🟢 Is there an ROI for AI replacing your developers?
A quick discussion on another aspect of AI hype and real costs: when AI is good enough, do you still need developers?
I was recently discussing the only currently proven business case for AI: replacing 80% of the developer base in small to large companies, in exchange for paying for "AI as a Service" through tokens, and found it hard to justify looking forward...
Quick and dirty maths...
Let's consider OpenAI as the key protagonist (OpenAI's 2025 numbers are public enough to do back-of-the-envelope calculations).
Their 2025 revenue came in at around $13B against total spending of roughly $22B, meaning OpenAI now spends approximately $1.69 for every dollar of revenue it generates. That's down from $2.35 in 2024. If you want to be optimistic, the loss ratio is improving quickly, driven by real revenue growth rather than cost discipline, as inference costs alone reached $8.4B in 2025 and are projected to rise to $14.1B in 2026.
If we consider a fair 25% B2B repricing multiplier, OpenAI should sell its tokens at ~2.1x current prices. And with projected cumulative cash burn through 2029 conservatively estimated at ~$115B before turning cash-flow positive, the "real" OpenAI token price will still be at ~2x the current price for the foreseeable future.
Of course, this is a well-known game where OpenAI tries to remove all the funding oxygen from the proverbial room, asphyxiate competitors, and then rapidly raise prices "à la Uber."
What about replacing developers with AI, then?
A mid-level software engineer's fully loaded annual cost is around: US SF/NYC ~$200K, US mid-market ~$135K, EU West (Paris, Munich, Amsterdam) ~$95K, and EU East (Warsaw, Tallinn, Lisbon) ~$50K. If we estimate the token load to approximate one productive engineer-year as of 200 meaningful tasks, full context loops, iteration, review, etc, we get into these numbers:
- Realistic consumption: 100-200M tokens/year
- At today's subsidised prices (GPT-4o blended $5/M): $500-1,000/year
- At real unsubsidized 2025 cost ($10.50/M): $1,050-2,100/year
The replacement ratio at real 2025 economics:
- US engineer vs. real token cost: 65-95x cheaper with AI
- EU West engineer vs. real token cost: 45-90x cheaper
- EU East engineer vs. real token cost: 24-48x cheaper
So does the replacement math work?
The ratio looks overwhelming in favor of AI even at real costs, but three things collapse the model in practice:
First, the human-in-the-loop cost doesn't exactly go to zero ; rather, it shifts upward in seniority. You remove mid-level engineers and immediately need more senior ones to supervise output, maintain coherence, and handle anything ambiguous. Senior engineers in the US run $250-300K loaded. One senior engineer managing AI output at 30% overhead is already erasing most savings if they're replacing two mid-level hires.
Second, token inflation at production scale is difficult to anticipate, let alone control. Replace 5 engineers with agents, and your context windows, orchestration, retry loops, and multi-agent coordination multiply actual token consumption by 3-5x easily. Your $2,000/year per engineer-equivalent quickly balloons to $8,000- $ 10,000 once the pipeline scales to meet real-world needs.
Companies currently building AI business cases on this year's API prices are embedding a 200-300% implicit subsidy into their unit economics. That subsidy depends entirely on OpenAI continuing to burn capital at scale. Nobody is stress-testing this assumption in investment committees.
All in all, the business case for AI replacing developers might hold up in one clear scenario: a high-volume, low-risk environment (where errors are manageable) and well-scoped tasks where the need for senior oversight is minimal. Everywhere else, you're really buying a productivity multiplier for your existing senior engineers, which can still be tremendous, but it's also a meaningfully different and much smaller ROI story.
When even Microsoft is throttling its third-party AI contracts, you should, too, pay attention:
And they're not alone:
Aditi Bharade
