross margin profitability", helped by its own model, at $2B in annualised revenue [C21]. GitHub Copilot reportedly lost about $20 per user per month on average in early 2023, with some users costing up to $80, while charging $10 [C22].
Every one of them kept shipping while losing money. Paul Graham's version of the lesson: "One of the most common types of advice we give at Y Combinator is to do things that don't scale" [Q1]. Reid Hoffman's is blunter: "If you're not embarrassed by the first version of your product, you've launched too late" [Q2].
There is a right time to sweat the infrastructure bill: when it's big, predictable and yours. Dropbox cut operating expenses by $74.6M over two years after moving off AWS [C16]. 37signals spent $3.2M on AWS in 2022 and projected about $7M in savings over five years by leaving. DHH's explanation: "we're paying as much to rent the donkey every few months as it would cost to buy it" [C18].
Both companies had years of real usage behind those decisions. That's what Donald Knuth was getting at in 1974 when he wrote that "premature optimization is the root of all evil" [Q3]. He didn't say never optimise. His point was that the effort belongs on the small, critical share of the work, once you can measure which share that is [Q3].
Losing money per customer isn't a strategy on its own. CB Insights looked at 431 VC-backed shutdowns since 2023: 70% ran out of capital, 43% cited poor product-market fit, and 19% cited unsustainable unit economics [C27]. MoviePass charged $9.95 a month for a movie a day, an analyst called it "selling dollars for nickels", and its parent company filed for Chapter 7 [C30]. Cinema tickets weren't getting 10x cheaper every year. That's the difference.
So here's where we'd draw the line:
If you're pre-launch, the cheapest move is getting something real in front of paying customers sooner, with a team you measure in working features, not hours logged. The hosting and token bills can wait. By then they'll probably be cheaper anyway. (When you get there, here's what it costs to run a SaaS once it's live: spoiler, it's mostly people's hours too.)
One thing to try before your next planning meeting: put two numbers side by side. What a week of your team costs, and what your per-customer gap costs in a year. If the second number is smaller, stop squeezing and ship.
That's the work yousource's Dev team does: building and shipping production software for founders, plus code audits and production hardening when a prototype needs to survive real customers. If you'd like a second opinion on where your build budget is going, talk to us.
On our illustrative model, a lean four-FTE US team (two developers, one QA tester, half a product manager, half a DevOps engineer) costs about $745k a year, or roughly $62k a month. That puts nine months to first paying customers at about $559k. This is our arithmetic from BLS and Payscale salaries, loaded by about 1.43x for benefits [C1][C2][C3][C5][C32].
Per hour, yes. Accelerance puts 2026 rates in Asia at $24–31 an hour for junior developers and $31–41 for seniors [C7]. Info-Tech found hidden costs add about 25% to an outsourcing engagement on average [C9]. The figure that matters is cost per shipped, working outcome: on our illustrative figures, a team costing 40% as much that takes 2.5 times as long costs the same.
a16z found that for an LLM of equivalent performance, cost is falling about 10x a year, from $60 per million tokens in November 2021 to $0.06 three years later [C23]. Epoch AI found the price of reaching GPT-4-level performance on a PhD-level science benchmark fell 40x per year [C24].
The Stanford AI Index estimated GPT-4 took about $78M of compute to train [T7]. Estimates for GPT-3's final training run range from about $2.1M (Epoch AI) to $4.6M (Lambda) [T4][T5]. SemiAnalysis estimated ChatGPT cost about $694,444 a day to operate in compute hardware in February 2023 [T10]. Small models are far cheaper: Karpathy's nanochat trained for $92.40 [T14].
When the bill is big, predictable and yours. Dropbox cut operating expenses by $74.6M over two years after moving off AWS [C16]. 37signals spent $3.2M on AWS in 2022 and projected about $7M in savings over five years by leaving [C18].
