OpenAI Burns $17 Billion a Year. The AI Business Model Might Be Impossible.

A few weeks ago, Bloomberg reported that OpenAI is finalizing a funding round expected to exceed $100 billion. The round would value the company at more than $850 billion. This would be the largest private funding round in history — more than double the previous record.

Let me put that number in context. $850 billion is larger than the market capitalization of Johnson & Johnson, or JPMorgan Chase, or Walmart. It's roughly the GDP of the Netherlands. It's being assigned to a company that, by its own financial disclosures, burns approximately $17 billion in cash per year.

OpenAI generates roughly $20 billion in annual revenue. It spends approximately 70% of that on compute costs alone — training new models and serving inference to hundreds of millions of users. Add in the 1,500+ employees, the research operations, the data licensing deals, and the legal costs, and the cash burn exceeds the revenue by a wide margin.

The technology works. ChatGPT is used by hundreds of millions of people. The API powers thousands of applications. GPT-5 is, by most accounts, genuinely more capable than its predecessor. OpenAI has built something extraordinary.

But "extraordinary technology" and "viable business" are not the same thing. And the gap between them, in OpenAI's case, is $17 billion per year.

The Unit Economics Problem

Every business has unit economics — the relationship between the cost of delivering a product and the revenue it generates per customer. For SaaS companies, the unit economics are straightforward: build the software once, sell access to many customers, marginal cost near zero. This is why SaaS companies achieve 70-80% gross margins.

AI model companies have fundamentally different unit economics. And this difference is not a phase that will be overcome with scale. It's structural.

The Cost of Training

Training a frontier AI model is a capital expenditure that has grown by approximately 10x per generation:

• GPT-3 (2020): estimated $5-10 million to train
• GPT-4 (2023): estimated $100 million to train
• GPT-5 (2025): estimated $2-5 billion to train (including failed runs and restarts)

The next generation will cost more. Each model generation requires more data, more compute, and more time. Unlike software development — where you build once and iterate — model training is a recurring capital expenditure. You don't train GPT-5 once and sell it forever. You train GPT-5, then train GPT-6, then train GPT-7. Each training run is a new multi-billion-dollar expense.

This is not a startup cost that amortizes over time. It's a perpetual R&D expense that grows with each generation.

The Cost of Inference

Training is an upfront cost. Inference — the cost of actually serving responses to users — is a variable cost that scales with usage.

When a user sends a query to ChatGPT, that query is processed by GPU clusters that consume electricity, require cooling, and depreciate. The cost per query varies by model and complexity, but estimates range from $0.01 to $0.15 per interaction for consumer queries and significantly more for complex API calls.

At 100 million daily active users, even at the low end, that's $1 million per day in inference costs — just for the consumer product. The API, which serves thousands of applications making millions of calls, adds significantly more.

The critical difference from SaaS: in traditional software, serving an additional user costs essentially nothing. The code runs on the same servers whether there are 100,000 users or 10 million. In AI, every additional user, every additional query, every additional token costs real compute. Revenue scales linearly. Costs scale linearly. Margins don't improve with scale in the way SaaS margins do.

The Pricing Pressure

OpenAI faces pricing pressure from two directions:

From open-source models. Meta's Llama, Mistral, DeepSeek, and other open-source models are free. They're not as capable as GPT-5 — but for many use cases, they're good enough. Every improvement in open-source model quality puts downward pressure on what OpenAI can charge.

From competitors. Anthropic, Google, and Amazon all offer competing API products. The market is moving toward commodity pricing for standard inference. OpenAI can maintain premium pricing only as long as its models are perceivably better — and that perception gap is narrowing with each competitor release.

The result: OpenAI needs to continuously increase the capability gap to justify premium pricing, but each capability increase requires exponentially more compute investment. It's an arms race where the cost of competing grows faster than the revenue from winning.

The Historical Parallels

The pattern of "revolutionary technology, unsustainable economics" is not new. Three historical comparisons illuminate the range of outcomes:

Amazon (The Bull Case)

Amazon was unprofitable for nine years after its IPO. Wall Street analysts wrote obituaries. The company was mocked as "Amazon.org" — a charity, not a business. Jeff Bezos was told repeatedly that the economics would never work.

But Amazon's unit economics actually improved with scale. Each additional sale was increasingly profitable because fixed costs (warehouses, logistics infrastructure, technology) were amortized over more transactions. The marginal cost of the next delivery declined. The company was building infrastructure that would eventually generate massive cash flows.

OpenAI bulls point to this parallel: invest now, build the infrastructure, capture the market, and margins will eventually follow.

The question is whether the parallel holds. Amazon's margins improved because the cost of shipping a box didn't increase with each generation of boxes. OpenAI's costs increase because each model generation requires more compute, and each served query requires real-time GPU processing that doesn't amortize.

Uber (The Mixed Case)

Uber burned approximately $25 billion before reaching profitability in 2023. The company revolutionized transportation, achieved massive scale, and eventually found sustainable unit economics — but only after dramatically cutting driver subsidies, raising prices, reducing service quality in unprofitable markets, and adding high-margin products (advertising, Uber Eats).

Uber's profitability didn't come from the original vision working. It came from abandoning the original vision — low prices, massive subsidies, global domination — and building a more constrained but economically viable business.

The Uber parallel for OpenAI: the company may eventually be profitable, but not in the way the current valuation implies. It may need to raise prices dramatically, reduce free-tier access, focus on high-margin enterprise contracts, and accept a smaller market than the current narrative promises.

The Telecom Bubble (The Bear Case)

In the late 1990s, telecom companies raised hundreds of billions of dollars to build fiber optic networks. The technology was real — fiber optic cable is genuinely superior for data transmission. Demand for internet bandwidth was genuinely exploding. The bull case was obvious: lay fiber everywhere, and the revenue will follow.

Approximately $2 trillion in value was destroyed when the telecom bubble burst. The technology worked. The infrastructure was built. The internet did become essential to modern life. But the economics of building the infrastructure didn't work for most of the companies that invested in it. The winners were the companies that used the infrastructure (Google, Amazon, Netflix) — not the companies that built it.

The bear case for OpenAI: the company is building the infrastructure layer (foundation models) while the real value accrues to the application layer (the companies building products on top of the models). OpenAI bears the cost. The application companies capture the margin.

The $850 Billion Math

Let's do the math that the $850 billion valuation implies.

To justify an $850 billion valuation using a standard discounted cash flow model with a 10% discount rate and a 25x terminal multiple, OpenAI would need to achieve approximately:

• $80-100 billion in annual revenue within 7-10 years
• 30%+ operating margins (currently negative 70%)
• Sustained growth at 30%+ annually during that period

For context, Google's annual revenue is approximately $350 billion. Microsoft's is approximately $260 billion. The entire global cloud computing market is approximately $600 billion.

OpenAI reaching $100 billion in revenue would require it to capture approximately 15% of the global cloud computing market — while simultaneously achieving margins that are currently nowhere in evidence.

Is this possible? Perhaps. If AI becomes the primary interface for all computing — replacing search, replacing traditional software, replacing significant portions of human knowledge work — then the total addressable market is enormous. But "enormous TAM" has justified a lot of value destruction in the history of technology investing.

What OpenAI Is Actually Betting On

OpenAI's implicit bet is that three things happen simultaneously:

1. Compute Costs Fall Faster Than Revenue Grows

Moore's Law historically reduced computing costs by approximately 40% per year. If this rate applies to AI-specific hardware (GPUs, TPUs, custom ASICs), then the cost of training and inference should decline dramatically over the next decade.

However, AI workloads have historically grown faster than cost reductions. Each model generation requires 10x more compute while hardware improves at 2x per generation. The net effect is that total compute spending increases even as per-unit costs decline. OpenAI is running up a down escalator — the escalator is getting faster, but so is the running.

2. The Application Layer Doesn't Capture the Value

OpenAI's model assumes that it can capture value at the model layer — that customers will pay premium prices for the best model rather than using the model through application-layer products that commoditize the underlying AI.

But the trend is the opposite. Developers increasingly access AI through application-layer products (Cursor, Lovable, Jasper, etc.) that abstract the model provider. The application decides which model to use based on price and quality. If Anthropic offers comparable quality at a lower price, the application switches. The model layer becomes a commodity.

3. Open-Source Doesn't Close the Gap

Meta has invested billions in Llama. Mistral, DeepSeek, and dozens of other companies are releasing competitive open-source models. If open-source models reach 90% of frontier model quality — which many analysts believe is 12-18 months away — OpenAI's pricing power collapses.

The precedent: Linux reached enterprise-grade quality and fundamentally disrupted commercial Unix. Red Hat built a profitable business on top of open-source, but the total revenue of open-source Linux companies was a fraction of what proprietary Unix vendors earned. The technology democratized, and the value shifted to the application layer.

What This Means for the Industry

The OpenAI economics question isn't just about one company. It's about whether the AI model layer — the foundation models that power the entire AI application ecosystem — can sustain a viable independent business.

If the Model Layer Is Profitable

If OpenAI proves that foundation models can be profitably operated as a business, it validates the entire "AI stack" thesis: model providers at the bottom, platform companies in the middle, application companies at the top. Each layer captures value. The ecosystem is stable.

This is the world most venture capitalists are investing in. It assumes that the model layer has pricing power, that differentiation is sustainable, and that the massive capital investment in training and inference infrastructure will eventually generate returns.

If the Model Layer Is Not Profitable

If the model layer turns out to be a commodity — because open-source closes the quality gap, because competition drives pricing to marginal cost, because the application layer captures the value — then the current investment in AI infrastructure is misallocated.

In this scenario, the winners are the companies building AI-native applications (vertical SaaS, AI agents, domain-specific tools) that use foundation models as a commodity input. The model providers become the equivalent of AWS — essential infrastructure, but not where the majority of value accrues.

The irony: OpenAI spent $17 billion building the technology that might make someone else rich.

The Case for Cautious Optimism

None of this means OpenAI will fail. The company has several structural advantages that could lead to long-term profitability:

Enterprise contracts. OpenAI's enterprise offerings command premium pricing with multi-year commitments. If enterprise revenue grows to represent the majority of total revenue, margins improve because enterprise usage is more predictable and can be served more efficiently.

Custom model training. Fine-tuning and custom model development for large enterprises is a high-margin service that leverages OpenAI's core capability without the marginal cost problems of consumer inference.

Platform economics. The GPT Store, the Assistants API, and the broader developer ecosystem create platform dynamics where third parties build on OpenAI's infrastructure. Platform businesses historically capture disproportionate value.

Hardware integration. OpenAI's investments in custom chips and data center infrastructure could dramatically reduce compute costs over time, similar to how Google's TPUs reduced its own infrastructure costs below market rates.

But each of these advantages requires years to materialize. And in the meantime, $17 billion per year is flowing out the door.

The Investor's Dilemma

The OpenAI funding round presents a clean version of a question that every technology investor must answer: do you invest in revolutionary technology with unproven economics, or do you wait for proof of profitability and risk missing the opportunity entirely?

History offers no clear guidance. The investors who backed Amazon at $1 billion when it was unprofitable made 2,000x their money. The investors who backed WeWork at $47 billion when it was unprofitable lost nearly everything. The technology was real in both cases. The economics were only real in one.

At $850 billion, OpenAI's investors are betting that AI is Amazon, not WeWork. They're betting that the technology is so transformative that the economics will eventually follow, that compute costs will decline, that pricing power will hold, and that the application layer won't commoditize the model layer.

They might be right. The technology is genuinely extraordinary. But "$17 billion in annual burn" and "the economics will eventually work" is a sentence that has been spoken before, about companies that no longer exist.

The technology works. The question — the $850 billion question — is whether the business ever will.