OpenAI’s own internal projections show a cumulative cash burn of $665 billion through 2030, revised upward by $111 billion from estimates made just two quarters earlier. They don’t expect to be cash-flow positive until 2029 or 2030.
Now here’s what nobody in the “AI wrapper business” conversation is connecting: you are building your margin on top of this.
Every $0.20 you pay OpenAI, every Anthropic API call, every Gemini token, you’re transacting with a company that is structurally, mathematically, and presently losing money at a scale that has no precedent in the history of technology companies.
This isn’t a moral argument. It’s a business conditions argument. And if you’re building on these platforms without understanding it, you’re flying blind.
#1 The Number Everyone Cites Is Probably Wrong
Here’s where it gets uncomfortable for anyone who’s been citing OpenAI’s revenue figures in their analyses, pitch decks, or Substack takes.
The “$3.4 billion in revenue” figure that circulated widely in 2024? It was an annualized run rate: one month’s revenue multiplied by twelve. OpenAI has never publicly defined how it calculates “annualized” or “ARR.”
The leaked Microsoft financial documents, the most granular independent data we have, tell a different story. Microsoft’s revenue share payments imply that OpenAI’s actual revenues for the first half of calendar year 2025 were approximately $2.27 billion, around $2 billion short of the $4.3 billion The Information reported.
There’s a possible explanation involving cloud credits Microsoft granted OpenAI that might not appear in these ledgers. The full picture is genuinely murky. But the signal is clear: OpenAI’s costs and revenues are potentially dramatically different to what we believed, and the significance of OpenAI’s inference spend alone raises serious questions about whether the prices it charges customers actually cover what it costs to serve them.
This is the foundation your $0.60 is sitting on.

#2 The Real Problem Is Inference. And It’s Structural.
You know what inference is. It’s the thing that happens when a user sends a prompt and gets a response back. It’s the moment the GPU wakes up, loads the model weights into memory, processes your tokens, and generates output.
It sounds like a simple operation. It is not a cheap one.
OpenAI posted a 33% gross margin in 2025, constrained by inference costs that reached $8.4 billion that year alone and are projected to rise to $14.1 billion in 2026.
Let that land. $8.4 billion. Just to run the models. Not to train them. Not to pay researchers. Not to run offices or sales teams. Just to respond to queries.
Inference costs quadrupled in 2025 as demand for ChatGPT and API services outpaced available compute capacity, forcing OpenAI to purchase computing resources at emergency prices. Their adjusted gross margin dropped to 33%, down from 40% the year before and well below their own internal target of 46%.
Here’s the brutal mechanics of why. When you send a long prompt, every single attention layer across every single transformer block needs to store intermediate computations in GPU memory, what’s called the KV cache. For a short message: negligible. For a 100,000-token document analysis, or a multi-turn agentic workflow, or a coding assistant with large context: the KV cache alone can consume gigabytes of high-bandwidth memory per request. Since the decoding process is memory-bandwidth-bound, not compute-bound, you can’t just throw more GPU cores at it. You need memory throughput. And memory throughput is expensive and physically constrained.
Every user you add makes the company incrementally more expensive to operate, not incrementally more profitable, at current margin levels. That’s the structural trap. Software companies are supposed to have the opposite property. The compute cost problem is structural, not temporary.
#3 Your $0.20 API Payment: What It Actually Pays For
When you route a request through OpenAI’s API, the $0.20 you budgeted covers something like this:
A fraction of a cent goes toward the raw GPU compute, the actual FLOPs to process your tokens. That part is manageable. On a per-transaction basis, with good batching and high GPU utilization, inference can be gross-margin positive.
But then there’s everything else that’s factored into that price, that you never see:
The model itself. Training costs alone are projected to reach nearly $440 billion through the end of the decade. OpenAI plans to spend $32 billion on model training in 2026 and around $65 billion in 2027. The intelligence embedded in the model you’re calling was paid for with capital you’ll never be directly billed for. It’s amortised across every API call in history and future, across consumer subscriptions, enterprise deals, everything. You’re getting the benefit of what may be the most expensive software artifact ever created, priced as a commodity.
The safety stack. Every output passes through classifiers, moderation filters, RLHF-trained reward models. The infrastructure that keeps your product from being a liability. Priced into your token cost. Not shown separately.
The optimization engineering. The teams who built Flash Attention, speculative decoding, continuous batching, custom CUDA kernels, the work that reduced the cost per token year over year so that competitive pricing was even possible. OpenAI spent $2 billion on sales and marketing alone in the first half of 2025. Hundreds of millions on pure research. You benefit from all of it. You pay tokens.
The scale discount you couldn’t buy yourself. Running your own frontier model inference would cost you 10 to 20 times per token versus what OpenAI charges. The gap between those two numbers is OpenAI’s purchasing power with NVIDIA, their relationship with Microsoft Azure, their utilization density across millions of concurrent users. That gap is a structural subsidy flowing to every API developer.
Your $0.20 is not just buying inference. It’s buying access to a capital stack that cost tens of billions to build and is still being funded by investor capital at a loss.
#4 The Anthropic Version Is Different in Degree, Not Kind
If your instinct is to switch to Claude and feel safer about this, the picture is similar, just with different financial discipline.
Anthropic expects to break even around 2028, the same year OpenAI projects $74 billion in operating losses. Anthropic is dropping its cash burn to roughly one-third of revenue in 2026 and down to 9% by 2027.
The philosophical difference is real. Anthropic is deliberately avoiding compute-heavy vanity products, no video generation, less consumer moonshot spending, 80% of revenue from enterprise customers who pay higher prices for compliance, governance, and reliability guarantees. They’re playing a tighter game.
But they’re still burning. Still structurally loss-making. Still building infrastructure that costs more to run than it generates. The same dynamic applies, your API margin is being partially enabled by their balance sheet.
#5 The Three Risks You’re Not Pricing In
This isn’t fatalism. You should still build. But here’s what should change about how you think.
Risk 1: The price floor has been fake.
Token prices fell 10x annually for multiple years. Every builder modelled their unit economics on a cost curve that was dropping because of competition and optimization gains. A broader AI price war has erupted, with Chinese models now priced at one-sixth to one-fourth the cost of comparable US systems, and DeepSeek built a rival AI for approximately $5.6 million (per their own technical report, though analysts note this excludes prior research costs and architectural experiments that likely push the true figure considerably higher) that challenged OpenAI’s flagship reasoning model.
Here’s the tension: competitive pressure keeps prices low, but structural costs keep rising. If it costs this much to run inference for OpenAI, it may be that OpenAI’s costs are dramatically higher than the prices it is charging its customers, which makes price increases necessary at some point, not optional.
The floor you’re building your margin on is priced below cost. That is, by definition, temporary.
Risk 2: Demand doesn’t save them the way the model assumes.
The OpenAI bet is simple: spend aggressively now, capture dominant market share, and when scale kicks in the unit economics flip. The company expects margins between 52% and 67% by 2029 and projects up to $280 billion in annual revenue by 2030. On paper, this is the Amazon or Netflix playbook.
The problem is scale isn’t helping yet. In software, you build the product once and serve a million customers at near-zero marginal cost. In LLM inference, every new user adds proportional compute cost. Every additional user makes OpenAI incrementally more expensive to operate rather than incrementally more profitable at current margin levels. This breaks the standard scale narrative. Profitability requires a technological breakthrough, better hardware efficiency, algorithmic improvements, architectural changes, not just more customers.
Risk 3: The company burning the most capital isn’t necessarily the one that survives.
DeepSeek’s approximately $5.6 million training run, which produced a model competitive with OpenAI’s best, caused Nvidia to lose $600 billion in market cap in a single day, the largest single-session loss in US stock market history. That wasn’t just a market anomaly. It was a signal about the durability of the assumption that the company spending the most on compute automatically wins.
If the compute advantage is not permanent, if open-source models or efficient Chinese labs can produce frontier-quality results at a fraction of the cost, the entire pricing structure that makes your $0.20 API cost sustainable collapses. Not because OpenAI goes bankrupt necessarily, but because the market reprices everything and you get a chaotic 18-month period of uncertain pricing, potential API access disruption, and platform consolidation.
Your product built on a single API dependency has no resilience to that scenario.
#6 What You Should Actually Do
None of this means stop building. The opportunity is real. The products being built are real. The user value is real.
It means build with this in your head.
Treat your current margin as a window, not a guarantee. The pricing structure you operate on today reflects subsidized infrastructure, competitive price suppression, and a capital bet that hasn’t paid off yet. Plan for the window to narrow. Know your minimum viable unit economics at 2x current API cost. Know which features survive that scenario and which don’t.
Route intelligently. Not every query needs frontier models. A 33% gross margin on inference implies OpenAI spent roughly $8.8 billion running models in 2025 alone. The difference between routing a simple classification task to a mini model versus a frontier model is the difference between a $0.001 API cost and a $0.02 API cost. That’s 20x. Across 100,000 daily queries, that’s the difference between a healthy business and a funding-dependent one.
Build depth that outlasts pricing shifts. Every layer of your product that isn’t “model + prompt + thin UI” is a layer that survives API repricing. Memory systems. Domain-specific fine-tuning. Integration with proprietary data sources. Workflow automation. Customer relationships. These are assets. A prompt wrapper is a liability.
Diversify your model dependency. The builder who runs 100% of their inference through one API is one pricing change, one service disruption, or one platform decision away from existential risk. The builder who can route across OpenAI, Anthropic, Gemini, and open-source models has optionality. Optionality is worth engineering complexity.
#7 The Honest Frame You Need
When you make $0.60 on a $1 AI wrapper, you are not extracting margin from a healthy platform.
You are extracting margin from an ecosystem that is, in aggregate, spending more money on inference, the literal act of serving your API requests, than the entirety of its revenue just twelve months ago.
The cumulative losses through 2030 are projected at $665 billion. Training costs alone, for the models powering your $0.20 API calls, are projected to hit $440 billion through the end of the decade.
Someone is paying for the intelligence in your product. That someone right now is a combination of Sam Altman’s fundraising ability, Microsoft’s strategic bet, SoftBank’s capital, and an army of institutional investors who believe the unit economics will eventually flip.
Your business model is implicitly a bet that they’re right.
If they are, the window stays open, prices stay competitive, and you have years to build something defensible on top of subsidised infrastructure.
If they’re wrong, the window closes, prices normalize to actual cost, and the only builders who survive are the ones who used the subsidised window to build something users genuinely cannot live without.
The question isn’t whether you’re ripping off OpenAI.
The question is: what are you building with the borrowed time?
Numbers sourced from: leaked Microsoft financial documents reported by Ed Zitron/Where’s Your Ed At (Nov 2025), The Wall Street Journal internal OpenAI projections, The Information H1 2025 financials, Sacra OpenAI analysis, Epoch AI inference profitability analysis, and The Decoder infrastructure reporting. Revenue figures carry inherent uncertainty due to OpenAI’s non-standard ARR definitions, ranges cited reflect the spread between implied figures from Microsoft revenue share data and OpenAI’s own reported ARR claims. DeepSeek training cost of $5.6M per their technical report; analysts widely note this excludes pre-training research expenditure.


