Selling Dollars for 85 Cents: The AI Revenue Illusion

The token economy isn't 1999. It's something new.

Anthropic hit $14 billion in annualized revenue this month. OpenAI ended 2025 at $20 billion. The gap is closing fast. Anthropic grew from $1 billion to $14 billion in 14 months. No enterprise technology company in recorded history has compounded at this rate at this scale.

But revenue is the wrong scoreboard.

Bill Gurley made this point in a short post this week: these companies both buy and sell tokens. Revenue comparisons without gross margin context are meaningless. In 1999 we called this “selling dollars for 85 cents.”

He’s directionally right. And the full picture is more interesting than that one-liner suggests. Because when you look at the actual economics, the story isn’t “this is 1999 all over again.” The story is “this is a new kind of business that’s figuring out its margin structure in real time, and the trendlines are encouraging.”

How the Token Economy Actually Works

The business equation behind every AI company is simpler than it looks.

An AI company rents GPUs. It uses those GPUs to run a model. Every time a user sends a query, the model reads tokens (input) and generates tokens (output). The company charges per token.

That’s the whole business. Buy compute. Sell tokens.

Here’s where it gets interesting. The cost side has three layers:

Layer 1: Training. Before you can sell a single token, you have to build the model. Anthropic spent roughly $4.1 billion on training in 2025. OpenAI spent $9 billion. These are sunk costs. You spend them before you earn a dollar. And they’re rising. Each generation of frontier model costs more than the last.

Layer 2: Inference. This is the cost of actually running the model for customers. Every query burns GPU-hours. Anthropic rents servers from Google and Amazon. OpenAI rents from Microsoft. This is your true cost of goods sold. Anthropic’s inference costs came in 23% higher than projected in 2025.

Layer 3: Everything else. Salaries, sales, marketing, office space. OpenAI spent an estimated $2.2 billion on sales and marketing and $1.2 billion on staff compensation for inference operations alone in the second half of 2025.

The revenue side is equally straightforward:

API revenue (the majority for both companies): charge per million tokens. Claude Sonnet runs $3 per million input tokens, $6 per million output. A single complex coding session can burn 5,000 to 20,000 tokens. A heavy enterprise customer processes millions of tokens per day.

Subscription revenue: $20/month for ChatGPT Plus, $20/month for Claude Pro. The margin here depends entirely on how much each user consumes. Light users are profitable. Power users cost more to serve than they pay.

Now do the math.

Anthropic’s gross margin in 2025 is roughly 40%. That means for every dollar of token revenue, 60 cents goes straight back to Google and Amazon for the compute that generated those tokens. The $14 billion in revenue becomes $5.6 billion in gross profit.

OpenAI’s gross margin is roughly 48%. Its $20 billion becomes $9.6 billion in gross profit.

The revenue gap looks like $6 billion. The gross profit gap is $4 billion. Same direction, different magnitude. Neither company is profitable yet once you add training costs, R&D, and operating expenses.

This is the core tension. The top line is extraordinary. The bottom line doesn’t exist yet. But “yet” is doing real work in that sentence.

Why the 1999 Comparison Is Too Cynical

The dot-com analogy is useful shorthand. In 1999, companies like Buy.com and Kozmo.com grew revenue by selling products below cost. Revenue went up. Value went down. Investors confused the two until they didn’t.

The AI token economy has surface-level similarities. Companies buy compute and sell tokens. At the application layer, Cursor reportedly paid $650 million to Anthropic while generating $500 million in revenue. The whole stack has moments where someone is hoping the layer below gets cheaper before the money runs out.

But here’s why the cynical reading is probably wrong.

The cost curve is real and steep. This is the single most important difference from 1999. Buy.com had no credible path to improving unit economics. Shipping physical goods below cost doesn’t get cheaper at scale.

AI inference does. And it’s happening fast.

OpenAI’s compute margin went from 35% in January 2024 to 70% by October 2025. That’s a doubling in less than two years. Token costs have dropped by orders of magnitude. Custom silicon is arriving: Anthropic signed a $21 billion TPU deal with Google that should cut per-token costs dramatically. OpenAI is building its own inference chip. Smaller, distilled models already handle the majority of workloads at a fraction of frontier cost.

Anthropic projects gross margins above 60% by 2027 and 77% by 2028. Those aren’t fantasy numbers. They’re extrapolations from a cost curve that has been declining faster than almost anyone predicted. When inference costs drop 10x in two years, the margin math changes completely.

This is not the same as hoping shipping costs will magically drop. This is Moore’s Law with a tailwind.

The subsidy is strategic, not structural. The “below cost” framing may apply to frontier reasoning models and agentic workflows that burn 10-100x more tokens per task. But the workhorse models (Sonnet, GPT-4o) run at healthy margins. OpenAI’s API gross margins were estimated at 75% for GPT-4o in mid-2024.

The blended picture isn’t “selling dollars for 85 cents.” It’s more like investing heavily in frontier R&D while generating real profit on the volume models that carry most of the traffic. That’s a portfolio with a loss leader, not a broken business model. Amazon sold books below cost to build a customer base that eventually bought everything. The playbook has precedent.

The demand signal is unlike anything in 1999. Buy.com customers had zero switching costs. Next purchase, cheapest site wins.

AI infrastructure is different. Developers build on specific APIs. They fine-tune models. They construct evaluation frameworks and prompt libraries. Enterprise customers sign multi-year contracts. Anthropic reports 80% of revenue from enterprise, with 500+ customers spending over $1 million annually. Eight of the ten largest Fortune 10 companies are Claude customers.

This is not manufactured demand sustained by subsidies. This is enterprise adoption at a pace that typically takes a decade, compressed into months. Companies are paying because the value is real. When a coding assistant saves an engineering team 30% of their time, the ROI on token spend is obvious.

Revenue per customer is expanding, not contracting. In token-based models, expansion doesn’t come from selling more seats. It comes from customers building bigger applications and consuming more tokens. One customer’s successful product launch can 10x their token usage overnight. That’s a fundamentally different dynamic than the dot-com era, where growth required constantly acquiring new money-losing customers.

The Honest Risk

None of this means the concerns are baseless. There’s a real tension:

Anthropic is valued at $380 billion. At 40% gross margins, that’s 68x gross profit. For that to make sense, you need revenue to keep growing, margins to expand to 60-70%, and R&D spending to not scale linearly with capability.

Each is plausible. None is certain. And the models customers want most (reasoning, deep research, coding agents) are the ones with the worst unit economics today.

But “the most expensive products have the worst margins” is true of almost every technology in its early phase. Early cloud computing was expensive. Early mobile data was expensive. The pattern is consistent: costs come down, margins expand, and the companies that invested through the expensive phase end up owning the market.

Dario Amodei told Fortune that a twelve-month delay in AI progress would make him bankrupt. That sounds alarming in isolation. In context, it’s the same thing Jeff Bezos was saying in 2001. When you’re investing ahead of a cost curve you believe in, the risk of stopping is greater than the risk of continuing.

What This Means If You’re Building

If you’re building on top of these models, the margin question applies to you with even more force. Your cost of goods sold is someone else’s price card.

But the opportunity is also enormous. The companies that will build the most durable businesses are the ones that decouple their economics from raw token throughput. That means workflow depth, proprietary data advantages, intelligent model routing (cheap models for 80% of tasks, frontier for the 20% that matter), and enough product surface area that the token cost becomes a minority of the value delivered.

The fact that Cursor crossed $1 billion in revenue and launched its own models isn’t a cautionary tale. It’s proof that the application layer can work if you build enough depth to control your own economics.

The Bottom Line

The revenue growth at Anthropic and OpenAI is real and unprecedented. $1 billion to $14 billion in 14 months. Nothing in enterprise tech has done that.

Revenue in a token economy is not the same as revenue in a software economy. When your marginal cost of delivery is 50-60 cents on the dollar instead of 5 cents, the comparison to traditional SaaS breaks down. Gurley is right that gross profit is the better scoreboard.

But the trajectory of that scoreboard matters more than today’s snapshot. And the trajectory, for the first time in the history of AI companies, is pointing in the right direction. Compute margins are doubling. Custom silicon is arriving. Enterprise demand is accelerating. The cost curve that needs to break is breaking.

The question isn’t whether AI companies can build real margin structures. The evidence increasingly says they can. The question is how fast, and whether the current valuations have priced in the timeline correctly.

That’s a valuation debate, not an existential one. And there’s a world of difference between the two.