OpenAI lost $5 billion in 2024 on $3.7 billion in revenue. Cursor reportedly spends 100% of its $2 billion annualized revenue on Anthropic API costs. GitHub Copilot was losing $20 per user per month when it charged $10. Every major AI coding tool on the market today is subsidized by venture capital, not by the prices you pay.
So what happens when the money runs out?
Developer Dani Akash published a detailed analysis of AI coding economics this week, and the numbers paint a more nuanced picture than the usual "VC subsidy bubble" panic.
The Real Cost of a Coding Session
A typical Claude Code session burns through roughly 592,000 tokens across 24 API requests. Over 90% of those are cache reads, which are charged at a fraction of normal rates, bringing the actual cost to somewhere between $0.20 and $0.65 per session. That sounds cheap until you consider heavy users. Power users of Claude Code can rack up $15-20 per day in actual compute costs - around $400-480 per month - while paying a $200 subscription.
Cursor's situation is even more stark. At its current usage patterns, the company's entire revenue goes straight to paying for the AI models behind the product.
Token Prices Are in Free Fall
Here's where the story diverges from Uber, MoviePass, and every other cautionary VC subsidy tale. GPT-4 cost $30 per million tokens when it launched in March 2023. Today, GPT-5 Nano delivers comparable performance for $0.05 per million tokens. That's a 99.8% price drop in three years.
More broadly, GPT-4-level performance went from $20 per million tokens in late 2022 to $0.40 per million tokens in early 2026 - a 50x reduction. Six factors are compounding simultaneously: custom AI chips (saving 30-60% on GPU costs), quantization (compressing models to run on less hardware), better GPU utilization (jumping from 30-40% to 70-80%), smaller models that match larger ones, distillation techniques, and open-source competition from projects like DeepSeek V3.
The conservative estimate is 5-10x annual cost reduction. The aggressive estimate is 50x per year.
AWS, Not Uber
The Uber comparison is the one everyone reaches for. Riders paid only 41% of actual trip costs in 2015, and prices eventually rose 92% between 2018 and 2021. But ride-hailing has physical cost floors: drivers, cars, fuel, insurance. Those costs don't collapse by 99% over three years.
AI inference is closer to cloud computing. AWS has cut prices 134 times since 2006, not because of subsidies, but because the underlying hardware genuinely gets cheaper. AI token costs follow the same pattern, driven by real engineering improvements rather than unsustainable discounting.
By 2028-2029, even aggressive AI coding usage - the kind that costs $400/month in actual compute today - could run $5-20 per feature. The "intentional builder" who writes careful prompts rather than firing off speculative requests already spends just $2-10 per feature at current API rates.
The subsidy period doesn't need to last forever. It just needs to last until the cost curve catches up. Anthropic projects cash-flow break-even by 2027. Google Cloud's AI division is already profitable. The math is tight, but it works.
One practical side effect worth watching: when flat-rate subscriptions eventually shift to usage-based pricing, "vibe coding" - that exploratory, unplanned approach where you throw half-baked prompts at an AI agent and see what sticks - gets expensive fast. Visible pricing tends to make people more intentional about what they ask for, which may quietly improve code quality across the board.