Quick question: What's the actual price of a "thought"?
Right now, somewhere in a data center, a GPU is burning electricity to generate a paragraph of text, a line of code, or a photorealistic image. And the frustrating part? Nobody can agree on what that output is actually worth. We've commoditized intelligence without ever establishing a unit of measure for it. It's like we invented electricity but decided to bill customers based on "vibes."
If you're building an AI product, investing in the space, or just trying to budget for your company's API costs, this ambiguity is a problem. You're pricing on guesswork. You're forecasting on FOMO. And the market is changing so fast that last quarter's benchmark is this quarter's cautionary tale.
What You'll Gain: By the end of this piece, you'll understand the emerging markets for model access, compute, and data. More importantly, you'll walk away with a mental model for what a "unit of intelligence" might actually be worth, and how to think about pricing in a world where the underlying technology refuses to stay scarce.
The Commodity Trap: Why GPT-4 Is the New Electricity
Here's the uncomfortable reality that foundation model providers don't want you to internalize: their product is converging on zero.
Not zero value. Zero differentiation.
Think about electricity. In the early days, cities had competing power grids. You'd see two sets of wires on the same street. Then standardization happened, and electricity became a utility. Nobody pays a premium for "premium electrons."
AI models are on the same trajectory. GPT-4, Claude, Gemini, Llama. The performance gaps are shrinking. The benchmarks are getting gamed. And open-source models are nipping at the heels of closed-source frontier systems with alarming speed.
The contrarian take: The "moat" around frontier models is shallower than the valuations suggest. The real moat isn't the model. It's:
Distribution (which platform owns the user relationship)
Proprietary data (which company has unique, high-quality training signals)
Workflow integration (which product is too embedded to rip out)
If your AI strategy begins and ends with "we use the best model," you're building on sand. The best model today will be the baseline model tomorrow, and it'll cost a fraction of what you're paying now.
The Three Markets of AI: Access, Compute, and Data
To understand pricing, we need to separate the AI economy into three distinct markets. Each has different dynamics, different scarcity drivers, and different pricing logic.
- The Market for Model Access (Tokens, Seats, Outcomes)
This is the most visible market. You pay per token, per seat, or per outcome.
Token-based pricing (OpenAI, Anthropic): You pay for input and output volume. Feels precise. Is actually arbitrary. The cost of inference is dropping 10x every 12-18 months, but token prices haven't dropped proportionally.
Seat-based pricing (Microsoft Copilot, Notion AI): You pay per user. This is legacy SaaS thinking applied to a technology that doesn't map to seats. It's a transitional model.
Outcome-based pricing (emerging): You pay per successful resolution, per closed ticket, per qualified lead. This aligns incentives but is hard to measure and easy to game.
The unit problem: What is a token? It's not a thought. It's not a word. It's a subword fragment. Pricing intelligence by the fragment is like pricing a novel by the ink.
- The Market for Compute (GPUs, TPUs, and the Scarcity Illusion)
This is where the real money is flowing. Nvidia's margins tell the story. When demand outstrips supply, the supplier captures the value.
But here's the thing: compute scarcity is partly manufactured. Not entirely. But partly.
Real constraints: Advanced chip fabrication capacity (TSMC), high-bandwidth memory (HBM), and energy infrastructure.
Manufactured constraints: Allocation strategies, long-term contracts, and the strategic hoarding of capacity by hyperscalers.
The compute market will eventually stabilize. When it does, the pricing power shifts from chip makers to the companies that own the demand (the applications that users actually want).
- The Market for Data (The Only Real Moat)
This is the least discussed and most important market.
Proprietary data is the only input that doesn't commoditize. Models can be replicated. Compute can be bought. But a decade of anonymized medical records? A proprietary corpus of legal outcomes? Real-time telemetry from a fleet of autonomous vehicles? That's not replicable.
The pricing mechanism for data is still primitive. Most companies either hoard it (creating no value) or give it away (capturing no value). The winners will be the ones who figure out how to license data as a recurring revenue stream, not a one-time sale.
What Is a "Unit of Intelligence" Worth?
We need a framework. Here are three lenses for pricing intelligence, each useful in different contexts:
Lens 1: The Cost-Plus Floor
What does it cost to run the inference? This is the floor. If you're pricing below this, you're subsidizing your customers.
Rough heuristic: Current inference costs range from $0.001 to $0.10 per 1,000 tokens, depending on model size and provider. But this is dropping fast.
Lens 2: The Value-Based Ceiling
What would a human charge for the same output? This is the ceiling.
If a junior lawyer bills $200/hour to review a contract, and your AI does it in 30 seconds, the value is enormous. But you can't charge $200 for 30 seconds of compute. You charge a fraction, and you capture a fraction of the savings.
The insight: AI pricing will settle somewhere between the cost of compute and the cost of human labor. The split depends on how much trust you've earned.
Lens 3: The Outcome Multiplier
What is the measurable result? This is the future.
If your AI sales agent books a meeting that closes a $50,000 deal, what's that worth? A percentage of the deal, or a flat fee per meeting?
Outcome-based pricing is the holy grail because it aligns incentives and removes the "so what?" objection.
The Pricing Paradox: Cheaper Models, Higher Bills
Here's the counterintuitive trend that's catching CFOs off guard: As model costs drop, total AI spending is increasing.
Why? Because cheaper intelligence unlocks new use cases. When something gets 10x cheaper, you don't use the same amount. You use 100x more.
Customer support bots that handle 80% of tickets instead of 20%.
Code generation that writes 50% of your boilerplate instead of 5%.
Document analysis that processes every contract instead of a sample.
The unit price is falling. The unit volume is exploding. And the companies that win will be the ones that can scale their usage without scaling their costs linearly.
The contrarian take: The biggest risk for AI companies isn't that models get more expensive. It's that they get so cheap that the value shifts entirely to the application layer. If intelligence is free, you can't sell intelligence. You can only sell the workflow, the trust, and the distribution.
Actionable Takeaways: Your AI Pricing Playbook
Whether you're buying, selling, or building with AI, here's how to navigate the pricing chaos:
Stop Pricing by Input. Start Pricing by Outcome. Tokens and seats are proxies. They're easy to measure but don't reflect value. If you can tie your price to a customer result (tickets resolved, leads qualified, hours saved), you escape the race to the bottom.
Build Your Data Moat Now. The model is a commodity. Your data is not. If you're not systematically collecting, cleaning, and structuring proprietary data, you're leaving your only durable advantage on the table. License it. Monetize it. Protect it.
Assume Inference Costs Will Drop 10x. Plan Accordingly. If your business model only works at current inference prices, it's fragile. Build for a world where intelligence is nearly free. What's your value proposition then? That's your real business.
The Verdict
We're in the awkward adolescent phase of the AI economy. The technology is real. The value is real. But the pricing mechanisms are borrowed from a world that no longer exists.
The companies that will thrive are the ones that stop trying to price "intelligence" as an abstract unit and start pricing outcomes. They're the ones that treat models as interchangeable inputs, data as the crown jewel, and workflow integration as the ultimate lock-in.
The unit of intelligence isn't a token. It isn't a parameter. It's the value of a decision made, a task completed, a problem solved. Everything else is just accounting.
What's your take? Have you seen any AI pricing models that actually make sense? And for the builders out there: Are you pricing on cost, value, or something else entirely? Let me know in the comments.
Top comments (0)